Africa’s Green Office Divide Is Becoming an Investment Risk

Africa’s commercial property market is entering a period in which the environmental performance of buildings is becoming increasingly connected to their financial performance. What was once primarily treated as a sustainability commitment is beginning to influence leasing decisions, operating costs, access to capital and the ability of older properties to compete with newer office developments. The change is not occurring evenly across the continent, but evidence from South Africa and developments in Nairobi, Lagos, Cairo and Casablanca suggest that energy efficiency and building resilience are moving closer to the centre of commercial real estate investment decisions.

The timing is important because Africa still has an enormous amount of urban development ahead of it. Buildings and construction already account for roughly 37% of global emissions, while African cities are expected to accommodate hundreds of millions of additional residents over the coming decades. Much of the property that will serve this population has yet to be developed. Decisions being made today about energy consumption, cooling, water, materials and infrastructure resilience could therefore determine the operating costs and competitiveness of African buildings for decades.

For property investors, however, the immediate argument for greener buildings is becoming increasingly financial. South Africa provides the clearest evidence that higher-performing offices can produce better investment outcomes. Long-term market data covering hundreds of prime and A-grade properties show that certified offices have outperformed comparable conventional buildings over the past decade. During 2025, vacancy among certified properties stood at approximately 10.3%, compared with 13.1% for comparable buildings without certification. Certified properties also produced substantially stronger net operating income per square metre and carried considerably higher average valuations.

Over the ten years to the end of 2025, certified South African offices generated annualised total returns of almost 7%, compared with just over 5% for conventional properties. These figures do not mean that environmental certification alone created the difference. Many certified offices are newer, situated in stronger locations, operated by institutional landlords and equipped to higher overall specifications. Nevertheless, the results indicate that sustainable characteristics are increasingly associated with the segment of the market performing most successfully.

This is particularly visible in Johannesburg, where corporate occupiers increasingly concentrate on better-quality properties in established business districts such as Sandton and Rosebank. The definition of a high-quality office has also changed. Modern interiors and prestigious addresses remain important, but electricity security, water resilience, efficient cooling and increasingly on-site renewable energy have become part of the building proposition.

This introduces an important African dimension to the green-building investment case. In European markets, reducing energy consumption is frequently discussed primarily in terms of emissions and operating costs. In several African cities, it can also be a question of business continuity. A property capable of producing some of its own electricity, storing water and reducing dependence on unreliable infrastructure can offer occupiers protection against disruption as well as lower consumption.

Cape Town is developing along similar lines, although its property economics differ from Johannesburg. Demand for high-quality offices remains strong, but land and development costs can be considerably higher. New construction therefore has to justify a larger capital commitment, increasing the importance of securing strong tenants and protecting long-term property income. Efficient buildings with lower operating requirements and modern infrastructure can become more attractive in that environment, particularly to multinational and larger domestic companies.

The implications for existing office stock could be significant. Older buildings were frequently designed when electricity and water security were less prominent leasing considerations and when corporate environmental requirements were considerably weaker. Properties with inefficient cooling, poor insulation and limited backup infrastructure are increasingly competing against newer buildings offering tenants lower consumption and greater operational reliability. The commercial consequence may not necessarily appear as a simple green rental premium. Instead, the more important effect could be a growing discount applied to inefficient buildings through higher vacancy, weaker tenant demand and increased capital expenditure requirements.

This is why applying a single green rent premium across Africa would be misleading. There is insufficient evidence to support the idea that every certified African office automatically commands rents 5% or 10% above comparable conventional buildings. Rental performance depends heavily on location, specification, age, tenant demand and market supply. Sustainability is increasingly one component of Grade A quality rather than an isolated characteristic capable of determining rent by itself.

Nairobi illustrates this relationship particularly well. Kenya has developed one of Africa’s more established green-building markets outside South Africa, and international certification has become increasingly common among newer high-quality office developments. Multinational companies, international organisations and larger professional occupiers increasingly expect efficient buildings with reliable utilities as part of their accommodation requirements.

Nairobi also remains a competitive office market. Occupancy varies considerably between districts, meaning landlords need to differentiate properties from competing Grade A developments. Environmental performance can contribute to that differentiation, but it cannot compensate for a poor location or excessive supply. A certified building in a weak office district can still struggle to attract tenants, while a well-positioned conventional building can continue to perform strongly.

The more accurate conclusion is that sustainability increasingly strengthens the competitiveness of an already attractive building. It should not be treated as a guarantee of occupancy. This distinction is important because claims that certified Nairobi offices automatically achieve occupancy rates 10 or 15 percentage points above conventional properties cannot be supported reliably by current market evidence.

Lagos presents a different sustainability argument. Nigeria’s commercial property market operates against a background of high energy costs and infrastructure constraints, giving efficient building systems an immediate economic value. Offices that can reduce dependence on externally supplied electricity through solar generation, efficient cooling and improved building management can lower exposure to volatile operating expenses.

For Lagos landlords, the investment case can therefore be less about achieving a theoretical environmental premium and more about controlling the cost of operating the property. Tenants occupying large offices are similarly exposed to electricity and backup-generation expenses, making buildings capable of reducing those requirements potentially more competitive.

International development capital is beginning to reinforce this direction. New projects receiving institutional and development financing increasingly incorporate environmental standards and resilience requirements into their designs. Developments associated with Lagos Free Zone, for example, include plans for certified commercial property alongside broader infrastructure improvements. However, Lagos remains at an earlier stage than Johannesburg when it comes to demonstrating a measurable market-wide relationship between certification and rents or valuations.

Cairo represents another potentially important market. The Egyptian capital’s office stock approached three million square metres during 2026, with new Grade A development increasing competition between landlords. New Cairo and other expanding commercial districts are adding modern office buildings, requiring developers to differentiate projects through specification, amenities, technology and increasingly environmental performance.

Egypt’s green-building market has also expanded significantly. Around 1.8 million square metres of development has obtained EDGE certification, while financial institutions are increasing the availability of capital intended for energy-efficient buildings and related projects. This could gradually influence both new development and refurbishment economics.

The financing issue requires careful interpretation. A certified building does not automatically receive cheaper debt. Lending costs continue to depend on the borrower, currency, project risk and wider financial environment. The emerging advantage is that qualifying projects can access dedicated sources of capital that may not be available to conventional developments. Green loans, sustainability-focused investment programmes and development finance can therefore broaden the financing options available to property owners.

Morocco is developing along a similar path. International financial institutions and domestic lenders have established programmes intended to support energy efficiency and environmentally improved buildings. Casablanca, as the country’s dominant commercial centre, could become an important testing ground for whether this capital produces a larger pipeline of sustainable office development and refurbishment.

This financing could become particularly important for existing properties. Africa’s future green-building market cannot depend exclusively on constructing new certified offices. Large quantities of existing commercial stock will remain in operation for decades, meaning landlords will increasingly have to decide whether refurbishment can extend the competitive life of those assets.

The investment requirements can be substantial. Improving glazing and insulation, replacing cooling systems, installing solar generation, introducing battery storage, reducing water consumption and upgrading building-management systems all require capital. The financial return arrives gradually through lower operating expenditure, stronger tenant retention and potentially improved property values.

Access to longer-term sustainability-linked capital can help make those refurbishment programmes more feasible. This could prove particularly important in markets where conventional financing remains expensive and landlords might otherwise postpone major upgrades.

For investors, the calculation is increasingly becoming one of future income protection. A property may still be well located and structurally sound while gradually losing competitiveness because its operating costs and infrastructure no longer meet occupier expectations. Environmental obsolescence can therefore develop before physical obsolescence.

This creates a potentially significant challenge for owners of ageing office portfolios. Delaying investment may preserve cash in the short term but increase the amount of capital eventually required to reposition the building. In extreme cases, refurbishment costs could become difficult to justify relative to the property’s value, leaving landlords with increasingly stranded secondary assets.

The most sophisticated investors are therefore likely to examine environmental performance as part of normal asset management rather than treating it as a separate sustainability exercise. Energy consumption, water security, cooling efficiency, renewable generation and resilience can increasingly be evaluated alongside rent, vacancy, lease expiry and capital expenditure.

Corporate occupiers are reinforcing this change. International companies attempting to reduce emissions across their operations increasingly need information about the environmental performance of the properties they occupy. As more high-quality certified alternatives become available, leasing inefficient space may become harder for those companies to justify internally.

That creates an advantage for landlords capable of providing credible performance data. Certification can support this process by offering independent verification, but the underlying performance of the property remains more important than the label itself. A building that consumes less electricity, manages water efficiently and maintains operations during infrastructure interruptions provides tangible benefits regardless of the certification displayed at its entrance.

This could gradually create a more pronounced two-tier office market across Africa. At the upper end will be modern and successfully refurbished properties combining strong locations with efficient building systems, dependable infrastructure and credible environmental performance. These buildings will be best positioned to compete for multinational companies, financial institutions and larger domestic occupiers.

The second tier will contain older buildings requiring increasing investment to maintain their position. Some will be successfully refurbished. Others may have to compete primarily through lower rents, while the weakest properties could eventually require conversion or redevelopment.

South Africa is already providing measurable evidence of this divide. Nairobi is demonstrating how sustainability can strengthen a building’s position in a competitive leasing market. Lagos shows how energy resilience can become a direct operating advantage, while Cairo and Casablanca illustrate how green financing can begin influencing the development and refurbishment pipeline.

The markets remain at very different stages, making continent-wide assumptions about rental premiums or occupancy advantages inappropriate. The direction of travel, however, is increasingly visible.

Africa’s green-building transition is becoming less about whether a property can carry an environmental label and more about whether the building can remain financially competitive as tenant requirements, operating costs and financing standards change.

For commercial property investors, that changes the central question. The issue is no longer simply whether a green office can command a higher rent. It is whether an inefficient building will increasingly have to accept lower rents, spend more capital and face greater vacancy simply to compete.

If that trend continues, sustainability will become more than a competitive advantage for Africa’s best buildings. For the continent’s ageing commercial property stock, it could increasingly determine which assets protect their income and which begin to lose value.

Source: © CIJ.World Africa Research & Analysis Team

AI Is Starting to Rebuild HR From Hiring to Retention

AI is starting to rebuild HR from hiring to retention by connecting recruitment, employee support, workforce analysis and retention through systems that can interpret large volumes of workforce information while still leaving critical employment decisions in human hands. The shift is becoming more significant than simple automation. Traditional HR software was largely designed around standardised processes serving the majority of customers. AI is making it possible to create more adaptable systems that can respond to the structure, policies and needs of individual organisations rather than forcing every company into the same workflow.

That transition was central to a discussion at AI4 2026 involving representatives of HireVue, BambooHR and satellite manufacturer Astranis. Despite approaching AI from different positions, the participants broadly agreed that its greatest value in HR lies not in replacing professionals but in improving access to information, removing repetitive administrative work and allowing human judgement to concentrate on decisions where context matters most. Recruitment is already one of the clearest examples. Generative AI has made it far easier for candidates to produce polished résumés, application responses and cover letters, which has reduced the effort required to apply for jobs and contributed to employers receiving large volumes of increasingly professional-looking applications.

The unintended consequence is that the résumé itself is becoming a weaker indicator of genuine ability. When candidates can use AI to improve how their experience is presented, recruiters need better ways of determining who can actually perform the job. This is pushing recruitment technology away from simply screening applications and towards validating skills. HireVue has been developing AI-supported interviewing around this problem, using technology to expand the number of candidates who can be assessed while maintaining human oversight over final hiring decisions.

The company launched a voice-based AI Interviewer in 2026 that can conduct structured conversations with candidates without requiring a recruiter to personally perform every initial interview. For organisations involved in large-scale recruitment, this can change the economics of the process. Employers may need to handle tens of thousands of applicants within short hiring periods, and conventional recruiter capacity inevitably forces them to reduce those numbers before meaningful conversations can take place.

AI can potentially loosen that constraint. Instead of deciding which applicants deserve an interview primarily from résumés and application forms, employers can conduct structured initial assessments across much broader candidate groups. That could give more people an opportunity to demonstrate their capabilities before a recruiter decides where to invest limited human time. It does not mean everyone reaches a final interview, but it can increase the amount of information available before candidates are eliminated from the process.

The opportunity is accompanied by another problem: candidates now have access to increasingly sophisticated AI themselves. Applicants can use AI to prepare résumés, anticipate interview questions and generate suggested responses. Tools capable of assisting people during interviews are also becoming more accessible. Recruitment is therefore entering an unusual arms race in which employers use AI to evaluate candidates while candidates use AI to improve how they appear to employers.

The result is likely to increase the importance of skills validation. Employers increasingly need evidence that candidates can do what they claim rather than relying primarily on how effectively they describe their experience. AI could also improve matching between applicants and vacancies. Someone might apply for one role while possessing skills better suited to another position within the same organisation. Historically, identifying that opportunity depended heavily on a recruiter recognising it. Systems capable of analysing capabilities across a company’s entire vacancy portfolio can potentially surface alternative matches automatically.

This shifts recruitment from processing applications towards understanding capabilities. The same principle extends beyond the hiring stage. One of the larger opportunities discussed at AI4 involved connecting recruitment information with what happens after someone joins the company. Businesses collect substantial amounts of workforce information throughout the employment lifecycle, including interview feedback, assessment results, performance reviews, engagement surveys, management evaluations and exit interviews. Much of this information has historically remained fragmented across different systems or stored as unstructured text that is difficult to analyse systematically.

Generative AI changes the economics of working with that information. Organisations can potentially examine patterns across thousands of written comments without requiring someone to manually read every document. That makes it possible to ask more complicated questions about employee performance, management quality, retention and organisational behaviour.

Management performance provides one example. Companies have traditionally evaluated managers through financial results, formal appraisals and perhaps employee surveys. AI-supported analysis can potentially combine those measures with turnover patterns, peer comments, direct-report feedback and exit interviews over time. If unusually large numbers of employees repeatedly leave the same team, management can investigate whether the explanation lies with recruitment, workload, leadership or another organisational issue.

The technology does not automatically provide the correct answer. Its value lies in making previously difficult qualitative information easier to examine. HR has never lacked information. The problem has been getting the right information to the right person at the moment it becomes useful. AI can increasingly act as an intelligence layer between workforce databases and managers.

BambooHR is moving in this direction with a connected AI architecture designed around organisational context rather than individual AI features. Its approach combines workforce information with AI systems capable of analysing data, answering employee questions and completing selected actions while operating within existing permissions and governance controls. That reflects a broader transformation occurring throughout enterprise software.

Traditional software-as-a-service platforms were generally built around the requirements shared by the largest possible number of customers. Developers identified what most users needed and constructed relatively standardised products around those workflows. Customers could configure them, but usually only within boundaries established by the software provider. AI makes it possible for the underlying platform to remain standardised while the user experience becomes more individualised.

Different organisations can potentially interact with the same HR platform in different ways according to their structures, policies and workforce characteristics. Individual employees may also receive different information depending on their role, permissions and circumstances. From the customer’s perspective, the number of AI agents operating behind the software matters far less than whether the system improves the experience and produces useful results.

An HR director is unlikely to care whether a provider operates ten agents or several hundred. The important questions concern whether employee requests are answered correctly, managers receive useful information, repetitive administration is reduced and important decisions become better informed. This creates pressure on technology companies to move beyond simply attaching AI features to established software in order to market the product as AI-enabled.

The underlying architecture is becoming more important. HR systems contain some of the most sensitive information in an organisation, including compensation, performance data, employment records, benefits and recruitment information. AI systems operating across these datasets require strict controls governing which information can be accessed and what actions can be taken.

Data architecture is also becoming a competitive issue. Many established enterprise platforms were designed long before generative AI existed. Their databases expanded gradually as companies added customers, integrations and new functions. Information can therefore be fragmented across numerous systems and tables that were never designed to support AI-driven analysis. Placing an intelligent interface above poorly organised information does not solve the underlying problem.

This could trigger a significant investment cycle across enterprise software. Existing platforms may need to redesign data structures, improve interoperability and create more consistent information models if they want AI systems to operate effectively across their products. At the same time, new HR technology companies have an opportunity to design systems around AI from the beginning.

Legacy software providers retain major advantages in customer relationships, historical information and established workflows. New entrants, however, can build platforms specifically around connected data, adaptable workflows and AI without needing to accommodate decades of older architecture. This may create a new divide within HR technology between companies that place AI above existing systems and those that redesign the underlying infrastructure around it.

Corporate customers are simultaneously gaining another option: building applications themselves. This could become one of the more disruptive consequences of generative AI for enterprise software. Historically, an HR department wanting specialised technology generally had two choices. It could purchase a commercial product or ask an internal engineering team to create something. Custom development was normally realistic only for larger organisations with substantial technical resources. That barrier is falling.

At Astranis, members of the people organisation have been building internal AI tools to support recruitment and administrative workflows. According to the AI4 discussion, one system can help recruiters prepare job descriptions, create educational material, map talent pools, develop sourcing strategies, produce outreach campaigns and organise interview plans through a common interface. Tasks that once required substantial manual research can therefore be compressed significantly.

More importantly, the people creating these tools understand the processes themselves. This creates a new form of enterprise development in which domain specialists become builders. HR professionals, recruiters, finance teams and operational managers no longer necessarily need to translate every requirement into a specification for an engineering department. AI-assisted development can allow them to prototype or create applications directly.

The person who understands the problem and the person capable of constructing the solution are therefore moving closer together. That could reshape enterprise software procurement. Companies will still buy sophisticated platforms where reliability, security, compliance and scale justify specialist providers, but they may increasingly create smaller applications around those platforms themselves rather than purchasing another standalone software product for every process.

The future may therefore be neither entirely build nor buy. Businesses could continue purchasing core systems of record while developing customised intelligence and automation layers around them. This places pressure on software providers to make their platforms more open and extensible. Systems that allow customers to connect data, create agents and integrate external AI environments may become more valuable than closed products offering only predetermined workflows.

The implications extend beyond HR technology. Over the past two decades, companies have bought specialised applications for increasingly narrow business requirements. AI could consolidate some of those functions because adaptable agents can potentially perform workflows that previously required separate products. That does not mean established software disappears. Systems responsible for payroll, compliance, financial transactions and authoritative workforce records remain difficult to replace. But the number of additional applications surrounding those core systems could eventually decline.

Value may instead migrate towards the quality of the underlying data and the intelligence capable of using it. HR provides a particularly clear example because context is essential. A general AI model can recognise patterns, generate text and answer broad questions, but it does not automatically understand an organisation’s compensation rules, management structure, recruitment policies, employee history or internal procedures. Connecting AI with that context dramatically increases its usefulness.

An employee asking about leave, for example, does not simply need a generic explanation of company policy. The relevant answer may depend on location, employment status, available allowance, previous leave and the permissions of the person making the request. An HR system capable of understanding those relationships can move from answering generic questions towards performing useful work.

This is also why companies need to improve their data before expecting major returns from AI. Automating a badly designed process does not make it good. Poor information hygiene, inconsistent records and unclear workflows can simply produce faster mistakes. The panel repeatedly returned to this point. Organisations should resist the temptation to automate everything at once and instead identify specific problems where AI can produce measurable value.

High-volume, repetitive processes with relatively limited downside are obvious starting points. Tasks involving irreversible consequences, complex judgement or substantial regulatory exposure require greater caution. Companies can begin with employee questions, reporting, recruitment research, meeting summaries and administrative preparation before progressing towards systems that influence employment decisions or take actions autonomously.

Hiring is particularly sensitive because the consequences directly affect people’s livelihoods and expose employers to regulatory and reputational risk. Human oversight therefore remains important. AI can organise information, conduct preliminary assessments and identify patterns that people might otherwise miss, but final decisions involving hiring, promotion, discipline or dismissal are much more difficult to automate responsibly.

The challenge is finding the appropriate boundary between machine assistance and human authority. There is also an important distinction between productivity claims and actual productivity. AI adoption alone does not guarantee efficiency. Organisations need to determine whether systems genuinely reduce work rather than simply moving it elsewhere. An automated process that requires extensive checking may deliver far less value than the initial time-saving estimate suggests.

This is especially relevant as companies begin deploying increasing numbers of agents. Counting agents is not a useful measure of transformation. Business outcomes are. A company with several carefully designed AI workflows solving important problems may achieve more value than an organisation deploying hundreds of agents without clear objectives.

HR departments therefore need to approach AI as an operational redesign rather than simply a technology purchase. The starting point should be the business problem. Companies can identify a high-friction process, determine why it performs poorly, clean the relevant information and then decide whether AI is an appropriate solution.

Retention could become one of the most important areas. Companies invest substantial resources in recruitment but often possess limited ability to identify early signs that valuable employees are disengaging or likely to leave. Workforce systems can increasingly analyse patterns associated with absence, declining engagement, turnover or other retention risks. Used carefully, these signals could allow managers to intervene earlier.

The danger is allowing probabilistic analysis to become an unquestioned judgement about an individual employee. A model identifying a possible retention risk is very different from knowing why someone may leave. This reinforces the broader principle emerging across HR technology: AI is most useful when it expands human understanding rather than pretending to replace it.

The same applies to recruitment. AI can allow more applicants to be evaluated, but scale creates value only if the assessment remains meaningful. It can analyse years of workforce information, but patterns still require interpretation. It can automate employee support, but sensitive situations continue to require empathy and judgement.

The future of HR may therefore become both more automated and more human. Routine administration can increasingly move into the background while HR professionals concentrate on decisions, relationships and organisational problems requiring context. The skills needed inside HR departments will change as a result.

Process knowledge alone will no longer be enough. Professionals will increasingly need to understand data, automation and the capabilities and limitations of AI systems. Some will effectively become technology builders within their own functions. That transition is already visible in emerging roles combining domain expertise with technical implementation.

Instead of engineers attempting to understand every detail of HR operations, organisations can develop specialists capable of translating directly between workforce problems and AI systems. The broader economic implication is that generative AI may weaken the traditional division between business departments and technology departments.

Marketing professionals can increasingly build marketing applications. Finance professionals can construct analytical workflows. HR teams can create recruitment tools. Engineering does not disappear, but its role shifts towards architecture, security, reliability and helping business specialists move prototypes into production.

That final step remains critical. Building a demonstration has become dramatically easier. Maintaining reliable enterprise software has not. Applications still fail. Models change. Data requires protection. Access permissions need management. Systems require monitoring and support, while regulatory obligations remain.

The democratisation of software development therefore does not eliminate engineering discipline. It makes that discipline relevant to many more people. For HR technology companies, the strategic contest will increasingly concern who can combine flexibility with trust. Customers want systems capable of adapting to their organisations, but they also need confidence that sensitive workforce information remains protected and employment decisions can be explained.

That combination will be difficult to achieve, but it could determine which platforms become the foundations of the next generation of HR technology. AI is already making recruitment faster and administrative work easier, yet the larger transformation is only beginning. Workforce systems are moving from databases that record what happened towards intelligence layers that attempt to understand what is happening and recommend what should happen next.

If that transition continues, the most important HR technology may no longer be the application employees open. It may instead be the intelligence connecting recruitment, performance, workforce data and employee experience behind the scenes. The companies that succeed will not necessarily be those deploying the most AI. They will be those that combine reliable data, organisational context and human judgement to solve specific workforce problems.

That makes the next phase of HR technology less about replacing people than redesigning the systems around them.

Source: CIJ.World Research & Analysis Team

When AI Starts Doing the Shopping, Consumer Brands Will Have to Win Over Algorithms as Well as People

Consumer brands have spent decades learning how to influence people at the moment they decide what to buy. Artificial intelligence is beginning to change that relationship because the next important shopper may not always be a person navigating a supermarket aisle, website or mobile application. Increasingly, an AI assistant could help decide which products consumers see, compare and eventually purchase. That emerging shift was at the centre of a presentation by Lizbeth James of Mars at Ai4 2026 in Las Vegas, examining how artificial intelligence could transform shopper intelligence within the consumer packaged goods industry.

The change is taking place on two sides simultaneously. Retailers are giving manufacturers access to increasingly detailed first-party commerce information, while AI assistants are becoming capable of helping consumers discover, evaluate and purchase products. Together, those developments could fundamentally alter how brands understand demand and compete for sales. Consumer-goods companies have traditionally depended on market research, household panels, retailer reports and periodic sales information, but large retailers can now provide much more detailed data about customer behaviour across physical stores, websites, advertising and digital transactions.

The scale and speed of this information creates a new problem. Brands may receive hundreds of measurements covering sales, inventory, search visibility, advertising performance, product-page engagement, customer switching and other indicators across thousands of products. Giving managers more dashboards does not necessarily solve the problem because humans still need to identify which signals matter, understand how they are connected and decide what action should follow. James argued that the next stage is therefore a move from business intelligence that explains what happened towards systems capable of continuously detecting changes, investigating possible causes and proposing commercial responses.

A decline in one coffee product, for example, might initially appear to be a straightforward sales problem. But the underlying explanation could involve several factors occurring simultaneously, including customers switching brands, changes in price, reduced online visibility, lower availability, stronger competitor advertising, differences between online and in-store purchasing or changes within particular customer groups. Traditionally, different teams might investigate each part separately. An AI system could potentially examine those signals together and identify relationships that are difficult to see through individual reports.

The commercial opportunity lies in connecting that diagnosis directly with action. If a product is losing visibility online, the system might recommend changes to product information or advertising. If consumers are moving towards a competitor because of price, it might model different promotional responses. If certain products are frequently substituted during online fulfilment, the manufacturer and retailer could investigate assortment or inventory. The objective is to shorten the distance between detecting a problem and doing something about it.

This does not mean conventional analytics are disappearing. Traditional statistical methods and machine-learning models remain important alongside generative AI. Forecasting, segmentation, clustering and other established techniques can provide the quantitative foundation, while generative and agentic systems make it easier for employees to interrogate the information, connect different sources and translate results into potential actions. A reliable data foundation therefore becomes more important rather than less important as AI adoption increases, because retailer information, advertising results, inventory data, digital-shelf performance and pricing all need to be harmonised before autonomous systems can make useful decisions.

Privacy creates another important boundary. Consumer-goods companies generally do not need to know the identity of every individual shopper to identify useful patterns. Retailer environments can use privacy-protected identifiers and aggregated groups to understand behaviour without simply handing manufacturers personally identifiable customer information. The resulting intelligence can show how groups of customers purchase, switch brands or respond to promotions while operating within retailer-controlled data environments and applicable privacy rules.

The more disruptive change, however, is occurring on the consumer side. Shopping assistants are moving beyond answering questions towards performing parts of the purchasing process. Consumers can increasingly use AI to research products, compare alternatives, examine prices, assemble shopping carts and automate some routine purchases. As these systems become more capable, people may increasingly delegate parts of product discovery and selection to software.

That development changes the traditional idea of the digital shelf. Until recently, a brand mainly needed to perform well in retailer search results, category pages, advertising placements and conventional search engines. In an agent-mediated environment, the consumer may instead ask for a week’s groceries within a particular budget, a coffee matching certain preferences or a replacement household product offering the best combination of price and quality. The AI assistant then decides which products deserve consideration.

Brands would consequently be competing for algorithmic recommendation as well as human attention. Packaging, imagery and advertising will remain important when consumers make visual choices, but product information will also need to be sufficiently structured, accurate and credible for AI systems to understand. Availability, price, reviews, product attributes, retailer information and external sources could all influence what an assistant chooses to recommend.

This creates an emerging commercial challenge around making products understandable and discoverable by answer engines and shopping agents. The concept remains immature, and brands should be cautious about claims that AI recommendations can simply be manipulated in the same way companies once optimised webpages for search engines. Nevertheless, the direction is significant. If an AI assistant becomes the interface between a consumer and millions of products, being accurately represented by that system becomes commercially important.

Retail media could change alongside it. Advertising has traditionally influenced people while they browse, search or consume media. Agentic commerce raises the question of what advertising means when software is helping make the decision. Retailers and technology platforms will have to determine how sponsored recommendations, organic recommendations, consumer preferences and commercial incentives coexist without undermining trust in the assistant.

For manufacturers, this creates a new competitive dimension. Large consumer-goods groups possess enormous amounts of historical information, established retailer relationships and substantial marketing resources, but they can also carry complicated organisational structures. James argued that this can slow implementation because insights frequently have to travel through multiple departments before action is approved. Smaller brands may have fewer resources but can sometimes respond more quickly, meaning AI does not automatically strengthen the largest companies and could instead reward businesses capable of converting information into action faster.

This helps explain why the organisational side of autonomous intelligence may prove harder than the technology. James described internal experiments in which AI systems improved their ability to answer business questions after receiving repeated feedback from employees. But even an apparently high level of accuracy leaves important questions about the remaining errors, particularly if a system is eventually allowed to take commercial actions without human approval.

Human oversight is therefore likely to remain essential for many decisions. Automatically identifying a sales anomaly is relatively low risk. Changing a product description may carry somewhat more risk. Altering pricing, reallocating substantial advertising budgets, changing assortment or making decisions affecting customers and supply chains can have much larger consequences. The degree of autonomy should therefore depend on the potential impact of the decision rather than on whether the technology is technically capable of executing it.

The transformation may initially be less dramatic than the idea of an autonomous commercial organisation suggests. Instead of removing employees entirely, AI is more likely to compress the time required for analysis. Work that previously required several analysts to collect reports, reconcile data and prepare presentations could increasingly be produced through a conversational interface that alerts different executives to the issues relevant to their responsibilities.

This could eventually reduce corporate dependence on conventional dashboards. A sales executive may not need to open a series of reports every Monday morning if an AI system can identify the most significant changes from the previous week, explain likely causes and suggest which issues deserve investigation. Brand managers, sales teams and senior executives could interact with the same underlying information through different interfaces tailored to their responsibilities.

The economic value comes from speed. Consumer demand can change quickly, competitors can adjust prices or advertising almost immediately and digital availability can fluctuate throughout the day. A company that requires several weeks to understand what happened may be reacting to conditions that have already changed again. Systems capable of reducing that cycle to hours or minutes could provide a meaningful competitive advantage.

But the most consequential development may ultimately take place outside the consumer-goods company itself. If consumers increasingly ask AI assistants to replenish household products, compare alternatives and choose between brands, manufacturers will have less control over the point at which purchasing decisions are made. The traditional battle for shelf space will increasingly be accompanied by a battle for inclusion in the recommendations generated by machines.

That does not mean consumers disappear from the decision. People will continue to have favourite brands, budgets, tastes and values, and they can override recommendations. But AI could increasingly filter the enormous number of choices before a person sees them. In categories involving routine purchases, consumers may eventually delegate much of that filtering altogether.

The implication for Mars, Nestlé, Coca-Cola and thousands of other consumer brands is significant. The next generation of shopper intelligence will not simply be about understanding what people bought yesterday. It will involve understanding how retailer algorithms, advertising systems, consumer preferences and AI shopping assistants interact to determine what gets purchased tomorrow.

Consumer companies have spent generations learning how to win the physical shelf and the past two decades learning how to win the digital shelf. They may now have to master a third environment: the algorithmic shelf. In that market, the brands with the largest advertising budgets will not necessarily have an automatic advantage. Success may increasingly depend on which companies can organise their data, detect changes quickly, make their products understandable to AI systems and respond before competitors recognise that shopper behaviour has changed.

The transition is still at an early stage, and fully autonomous commercial decision-making remains considerably more ambitious than today’s deployments. But agent-assisted shopping is no longer theoretical. As retailers and technology companies give AI systems greater roles in product discovery and purchasing, consumer brands face a strategic question that barely existed a few years ago: when the shopper’s AI starts deciding what deserves to be bought, how does a brand make sure it remains part of the choice?

Source: CIJ.World Research & Analysis Team

Upwork Is Betting That AI Will Restructure Work Rather Than Eliminate It

Artificial intelligence is widely portrayed as a threat to knowledge workers, but Upwork is preparing for a different outcome. Rather than assuming AI will simply replace the freelancers operating on its marketplace, the company is rebuilding its platform around a future in which people and autonomous software increasingly work together. The shift represents a fundamental change in what Upwork is trying to become. Historically, the platform connected businesses seeking particular skills with independent professionals capable of completing the work. In the emerging model, the customer may increasingly begin with an objective rather than a job description, while artificial intelligence determines what work needs to be completed, breaks it into individual components and identifies whether each component should be handled by a person, an AI system or a combination of the two.

Andrew Rabinovich, Upwork’s Chief Technology Officer and Head of AI and Machine Learning, described this transition during a discussion at AI4 2026. Having worked in artificial intelligence and machine learning for more than two decades, Rabinovich challenged the assumption that current systems are close to eliminating the need for human expertise across the economy. His argument is not that AI will leave employment unchanged. Quite the opposite. Many routine tasks are already becoming automated. Simple translation, basic summarisation, straightforward graphic production and other relatively predictable assignments can increasingly be performed by machines. What remains much more difficult is work involving ambiguity, judgement, taste, context and responsibility for determining whether the final result is actually good enough.

That distinction is becoming central to Upwork’s strategy. The company is moving beyond its historical role as a talent marketplace towards something resembling an operating system for knowledge work. Instead of asking a customer to decide in advance whether a project requires a designer, software engineer, marketing specialist or another professional, its AI systems increasingly attempt to understand the desired business outcome first. A company might know that it wants to increase online sales, for example, without knowing whether the solution requires changes to its website, advertising strategy, customer experience, pricing or product presentation. AI can potentially interpret that objective, divide it into smaller pieces of work and then organise the resources required to complete them.

At the centre of this strategy is Uma, Upwork’s AI work agent. The system is being developed not primarily as a replacement worker but as an orchestration layer connecting customers, independent professionals and AI capabilities. The distinction matters. Instead of asking one AI model to complete an entire complex project, the platform can potentially decompose the work into a sequence of smaller tasks. Some may be suitable for automation. Others may require a specialist. Some may involve an AI producing an initial result before a professional reviews, corrects or develops it further.

Upwork has an unusual advantage in attempting this because it possesses years of information about actual commercial work. Its marketplace has recorded project descriptions, skills, hiring decisions, completed assignments, payments and customer feedback across millions of engagements. That historical information can help the company understand how successful projects are structured and which combinations of skills tend to produce satisfactory outcomes. In an AI economy, this type of proprietary workflow information may become considerably more valuable than conventional marketplace data because it provides examples not simply of what people say they can do, but of work that customers were willing to pay for and subsequently accepted.

This is particularly important because evaluating professional work remains one of AI’s largest unresolved problems. Certain assignments can be checked relatively easily. Software either functions or it does not. A mathematical proof can be tested. A structured data transformation can be compared against a defined result. Most professional work is considerably less clear. There may be no objectively correct advertisement, architectural concept, marketing strategy, presentation or piece of writing. Whether the result is successful depends partly on the expectations and judgement of the person commissioning it.

Upwork estimates that much of the activity on its platform falls into these more qualitative categories. That creates a problem for fully autonomous AI because producing something is not the same as knowing whether the result meets the client’s actual requirements. The company’s Human+Agent Productivity Index is intended to investigate this gap using real commercial projects rather than traditional academic AI tests.

The initial study examined more than 300 previously completed Upwork assignments across fields including writing, marketing, consulting, data science, engineering, administrative support and software development. The company deliberately selected relatively straightforward, clearly defined projects where AI agents would have a realistic opportunity to succeed. Even with that advantage, agents operating independently struggled with many tasks. Upwork found that introducing experienced human professionals to review outputs and provide additional direction could increase completion rates by as much as 70% compared with agents working alone.

The result supports a very different interpretation of AI productivity from the idea that machines simply replace workers. A substantial part of the value may instead come from reducing the amount of human time required to produce an acceptable result while preserving human judgement at critical points. The reason becomes clearer when examining how knowledge work is specified. Clients frequently do not provide perfect instructions because they do not know precisely what they want until they see an initial result. A designer may need to show several versions before a customer can articulate what feels wrong. A marketing strategy can require multiple discussions before the true commercial objective becomes apparent.

Humans handle this ambiguity through experience, intuition and conversation. Current AI systems remain much less reliable when requirements are incomplete, contradictory or dependent on unstated preferences. The problem becomes greater when assignments involve several forms of information simultaneously. A digital project might combine written content, visual design, software, customer experience, marketing and business strategy. AI systems can perform individual parts increasingly well but still struggle when responsibility extends across a long sequence of interdependent decisions.

This suggests that the future division between people and AI may not follow traditional job titles. Instead, individual occupations could be broken into components, with machines handling some portions while people become increasingly responsible for direction, judgement, verification and unusual cases. That would significantly change the structure of freelance work.

Lower-value production tasks are likely to experience substantial pressure. If an AI tool can create a basic logo, translate routine material or prepare an elementary summary within seconds, businesses will have progressively less reason to pay a professional to perform the entire process manually. But reducing the cost of those individual tasks does not necessarily reduce the overall amount of work companies undertake.

The history of technology contains repeated examples of falling production costs increasing demand. Cheaper computing did not eliminate software development. It dramatically expanded the number of applications businesses could afford to create. More capable AI could have a similar effect on digital work by making projects economically feasible that companies would previously have considered too expensive. Personalised software provides one example. Historically, organisations usually purchased standard products because developing custom applications for individual teams or users was uneconomic. AI-assisted development could substantially lower those costs, creating demand for far more specialised software than the traditional technology industry could realistically produce.

The same logic could apply across marketing, design, research, consulting and other professional services. If AI reduces the cost of delivering a project, businesses may choose to undertake substantially more projects rather than simply spending less on the work they already commission. This is where Upwork sees another potentially important change emerging. AI agents may become customers of human labour as well as competitors with it.

The conventional marketplace has humans on both sides: a person or company posts a requirement and another person accepts the assignment. In an agentic economy, the party deciding that human expertise is needed could increasingly be software. An AI assistant attempting to solve a complicated problem may reach a stage at which it cannot confidently evaluate its own output. Instead of simply failing or returning an uncertain answer, it could obtain specialist human assistance.

Upwork has already begun connecting its marketplace directly into major AI environments. Businesses can discover professionals through AI interfaces, while integrations are expanding the ability to reach human expertise without leaving the software environment in which the task began. The longer-term implication is more significant than simply making online recruitment more convenient. AI agents could eventually identify when they require external human judgement and locate an appropriate expert themselves.

A legal research agent could request specialist review. A software development agent could obtain a security assessment. A marketing agent could ask a human strategist to evaluate several campaign concepts. A design system could request aesthetic judgement before delivering its final recommendation. This creates the possibility of agentic hiring: machines purchasing small units of human expertise whenever their own capabilities become insufficient.

If this develops at scale, the traditional freelance project could begin to fragment. Instead of one professional spending days completing an entire assignment, experienced workers may increasingly provide shorter, higher-value interventions across many AI-managed workflows. Their economic role shifts from producing every element to resolving ambiguity, providing expert judgement and verifying that machine-generated outcomes are acceptable. A specialist might consequently participate in many projects simultaneously rather than working continuously on one assignment.

This could increase the premium attached to genuine expertise while reducing the value of routine production skills. Knowing how to produce a first draft becomes less valuable when machines can generate drafts instantly. Knowing why the draft is wrong, how it should change and whether the result solves the customer’s actual problem becomes more important.

For employers, this has potentially profound implications. Companies have traditionally organised work around jobs and departments. AI systems may increasingly organise it around outcomes and tasks. A project can be decomposed into components before resources are assigned dynamically according to the capabilities required. That model could alter corporate workforce planning. Organisations might retain smaller groups of permanent employees responsible for strategy, relationships, judgement and institutional knowledge while combining them with AI agents and externally sourced specialists brought into workflows as required.

The boundary between employee, freelancer, software supplier and AI agent could therefore become less distinct. This is not necessarily a simple reduction in labour. It represents a change in how labour is assembled.

For professional-services businesses, the implications are particularly important. Consulting, legal, accounting, marketing, engineering and other knowledge-intensive companies have historically linked revenue partly to the number of professional hours required to deliver work. AI threatens that model because the production time behind many deliverables is falling rapidly. Clients may increasingly become unwilling to pay traditional fees for tasks they know can be partially automated. Professional firms will therefore need to demonstrate value through expertise, accountability, judgement and outcomes rather than the amount of human effort involved.

Freelancers face a similar transition. Competing against AI on tasks that machines can perform cheaply is unlikely to be sustainable. The stronger position may be using AI to increase personal capacity while developing expertise in areas where customers still require human responsibility and judgement.

That does not remove concerns about the labour market. Greater productivity can increase expectations. If professionals become capable of producing several times as much work, businesses may simply demand more output rather than allowing people to work less. AI agents also operate continuously, potentially creating a workplace where projects remain active around the clock and employees feel pressure to supervise automated systems long after traditional working hours. The growth of gig and project-based employment creates additional concerns around income stability, employment protections and the increasingly fragmented nature of professional careers.

Technology-driven labour transitions rarely develop smoothly. Companies overreact, reverse decisions and experiment with organisational structures before more stable models emerge. Customer service provides a visible example. Some companies initially predicted that generative AI would dramatically reduce the need for human support teams, only to discover that automated systems could not reliably handle every customer situation. The resulting model increasingly combines automation for straightforward interactions with people handling complex, emotional or unusual cases.

Knowledge work may follow a similar trajectory. AI is likely to eliminate particular tasks much faster than it eliminates entire professions. The important question then becomes what happens to the remaining human work and whether workers can move towards the higher-value components quickly enough. Upwork is effectively betting that they can.

The strategy also suggests why marketplaces with proprietary data may occupy an unexpectedly important position in the AI economy. Large language models have access to enormous amounts of general information, but they have much less knowledge about the commercial process through which real work is commissioned, evaluated and accepted. Platforms that possess this information can potentially teach AI not simply how to generate an output but how professional work is structured and what customers consider successful.

Upwork says Uma is increasingly embedded throughout the process of defining projects, assessing potential talent, starting contracts and monitoring work. The marketplace is consequently becoming more deeply involved in the workflow itself rather than simply introducing buyers and sellers. That distinction is strategically important. Digital marketplaces traditionally make money by facilitating transactions. An AI-driven work platform could participate much earlier, helping determine what needs to be done, how the project should be structured, which resources should perform it and whether the result is acceptable.

If successful, the platform moves from matchmaking towards orchestration. For businesses, that could reduce one of the largest inefficiencies associated with external talent: managers frequently know the outcome they need but lack sufficient technical knowledge to define the exact combination of specialists required to achieve it. An intelligent orchestration layer could potentially bridge that gap.

The result could also broaden access to professional expertise. Small businesses that cannot employ permanent teams of designers, developers, analysts and marketing specialists could combine AI systems with independent professionals on demand. The cost of accessing sophisticated capabilities could therefore decline substantially. This has implications far beyond the freelance economy. The same model could eventually influence how companies organise project teams, outsource specialist work and manage temporary expertise across almost every knowledge-intensive industry.

It also challenges one of the most common assumptions surrounding AI and employment. The debate is frequently framed as a contest between human workers and machines. In practice, the emerging labour market may be considerably more complicated. AI systems can replace certain tasks, enhance the productivity of workers, create new categories of work and simultaneously become customers purchasing human expertise.

That final development may prove especially important. An internet populated by autonomous agents will still encounter situations that require judgement, verification and responsibility. If those agents can locate and purchase human expertise dynamically, people could become an on-demand layer within machine-driven workflows rather than simply being displaced by them. The economic value of human work would then move progressively away from producing routine outputs and towards deciding what should be done, resolving ambiguity and determining whether the result is acceptable.

The transition is unlikely to be painless. Some categories of work will shrink, skills will become obsolete and companies will experiment aggressively with replacing labour. Productivity gains could also place new pressure on workers as expectations increase. Yet the evidence emerging from real commercial assignments suggests that current AI agents still perform very differently when left alone compared with when knowledgeable professionals guide them.

For companies, the strategic lesson may therefore be less about choosing between employees and AI and more about redesigning work so that each performs the functions where it creates the greatest value. The organisations that succeed may not be those that automate the largest number of jobs. They may be those that become best at dividing work between machines and people.

Upwork’s transformation provides an early example of what that could look like. The company that once built its business connecting businesses with freelancers is increasingly preparing for a world where the customer, the worker and the manager of the project may each be some combination of human and machine. If that model spreads, AI will not simply change how people work. It will change the architecture through which work is commissioned, organised, delivered and valued.

Source: CIJ.World Research & Analysis Team

AI Is Turning Corporate Governance Into a Continuous Management System

Artificial intelligence may be forcing companies to rethink not only their technology infrastructure but the way boards and senior management govern risk itself. Traditional oversight structures built around committees, periodic reports and relatively static risk assessments can struggle when AI applications are introduced, modified and expanded across an organisation far faster than conventional governance processes can respond. That was the central argument presented at Ai4 2026 in Las Vegas by Brian Allen, Executive Director of the Center for AI Oversight, who argued that organisations should stop thinking primarily about governing AI technology and instead govern the management system surrounding its use.

The distinction is important. A board does not need to understand every model, agent or technical architecture operating inside the business. It needs confidence that management has established a repeatable process for deciding where AI can be used, what risks the organisation is prepared to accept, when problems or missed opportunities must be escalated, and whether the expected economic benefits justify the exposure. Allen framed this as part technology revolution and part management revolution, arguing that companies are introducing a rapidly changing technology into governance structures designed for a much slower environment.

This also changes the relationship between risk and strategy. Traditional cybersecurity discussions have often concentrated on downside: preventing breaches, protecting information and avoiding disruption. AI creates another dimension because moving too slowly can itself become a competitive risk. A company that blocks useful automation while competitors reduce costs, improve margins or deliver services faster may be accepting a different kind of exposure. Governance therefore needs to balance the danger of deploying AI too aggressively against the economic cost of not deploying it quickly enough.

Quarterly reporting illustrates the problem. A board may receive an AI risk update every three months, yet employees and business units can introduce new models, applications and agents in a fraction of that time. By the time a traditional report reaches senior management, the company’s actual AI environment may already have changed materially. Allen therefore challenged what he called the assumption that having an AI committee automatically means an organisation has effective governance. A committee is a governing body, but the more important question is what management system that body is actually governing.

If the committee’s main role is simply to receive periodic information about projects and risks, executives may be well informed without actively governing how the organisation balances AI opportunity and exposure. Effective oversight instead requires a mandate defining responsibility, authority, risk tolerance, escalation procedures and the evidence necessary to determine whether management is operating within those expectations. This broadly reflects the direction of established frameworks such as the NIST AI Risk Management Framework, which treats governance as something extending across the entire lifecycle of an AI system rather than as a one-time technical approval.

Risk tolerance is central to this approach because it gives management something concrete against which investment and behaviour can be measured. Broad statements that a company has a low, moderate or high appetite for AI risk provide limited practical guidance. A more useful model is to establish tolerances around particular applications and consequences. An internal productivity assistant may justify a different tolerance from an AI system influencing lending decisions, recruitment, pricing, customer data or safety-critical infrastructure.

Those tolerances should function as decision points rather than automatic stop signs. If an AI application moves beyond an agreed boundary, management can decide whether to reduce the risk, introduce additional controls, accept greater exposure or discontinue the activity. The same principle can work in the opposite direction. If controls are preventing a low-risk AI initiative from generating measurable savings or competitive benefits, that missed opportunity can also be escalated. Governance then becomes a mechanism for balancing risk and return rather than merely stopping activity.

That makes AI oversight closely connected to resource allocation. Once a company understands the economic importance of a particular use case and the exposure associated with it, management can decide how much investment in security, data quality, monitoring, human oversight and resilience is justified. The objective is not zero risk, which is unrealistic in any business environment. It is to determine how much risk the company is prepared to accept in exchange for a particular strategic benefit.

Allen proposed an oversight model built around several interconnected areas: governance responsibility, a risk-informed decision system, trust and assurance, risk-based strategy, and escalation and disclosure. The precise structure will differ between organisations, but the underlying principle is that these elements need to operate as one programme rather than as a collection of disconnected activities. Many large companies already have much of the required infrastructure through legal, compliance, cybersecurity, audit, data governance and enterprise-risk teams. The challenge is connecting those functions into a repeatable system that can make AI decisions consistently and quickly.

Clear ownership is therefore essential. Companies need to know who has the authority to establish the AI governance programme, who operates it and which matters must reach executive management or the board. Formal mandates and documented responsibilities become especially important when difficult decisions arise because employees cannot rely solely on informal relationships or assumptions that information will automatically move upwards through management. A codified governance structure provides something the organisation can lean on when business pressures become intense.

The board’s role should nevertheless remain different from management’s. Directors do not need to operate AI controls themselves. Their responsibility is oversight: understanding whether an appropriate system exists, whether management is executing it and whether significant risks and opportunities are reaching the board. Operational teams then determine how technical controls, cybersecurity measures, data processes and assessments are implemented. This separation allows directors to remain engaged without attempting to manage technology they are not equipped to operate.

Corporate law provides important context to that discussion, although it should not be interpreted as establishing a specific AI governance standard. Delaware’s Caremark line of cases has progressively examined directors’ and, more recently, corporate officers’ responsibilities for oversight and reporting systems, particularly where risks are central to a company’s operations. The lesson for AI is not that boards automatically face liability whenever an AI system fails. Rather, as AI becomes material to business operations, directors and executives may increasingly need to demonstrate that appropriate information and escalation mechanisms existed and that significant warning signs were not ignored.

This creates another reason for companies to document how AI decisions are made. Governance should produce evidence showing what management knew, which risks were considered, what tolerance was approved and why a particular decision was taken. The standard should not be that every AI decision must ultimately prove correct. Business involves uncertainty. A more realistic objective is demonstrating that material decisions were informed, appropriately authorised and consistent with the organisation’s established governance process.

Explainability illustrates the distinction. AI models are frequently described as black boxes because it can be difficult to reconstruct precisely how a complex model produced an output. That does not necessarily mean a company cannot provide meaningful accountability. Management can document what information entered the system, what it was intended to accomplish, which controls applied, what output was produced, how that output was validated and why the organisation considered it reliable enough for the intended purpose. The emphasis shifts from trying to expose every internal mathematical operation towards demonstrating that the overall decision process is controlled and defensible.

The same principle applies to AI strategy more broadly. Companies cannot know precisely how models, regulation and competitive pressures will evolve over the next several years. Governance therefore needs to accommodate uncertainty rather than pretending it can eliminate it. Organisations should establish boundaries that permit experimentation while creating clear mechanisms for escalating situations where financial, legal, operational or reputational exposure becomes significant.

For property companies and investors, these issues are likely to become increasingly relevant as AI enters investment analysis, valuation, leasing, development, building management, procurement and financing. An investment manager may use AI to screen acquisitions, a lender to analyse borrowers, a developer to assess sites and a property manager to automate tenant interactions. Each application creates a different balance between economic opportunity and risk, which means a single company-wide rule is unlikely to be appropriate for every use case.

Commercial real estate also demonstrates why governance needs to extend beyond the technology department. AI-generated investment analysis may depend on data controlled by asset managers and finance teams. Automated lease analysis involves legal departments. Building-management AI can affect operations and safety, while customer-facing systems create privacy and reputational considerations. Effective oversight therefore requires technology specialists to work with the people responsible for the underlying business processes rather than treating AI as an isolated IT function.

There is also a longer-term organisational issue. As companies automate more analytical and administrative work, they need to consider what happens to the institutional knowledge traditionally developed by employees performing those tasks. Junior professionals often learn by completing repetitive work, observing experienced colleagues and gradually understanding exceptions that are not obvious from formal procedures. If AI removes much of that work, businesses may need alternative ways to develop the judgement required for future senior positions.

This makes AI governance partly a question of organisational design. Companies are not simply deciding which software to buy. They are deciding which decisions remain with people, which can be delegated to machines, what knowledge must remain inside the organisation and how accountability operates when human and automated decision-making become intertwined.

The organisations best positioned for this transition may therefore be those that treat governance as an accelerator rather than a brake. A well-defined programme can allow employees to experiment faster because boundaries, responsibilities and escalation routes are already understood. Instead of requiring every new AI project to begin a fresh debate about risk, management can evaluate it against an established system and make decisions more quickly.

That may ultimately be the most important distinction between traditional technology governance and the emerging AI model. Companies cannot realistically govern every technological change individually at the speed at which AI is developing. What they can govern is the system through which those changes are evaluated and adopted.

For boards, the question is therefore moving away from whether the company has an AI committee or policy. The more important questions are whether management knows which risks it is willing to take, whether those boundaries can adapt as circumstances change, whether opportunities as well as threats are escalated and whether there is evidence that the organisation is actually operating according to those decisions.

AI governance at that level becomes much closer to strategy than compliance. It is a mechanism for deciding how quickly an organisation wants to move, where it is prepared to take risk and where stronger controls are necessary. As AI spreads across corporate operations, the ability to make those decisions repeatedly, transparently and at speed may become as important as the technology itself.

Source: CIJ.World Research & Analysis Team

Manufacturers Are Finding That the Best AI Strategy Starts Without AI

Artificial intelligence is spreading rapidly across manufacturing, but some of the companies deploying it most aggressively are discovering that the best place to start is not with the technology. Instead, manufacturers are identifying costly operational problems, simplifying the underlying process and involving factory workers before deciding whether AI is necessary at all. That was one of the clearest conclusions from a manufacturing panel at Ai4 2026 in Las Vegas, where executives discussed applications ranging from equipment maintenance and computer vision to production logistics and internal software development. Rather than describing a future dominated by fully autonomous factories, the discussion showed how AI is being introduced incrementally into existing industrial processes where downtime, quality problems, inventory, maintenance and inefficient workflows can be measured financially.

One example comes from Viaduct, the Silicon Valley AI company acquired by Japan’s Sumitomo Rubber Industries in 2025. Viaduct CEO David Hallac described the deployment of an AI-assisted maintenance system at Sumitomo Rubber’s Miyazaki tyre factory in Japan, where equipment failures could disrupt production while maintenance technicians searched historical records for solutions. The system brings together years of maintenance information and recommends potential repairs when problems occur, effectively turning accumulated factory knowledge into a resource that technicians can access during maintenance work. Viaduct says the implementation has reduced repair time significantly, while the broader significance extends beyond a single factory because Sumitomo Rubber sees predictive maintenance and AI-driven asset management as capabilities that can eventually be deployed across wider manufacturing operations.

The industrial value of such systems comes partly from preserving knowledge that traditionally remains inside experienced employees’ heads. Manufacturing plants frequently depend on technicians who have spent decades learning how individual machines behave, which sounds indicate developing problems and which sequence of events usually precedes a failure. Some of that information exists in maintenance systems, but much of it can be contained in shorthand notes, fragmented databases or individual experience. AI potentially changes that relationship by combining maintenance histories, sensor information, manuals, images, repair records and employee knowledge into systems that can identify recurring patterns. A technician receiving an alert can then be shown not simply that a component may fail, but which previous cases displayed similar behaviour and what actions resolved them.

That type of evidence is important because industrial workers are unlikely to trust a system simply because an algorithm tells them something is wrong. An experienced technician may reasonably ignore a warning if a machine appears to be operating normally. A system becomes more persuasive when it can demonstrate that several comparable machines showed the same combination of signals before a particular failure occurred. The technology then supplements the technician’s judgement rather than asking the employee to surrender that judgement to a black box.

Kaspar Companies, a family-owned Texas industrial group with businesses including truck beds, outdoor products and precious metals, offered a different example. The company has been experimenting rapidly with internally developed applications, but Peter Robinson said the organisation initially made the mistake of looking for places to deploy AI rather than beginning with operational problems. The company responded by reconnecting its AI programme with its existing continuous-improvement system. Factory teams identify a specific problem, examine the existing workflow and remove unnecessary steps before considering a technological solution. Employees performing the work participate directly in that process, meaning the people eventually expected to use the application are involved before it is developed.

One application involves computer vision at Kaspar’s Bedrock truck-bed operation. Different truck beds require different hardware packages when they are delivered to distributors for installation. Missing or incorrect components can cause problems after the original truck bed has already been removed. The company developed a camera-based system that identifies the truck bed and checks whether the correct hardware has been included before it leaves production. Another project began with excess finished inventory. Customers order combinations of truck beds that need to fit onto outbound trucks, but different product dimensions made it difficult to determine the optimal load in advance. Kaspar developed a system that helps optimise the combination before production and shipment.

An unexpected commercial benefit emerged from that second project. The same system could identify situations where additional truck beds could fit onto an order without requiring another vehicle. Customer-service employees could then offer customers the opportunity to add products to the shipment, turning a project initially intended to reduce inventory into a potential sales tool. That example illustrates why the return from manufacturing AI does not always appear where companies initially expect it. A project designed to reduce waste might increase revenue, while predictive maintenance could improve quality as well as reduce downtime. Better production information can also influence warranty costs, customer satisfaction or inventory requirements elsewhere in the organisation.

It also demonstrates why manufacturers need to remain close to factory operations when evaluating results. Some secondary benefits are difficult to predict from an initial business case and may only become visible after operators begin using the system. AI development therefore cannot be separated entirely from the physical environment in which the technology operates. Engineers need feedback from the people running machines, repairing equipment, loading trucks and inspecting products. The panel repeatedly returned to the idea that the most successful projects tend to be those built around an obvious operational pain point rather than around a desire to showcase AI.

The discussion also challenged the assumption that every industrial problem requires artificial intelligence. Kaspar said some of its most valuable internally developed solutions do not contain AI in their final operation, even though AI tools may have helped developers build them faster. The relevant question is therefore not whether a manufacturer can attach AI to a process, but what is the simplest reliable technology capable of solving the problem. That distinction is becoming increasingly important as generative AI makes software development faster and cheaper.

Manufacturers can potentially build specialised internal tools that previously would have required considerably more engineering resources. AI therefore has two roles in industrial transformation: it can become part of the application operating on the factory floor, or it can simply help engineers develop conventional software much more quickly. This could alter the economics of industrial software because manufacturers have historically purchased broad enterprise systems and adapted their processes around them partly because developing specialised applications internally was expensive. AI-assisted development lowers that barrier and may make it easier for small technical teams to create tools around individual production, inventory, quality or maintenance problems.

Kaspar’s approach illustrates this change. Rather than assembling a large AI department, the company has maintained a small development team and pushed it to experiment rapidly. Robinson said dozens of projects have been attempted, while acknowledging that a significant proportion have not generated measurable value. The failures are treated partly as learning exercises, with unsuccessful applications stopped rather than continually receiving resources simply because development has already begun. That willingness to abandon projects could become one of the more important disciplines in corporate AI adoption.

The technology is developing so quickly that an application which is difficult or uneconomic today may become straightforward several months later. Manufacturers therefore need to distinguish between problems worth solving immediately and projects that should be paused until the technology improves or the business is ready to adopt them. Speed still matters once the correct problem has been identified, but the panel’s message was that companies should spend considerable effort understanding the problem before development begins and then move quickly during execution.

Workforce strategy is equally important. Fear that AI will eliminate manufacturing jobs can undermine adoption before a system reaches the factory floor. Kaspar has attempted to connect its AI strategy with an existing policy of moving employees affected by productivity improvements into other roles rather than treating operational improvement primarily as a headcount-reduction programme. Employees can therefore see automation as part of continuous improvement and capacity expansion rather than an immediate threat to employment. Whether every manufacturer can adopt that approach will depend on its economics, workforce and competitive pressures, but the broader lesson remains that operators are more likely to support systems they helped design and that solve problems they experience personally.

This human factor may explain why the idea of the fully autonomous smart factory remains less important in practice than hundreds of smaller improvements. AI can help technicians repair equipment faster, identify defects, optimise shipments, retrieve production information and build software, but these applications still depend heavily on people understanding the process and deciding what outcome is valuable. Technology imposed from a central office without sufficient understanding of factory work is more likely to encounter resistance than a system developed together with the people expected to use it.

Robotics could extend that transformation further. Robinson expects physical automation and humanoid robots to become a much larger part of the manufacturing discussion as AI capabilities improve. That would move AI from analysing and coordinating industrial processes towards physically performing more tasks inside factories, potentially creating another major investment cycle in production facilities. The implications extend beyond manufacturing technology itself because factories adopting more robotics, computer vision, predictive maintenance and AI-driven production systems will require different digital infrastructure, connectivity, power capacity and technical skills.

Industrial property capable of supporting advanced automation may therefore become increasingly differentiated from older facilities that lack the electrical, data and physical infrastructure required for modern production. The manufacturing AI market may also become increasingly specialised, with broad horizontal AI platforms consolidating around a relatively small group of large technology providers while specialist companies concentrate on individual industrial problems. Manufacturers may consequently rely on common underlying models while building or purchasing specialised applications for maintenance, quality control, production planning, logistics and other operational functions.

The strongest lesson from the panel, however, was considerably simpler. AI should not become the starting point for industrial transformation. Manufacturers first need to understand where production is losing time, money, capacity or quality. They then need to simplify the process, involve the people performing the work and choose the simplest technology capable of solving the problem. Sometimes that solution will involve sophisticated predictive models, computer vision or AI agents. Sometimes AI will merely help engineers build ordinary software faster. And sometimes the correct answer will contain no artificial intelligence at all.

For manufacturers under pressure to demonstrate that their AI investments produce real economic value, that distinction could become increasingly important. The smartest factory may ultimately not be the one using the most AI. It may be the one that has become best at identifying exactly where AI is worth using.

Source: CIJ.World Research & Analysis Team

Banks Are Discovering That AI Value Depends Less on the Model Than on How the Business Uses It

The banking industry’s artificial-intelligence debate is moving from how much money institutions are investing towards a more difficult question: where is the return actually appearing? At Ai4 2026 in Las Vegas, BMO Financial Group’s Chief AI & Quantum Officer Kristin Milchanowski argued that AI is entering a new stage in financial services in which access to technology itself will become less differentiating. As banks increasingly obtain similar foundation models and cloud infrastructure, competitive advantage will depend more heavily on how effectively those capabilities are embedded into business processes, client relationships and decision-making.

That represents an important change from the first phase of generative AI adoption. Banks spent heavily on experimentation, employee tools, pilots and infrastructure, but a successful demonstration is not the same as an application capable of producing repeatable value across a large institution. Milchanowski described one of BMO’s central tests as whether AI has been connected to a workflow that genuinely matters to the business. If employees constantly need to be reminded that a particular system is using AI, the technology may not yet be sufficiently integrated. The objective is for AI increasingly to disappear into everyday banking processes rather than remain a separate product employees or customers have to consciously operate.

BMO has formalised much of this strategy through its Institute for Applied Artificial Intelligence & Quantum, created in 2026 as an enterprise-wide centre covering implementation, governance and emerging quantum capabilities. Milchanowski’s role therefore combines two responsibilities that are sometimes separated inside large organisations: building technology and establishing the framework under which that technology can operate. The economic challenge is that most banks do not have an accounting line showing AI revenue or AI profit. AI contributes to other activities, meaning the return has to be identified inside broader business outcomes such as faster credit processing, stronger client engagement, improved win rates or lower operating costs.

That makes baseline measurement critical. A bank needs to know how long a process took, what it cost and how effectively it performed before AI was introduced. Without that starting point, it becomes difficult to demonstrate whether the technology actually created value. Milchanowski described BMO’s work in commercial credit as one example, where generative AI is being used to collect information and help bankers prepare credit analysis more quickly. The bank has also developed Aura, a GenAI-powered assistant supporting commercial underwriting by retrieving information from approved systems and helping prepare parts of credit submissions while human underwriters remain responsible for the final assessment.

The underlying challenge is familiar across financial services. Large banks have accumulated technology over decades, leaving employees working across numerous applications and databases to complete a single task. A banker or financial-crime specialist may need to search several systems, review multiple documents and assemble the results into a memo before making a decision. Generative AI and agents can increasingly act as an information layer across those fragmented environments, retrieving relevant material and bringing it together for the employee. The human may continue making the consequential decision, but considerably less time is spent locating, comparing and reorganising information.

This is why many of the most credible banking applications currently involve document interpretation, information retrieval and the restructuring of content from large bodies of unstructured material. AI can locate clauses in legal documents, identify risk factors, compare information between reports or assemble material scattered across internal systems. These applications may appear less dramatic than autonomous banking, but they address workflows that consume significant employee time across lending, compliance, legal, operations and risk and can therefore produce measurable operational value.

The larger economic opportunity, however, may not come from efficiency alone. Milchanowski cautioned against treating AI as another cost-reduction programme. A strategy concentrated exclusively on doing the same amount of work with fewer resources could overlook the greater opportunity to generate revenue. Her argument was that banks should concentrate on improving client experience, increasing adoption and strengthening relationships, with efficiency emerging partly as a consequence of better-designed processes rather than becoming the sole objective.

That distinction matters because efficiency gains can quickly become part of the industry’s new cost base. If every major bank reduces the cost of producing a credit memo or answering a routine customer question, competitors will eventually be forced to make similar investments. The more durable advantage could come from using AI to improve how quickly institutions identify opportunities, respond to clients and make decisions. Milchanowski described this as decision velocity: reducing the administrative distance between information and the people responsible for acting on it.

Large organisations still rely heavily on periodic reporting processes in which information moves through multiple departments before senior management receives an aggregated picture. Forecasting can require several teams to prepare spreadsheets, reconcile assumptions and roll information upwards before a final view reaches leadership. If AI can continuously access approved information from core systems and assemble it into useful analysis, management could move towards a more continuous operating model. The value would not come from allowing AI to determine strategy independently, but from enabling people to make informed decisions sooner.

Before technology is introduced, however, BMO’s approach places responsibility on business teams to simplify existing processes. Milchanowski’s warning was straightforward: automating a poorly designed workflow simply allows the organisation to perform unnecessary work more efficiently. Banks therefore need to remove obsolete approvals, duplicated steps and bottlenecks before applying AI. Technology should then be used to redesign the remaining workflow rather than reproduce decades of accumulated complexity in automated form.

This principle also affects how AI programmes are organised and funded. Instead of building hundreds of unrelated applications, central technology teams can develop reusable capabilities that individual business lines adapt for their own products and workflows. BMO’s model combines enterprise-wide technology and governance with authority retained by the individual profit-and-loss businesses, which remain responsible for identifying client needs and allocating investment. Common infrastructure can therefore be developed centrally without requiring every business unit to create its own AI platform.

Governance becomes particularly important as banks move from assistants towards agents capable of performing actions. Milchanowski identified identity, orchestration and accountability as some of the main challenges. An institution needs to know who created an agent, why it exists, what information it used, which systems it may access and who is responsible for the results. As multiple agents begin interacting, banks will also need mechanisms determining which systems can communicate and under what circumstances.

Well-designed governance can eventually accelerate deployment rather than simply slow it. When responsibilities, thresholds and escalation procedures have already been defined, teams do not need to reconsider the same fundamental questions every time a new application is proposed. The institution can move more quickly because the boundaries are clearer. This also means employees need sufficient AI literacy to understand how tools are being used and remain accountable for decisions made with their assistance.

The technology strategy reflects a similar desire to avoid adding another fragmented layer to an already complicated banking architecture. Rather than buying large numbers of disconnected applications, AI increasingly needs to sit inside established infrastructure and data environments if banks want to improve decision speed across the organisation. Buying a platform does not eliminate development work because the technology still needs to be integrated with the bank’s own data, controls and operating processes.

A related shift may eventually favour smaller and more specialised language models. Milchanowski predicted that banks could increasingly use smaller models where their capabilities are sufficient because running the largest available model for every task can make little economic sense. A bank may need advanced reasoning for some activities but not for routine extraction, classification or internal Q&A. The return on AI can therefore improve partly through matching model capability and cost to the complexity of individual workloads.

Quantum computing occupies a much earlier position in this development cycle. BMO is researching quantum applications through its work with IBM and has explored areas including optimisation, risk and environmental modelling. One example involves earthquake forecasting, where the bank is researching ways to process extremely large and complex datasets. The broader logic is not that earthquake prediction will suddenly become a banking product, but that methods developed in demanding scientific problems could eventually have applications in capital markets and financial risk.

For financial institutions, the important contrast is therefore between today’s AI investment cycle and the much earlier quantum research cycle. Banks are already being asked to demonstrate measurable returns from AI, while quantum investments can still reasonably be justified through research, readiness and long-term capability building. Confusing those stages could lead organisations either to demand commercial returns too early from quantum research or to accept experimentation for too long in mature AI programmes.

The broader message from the Ai4 discussion is that banking’s AI race may ultimately become less about access to technology. Major financial institutions can increasingly obtain similar models, cloud platforms and development tools. What they cannot purchase as easily is the operating discipline required to redesign workflows, integrate proprietary data, establish governance, train employees and connect AI investment to commercially meaningful outcomes.

That is why AI may increasingly become a margin and execution issue rather than simply a technology issue. Banks using the same underlying models could still produce very different financial results depending on how those models are deployed. The winners may not be the institutions with the largest number of AI pilots, the highest token consumption or even the most advanced individual models. They are more likely to be the banks capable of turning intelligence into faster decisions, stronger client relationships and repeatable economic value across the organisation.

Source: CIJ.World Research & Analysis Team

AI Agents Are Moving From Experiments to Business Operations – but Data Still Determines Who Succeed

Artificial intelligence agents are beginning to move beyond experimental applications and into the everyday operations of large companies, but some of the organisations deploying them at scale are finding that the technology itself is only one part of the challenge. Data quality, employee trust, cost control and the ability to connect AI with measurable business outcomes are emerging as equally important factors determining whether projects move successfully from pilot programmes into production. That was the central message from an Ai4 2026 panel in Las Vegas bringing together executives working with AI across pharmaceuticals, healthcare, customer experience and digital businesses.

The discussion included Nitesh Soni of Sanofi, Anuj Maheshwari of CVS Health, Chris Han of Thinking AI and Kevin Lee of NiCE, with each describing how agentic systems are beginning to alter established business processes. The examples differed substantially, but a common pattern emerged. Successful projects were generally not being designed around the question of where a company could deploy an AI agent. They began instead with a business process that was slow, expensive or difficult to scale and then examined whether AI could improve the outcome.

At Sanofi, Soni described the development of a conversational data platform intended to give commercial teams faster access to information that previously required analysts to search through multiple dashboards and datasets. A brand manager or sales team might want to understand how a product is performing in a particular territory, where opportunities are emerging or where resources should be redirected. Traditionally, those questions can require analysts to combine information from several systems before producing an answer. The emerging model uses a network of specialised agents behind a conversational interface. According to Soni, the architecture currently involves roughly 10 to 12 agents, some performing general orchestration and others handling specific tasks. A user asks a question in ordinary language, an orchestration layer interprets the request and the appropriate agents retrieve and combine the necessary information.

The objective is not simply to generate an answer. The system can potentially suggest follow-up questions and recommend possible next actions, allowing employees without specialist analytics skills to interrogate enterprise information directly. This represents an important change from the traditional business-intelligence model, where organisations build dashboards and expect employees to understand where the relevant information sits and how it should be interpreted.

The CVS Health example takes a different approach. Maheshwari described what the company calls agentic twins, which are intended to simulate how different types of patients might respond to healthcare programmes before those programmes are introduced more widely. He said CVS Health has created hundreds of thousands of simulated agent profiles to test possible consumer responses while protecting patient privacy. The concept resembles a flight simulator for business decisions. Instead of launching a programme directly into the market and then discovering how patients react, an organisation can test different approaches in a virtual environment first.

Maheshwari connected the work to a longstanding healthcare challenge: patients who either do not collect prescribed medication or stop treatment earlier than intended. The purpose of the simulations is to understand the different barriers influencing those decisions and test whether alternative communications or programmes might improve outcomes before exposing real patients to the intervention. The significance extends beyond healthcare. Digital twins are already familiar in manufacturing, buildings and infrastructure, where companies create virtual representations of physical assets. Agentic AI introduces the possibility of creating behavioural simulations involving customers, employees or other groups, allowing companies to test decisions before committing resources in the physical marketplace.

For businesses, this could change the economics of experimentation. Marketing programmes, customer experiences, pricing strategies and operational changes could potentially be tested against large synthetic populations before companies launch them. The technology is still developing, and simulated behaviour cannot automatically be assumed to reproduce real human decisions accurately, but the approach demonstrates how agentic AI could move beyond task automation into decision modelling.

NiCE’s Kevin Lee approached the issue from the customer-experience perspective. His argument was that companies frequently begin with the technology rather than examining what customers actually need. A business may decide that a generative-AI model should solve a particular problem even when a conventional rules-based system or established automation could perform the task more reliably and at lower cost. That distinction becomes increasingly important as companies begin paying for AI usage at scale. A simple request such as checking the status of an order may not require an advanced generative model, while more complicated conversations involving several intentions, changing context or unstructured information may justify a more capable system.

The emerging enterprise architecture is therefore unlikely to involve sending every problem to the largest available AI model. Companies are instead beginning to route different tasks to different technologies depending on complexity, cost and risk. Some processes may use conventional software, others smaller AI models, while difficult reasoning tasks are sent to more powerful systems. An orchestration layer determines which approach is appropriate. This could become one of the most important economic questions surrounding enterprise AI because individual interactions with models can appear inexpensive, but the cost becomes considerably more significant when millions of customers, employees or automated agents are generating requests continuously.

The objective therefore changes from maximising the intelligence used for every task to finding the least expensive architecture capable of producing the required outcome reliably. The panel repeatedly returned to one factor underlying all of these systems: data. Soni argued that organisations unable to establish reliable, governed information foundations will struggle to scale agentic AI regardless of how capable the underlying models become. A prototype can sometimes operate on carefully prepared datasets, but enterprise deployment has to work across information created by different departments, systems and historical processes.

Large companies frequently have several versions of the same information stored in different places. Documents may be outdated, databases may use inconsistent definitions and responsibility for maintaining particular information may have disappeared as employees changed roles. AI does not automatically solve those problems. Connecting a powerful model to five conflicting sources does not necessarily produce a better answer. It may simply allow the system to choose between five different versions of reality. This is why the less visible work of cataloguing, governing and maintaining enterprise data is becoming increasingly important to AI adoption.

Sanofi’s approach, as described by Soni, separates information into different layers. Raw information from source systems is first organised before being aligned with business processes and eventually transformed into datasets suitable for analytics and AI applications. AI itself can also help perform some of that preparation, creating a cycle in which better data enables better AI while AI assists companies in improving their data.

Trust becomes the next challenge. An AI system can produce a highly accurate recommendation, but the business value remains limited if employees do not understand the result or refuse to act on it. This makes explainability and transparency important not only for regulatory reasons but for adoption. Employees need to understand where information came from, why a recommendation was produced and whether they remain responsible for the final decision. This is particularly important in healthcare and pharmaceuticals, where incorrect recommendations can have consequences far beyond lost productivity.

Systems therefore need observability, testing and traceability designed into them from the beginning rather than added after deployment. Maheshwari emphasised the need to distinguish between what AI is technically capable of doing and what an organisation should allow it to do. Human intelligence remains essential where judgement, safety or accountability are involved. The principle applies beyond healthcare because companies deploying AI agents increasingly need to decide which actions can be automated, which require human approval and which should not be delegated to AI at all.

Change management is another major part of the equation. Soni argued that business teams need to become involved from the beginning rather than being presented with a finished AI product and expected to adopt it. Continuous feedback from users can improve the technology while also creating internal advocates who understand why the system was built and how it should be used. This suggests that scaling AI may ultimately be as much an organisational problem as a technical one. Companies can purchase access to increasingly capable models relatively easily, but changing how thousands of employees work, how decisions are made and how responsibility is distributed across the organisation is considerably more difficult.

The measurement of success is also becoming more sophisticated. Early AI programmes frequently concentrated on technical accuracy or the number of tasks automated. The panel argued that companies increasingly need to connect those measurements with operational and financial outcomes. One useful metric is time to insight. Maheshwari described analytical work that could previously require several people working for days being reduced dramatically through AI-assisted processes. The relevant question, however, is not simply whether the AI completed the task faster but whether it maintained the required standards of accuracy, safety and reliability.

In pharmaceuticals, time can have even greater significance. Soni argued that reducing the time required across processes ranging from coding and commercial analysis to bringing therapies to patients can produce meaningful value when improvements are multiplied across a global organisation. Customer-experience businesses use different measurements, including satisfaction, successful resolution and the cost of individual interactions, while digital businesses may concentrate more heavily on customer retention and engagement. The common principle is that an AI project should ultimately be measured against the business problem it was supposed to solve.

This addresses one of the central difficulties surrounding corporate AI investment. The rapid development of generative and agentic AI has encouraged companies to experiment widely, sometimes without defining the financial or operational result expected from the project. As enterprise spending increases, that tolerance is likely to decline. Executives will increasingly ask how much a system costs to operate, which processes it improves, how much employee time it saves and whether those improvements ultimately influence revenue, margins, customer retention or another strategic objective.

Cost management is therefore becoming part of AI architecture itself. The panel described an emerging combination of buying, renting and internally developing AI capabilities. Companies may purchase common platforms for widely used functions, access specialist models or tools where capabilities change rapidly and develop their own systems where proprietary knowledge provides a competitive advantage. Maintaining flexibility could become important because the AI market is evolving too quickly for companies to assume that today’s leading model or architecture will remain dominant.

The same principle applies inside individual workflows. Not every task requires generative AI, and not every generative task requires the most sophisticated model available. Matching computational cost to the complexity and economic value of the task could become a basic discipline of enterprise AI management. The wider implication is that AI agents are beginning to change business architecture rather than merely automate individual jobs.

Companies historically organised workflows around humans moving information between software applications. An employee opened a dashboard, interpreted the information, transferred it into another system, made a decision and then initiated the next action. Agentic systems can increasingly connect those stages. One agent can retrieve information, another analyse it, another recommend an action and another execute an approved step, while humans increasingly supervise the workflow, resolve exceptions and make higher-value judgements.

That does not mean every company is close to autonomous operations. The panel repeatedly acknowledged that enterprise deployment remains difficult and that many organisations are still learning how to move projects beyond controlled pilots. But the direction is becoming clearer. The most successful enterprise AI strategies may not be those with the largest number of agents or the most advanced models. They may instead be those with the strongest information foundations, clearest business objectives and best understanding of where machines should operate independently and where humans should remain involved.

For companies investing heavily in AI, this represents an important shift in priorities. The question is moving away from how many AI tools an organisation has deployed towards whether those systems are improving the economics and effectiveness of the business. AI agents can make information easier to access, simulate possible decisions, automate workflows and reduce the time required to analyse complex situations, but none of those capabilities automatically creates value.

The underlying data still needs to be reliable, employees still need to trust the system, costs still need to be controlled and outcomes still need to be measured. As agentic AI moves deeper into large organisations, those fundamentals may ultimately determine which companies turn the technology into a genuine operating advantage and which simply build another layer of expensive software.

Source: CIJ.World Research & Analysis Team

Medicine’s AI Challenge Is Shifting From Adoption to Knowing Where the Machine Should Stop

Artificial intelligence has moved rapidly into everyday medical practice, but healthcare’s next challenge may be considerably harder than getting doctors to use the technology. As AI begins moving beyond documentation and administrative support towards diagnosis, clinical decision support and personalised treatment, physicians and medical schools increasingly need to determine which capabilities can safely be delegated to machines and which skills doctors must continue to develop themselves.

Speaking at Ai4 2026 in Las Vegas, American Medical Association CEO John Whyte said the profession has barely begun to use the full potential of AI. The AMA’s latest physician survey found that more than four in five doctors reported using some form of AI professionally in 2026, more than double the proportion recorded when the organisation began surveying physicians in 2023. Much of today’s activity remains concentrated around functions such as summarising medical information and preparing clinical documentation, but Whyte argued that the larger opportunity lies in identifying patterns, supporting earlier diagnosis and helping clinicians develop more personalised treatment strategies.

The AMA deliberately uses the term augmented intelligence to emphasise what it sees as the appropriate relationship between medicine and technology. AI should improve a physician’s ability to make decisions rather than become an autonomous replacement for clinical judgement. That distinction becomes more important as systems become better at interpreting complex information. Imaging is an obvious example, where algorithms can assist specialists in detecting abnormalities, prioritising cases and identifying patterns that might otherwise be missed. Rather than eliminating radiologists or other specialists, the more likely near-term effect is a redistribution of work in which machines handle more repetitive analysis while physicians concentrate on difficult findings, clinical context and decisions requiring judgement.

This development also challenges one of the traditional foundations of medical education. For generations, doctors were expected to accumulate and retain enormous amounts of information, while AI can now retrieve much of that knowledge almost instantly. The question therefore becomes whether tomorrow’s physicians should be trained primarily to remember information or to interrogate, interpret and challenge the information presented to them. Whyte argued that medical education cannot simply introduce AI from the first day of training without considering what students might fail to learn as a result.

If a student has always relied on an automated system to construct a patient history, suggest a differential diagnosis or interpret a physical finding, that physician may never develop the underlying skill independently. The danger is not necessarily that AI supplies poor information every time, but that doctors become unable to recognise the occasions when the system is wrong. Generative models can produce convincing answers even when the underlying conclusion is incomplete or incorrect, making it essential that medical professionals understand where recommendations came from, assess the quality of the evidence and recognise when a result conflicts with clinical experience or other information about the patient.

Medical schools are consequently beginning to rethink how AI fits into training rather than treating it as another software tool. The AMA has supported initiatives exploring how technology can personalise medical education while preserving the competencies doctors need in clinical practice. The challenge is similar to the arrival of calculators in mathematics: tools can remove repetitive work and improve productivity, but students still need to understand the underlying principles before they can judge whether the answer produced by the technology makes sense.

The same problem is appearing outside hospitals as consumers acquire unprecedented amounts of health information directly from technology. Watches, rings and other connected devices can now provide users with measurements related to sleep, heart rate, oxygen saturation, exercise and numerous other physiological indicators. This represents a significant change from the first generation of consumer trackers that primarily counted steps, but the availability of more data does not necessarily mean patients know what to do with it.

Whyte has warned that the distinction between a wellness product and a medical product can become increasingly difficult for consumers to understand when a device presents information that appears to indicate disease risk. Manufacturers may carefully describe a feature as an indicator or wellness measurement rather than a diagnostic tool, while users may nevertheless change their behaviour because of what the device reports. The regulatory framework is therefore becoming an increasingly important part of the AI and digital-health debate, particularly as consumer technology begins influencing decisions that previously depended on conventional medical testing.

Blood-pressure monitoring illustrates the issue. Continuous or frequent measurements could potentially improve cardiovascular care because physicians would gain a much richer picture than a reading taken during an occasional medical appointment. But that value depends on the reliability of the technology and on users understanding what the results mean. A patient who receives no warning from a consumer device might incorrectly assume that hypertension has been ruled out, potentially delaying conventional measurement or medical advice.

The problem therefore extends beyond whether a wearable produces a number. Healthcare needs to determine whether the number is sufficiently accurate, whether it is clinically meaningful, how it should be interpreted and what action the patient should take. As consumer health technology becomes more sophisticated, the separation between wellness information and clinical information is likely to become increasingly difficult to maintain. There is also the risk of turning health into a continuous monitoring exercise, with consumers checking sleep scores, recovery measures and other indicators throughout the day even when those numbers may not materially improve medical outcomes.

The issue of access creates another complication. Digital technology is often presented as a way of democratising healthcare by bringing expertise to patients regardless of location, but Whyte questioned whether that promise has yet been demonstrated consistently. Telemedicine and AI can reduce geographic barriers, but digital services can also become additional conveniences for people who already have strong access to healthcare. Insurance, affordability, broadband connectivity, digital literacy and the availability of clinicians still influence whether technology meaningfully expands access.

This means healthcare AI should not be evaluated only by the sophistication of the model. A diagnostic assistant that performs well technically but reaches mainly patients who already have extensive access to care may produce less social benefit than expected. The more meaningful question is whether technology changes outcomes for populations that currently struggle to obtain timely medical advice.

Administrative work remains one of the clearest immediate opportunities. Physicians spend significant time documenting care, completing forms, responding to messages and satisfying reporting requirements. AI can potentially absorb parts of this burden, allowing doctors to devote more attention to patients and complex cases. In this area, the risk of replacing clinical judgement is relatively limited while the potential productivity benefit is substantial.

Whyte’s broader argument, however, is that healthcare should resist allowing technology companies to determine medicine’s priorities simply because a particular tool has become technically possible. The industry should first define the clinical problem it wants to solve and then determine whether AI is the appropriate solution. Starting with technology and searching afterwards for a medical application risks creating impressive products without meaningful improvements in patient care.

That philosophy could increasingly shape how healthcare organisations invest in AI. The winners may not be hospitals or medical groups that deploy the greatest number of models. They may be those that identify where technology genuinely improves diagnosis, reduces administrative burden, expands access or strengthens the relationship between clinicians and patients while preserving accountability.

Medicine’s AI transformation therefore presents a different challenge from many other industries. In banking, logistics or corporate operations, automation can often be judged primarily through productivity, cost and speed. In healthcare, the consequences of an incorrect decision can involve a patient’s health or life. Efficiency matters, but it cannot become the only measure of success.

The question facing medicine is therefore no longer whether physicians will use artificial intelligence. Most already do. The more consequential question is how far that relationship should develop and what doctors must continue to know and decide for themselves. AI may ultimately allow physicians to practise with access to more knowledge, better pattern recognition and less administrative work than any previous generation, but the technology will create the greatest value only if healthcare preserves the ability of clinicians to question the machine rather than simply follow it. The future of medical AI may therefore depend as much on strengthening human judgement as on improving artificial intelligence.

Source: CIJ.World Research & Analysis Team

Insurance’s AI Problem Is No Longer Capability – It Is Economics

Artificial intelligence can perform an expanding range of insurance tasks, but that does not mean insurers should use it for everything. As carriers move from experimentation towards large-scale deployment, the more important question is becoming whether AI is actually the cheapest, most reliable and most defensible technology for the job. That was the central argument of a presentation at AI4 2026 in Las Vegas examining the economics of AI adoption in insurance. Rather than focusing on what the latest models can technically accomplish, the presentation challenged insurers to distinguish between work that should remain conventional software, work where AI can create genuine value and work where human judgement and relationships continue to provide the greatest competitive advantage.

The starting point was deliberately simple. The speaker compared the cost of asking leading AI models to solve a basic arithmetic problem with the cost of performing the same calculation using conventional software. All produced the correct answer, but the AI models were considerably more expensive. The exact cost multiple depends heavily on model choice, token usage, infrastructure and implementation, but the broader point is more important than the individual calculation: using a large language model for a deterministic task can be economically irrational even when the model performs it perfectly.

That distinction becomes significant in insurance because carriers process enormous numbers of transactions. Eligibility checks, coverage validations, rating calculations and many other routine processes operate at volumes where a small difference in the cost of an individual transaction can become material when multiplied across millions of policies or claims. The implication is that the industry’s growing enthusiasm for autonomous agents needs to be balanced against unit economics. A system that is inexpensive to demonstrate can become considerably more costly when deployed across a large insurance operation. The correct question is therefore not simply whether an AI agent can perform a task, but whether it should.

This represents an important change in the enterprise AI debate. During the initial generative AI boom, organisations concentrated heavily on technical capability and employee adoption. Companies measured access to AI assistants, token consumption and the number of pilots underway. By 2026, management teams are placing greater emphasis on measurable financial returns. A widely discussed MIT study published in 2025 found that the large majority of organisations examined were not yet generating measurable returns from their generative AI investments. The finding was sometimes interpreted as evidence that AI itself was failing, but subsequent discussion has focused more heavily on the mismatch between technology and business processes. The problem is frequently not whether AI works, but whether it has been applied to the correct problem and integrated into the organisation in a way that creates economic value.

Insurance provides an unusually clear environment in which to examine that problem because the industry’s work ranges from highly deterministic calculations to decisions that depend heavily on professional judgement. At one end are tasks where a given input should consistently produce the same output. At the other are activities such as complex commercial underwriting, difficult claims negotiations and broker relationships, where experience and interpretation can materially influence the outcome.

The presentation argued that insurers should examine work along this spectrum before deciding which technology to deploy. Deterministic processes that can be expressed through clear rules are generally better suited to traditional software, particularly when they operate at significant scale. Conventional code is predictable, comparatively inexpensive to execute and easier to audit because the same input should produce the same result. The problem is that much of this codifiable work has never been fully automated. Insurance companies have spent decades building core systems, yet employees still perform substantial amounts of rules-based work manually because information arrives in formats those systems were never designed to process.

Broker submissions may arrive as PDFs, spreadsheets, emails, scanned documents and attachments using inconsistent structures. Employees therefore become the bridge between unstructured information and highly structured insurance systems. This is one of the areas where modern AI can have an important role without necessarily becoming the system that ultimately makes the decision. AI can read and interpret unstructured material and convert it into a format traditional software can process. It can also help software engineers build conventional applications considerably faster.

The resulting architecture can therefore combine both technologies. AI handles ambiguity at the edge of the process, while deterministic software executes rules once the information has been structured. That approach can provide many of the advantages of generative AI without requiring an expensive probabilistic model to perform calculations that established software can complete reliably. AI-assisted software development could also accelerate this transition. Insurance companies historically avoided automating some smaller or specialised processes because the development cost could not be justified. If AI coding tools substantially reduce the cost and time required to build conventional applications, some of these previously uneconomic automation projects could become viable.

This creates a different way of thinking about generative AI investment. Instead of replacing traditional software, AI may help companies create more of it. The second category consists of cognitive work that conventional systems have struggled to perform. Commercial insurance submissions illustrate the opportunity. An underwriter may receive dozens of pages containing financial statements, previous losses, property information, policy details and correspondence containing important context. Reading and interpreting the submission can require a significant amount of professional time before the underwriter begins making the actual risk decision.

Large language models are well suited to this type of task because they can extract information, summarise documents, compare different sources and identify relevant passages from unstructured material. AI can therefore operate as an assistant that gives underwriters more time for the work where their judgement creates value. The potential productivity improvement can be substantial. An underwriter capable of reviewing considerably more submissions could increase the insurer’s capacity without proportionately increasing administrative resources. Alternatively, the underwriter could maintain the same workload while spending more time understanding difficult risks, negotiating terms or developing better pricing.

The distinction matters because reducing processing time is not necessarily the same as creating competitive advantage. If every insurer eventually uses similar document-processing tools, faster submission handling simply becomes the expected market standard. The more durable advantage comes from what professionals do with the time and information the technology provides. That could include more sophisticated underwriting, development of new insurance products or the ability to enter risk categories competitors find difficult to price. In that scenario, AI does not generate the competitive advantage by itself. It releases human capacity that the insurer can redeploy into activities that create differentiation.

This becomes particularly relevant as insurers accumulate larger volumes of historical information. Many companies possess decades of claims files, underwriting documents and risk data that have never been fully analysed because much of the information remains trapped in unstructured records. AI can potentially make that material accessible to actuaries and underwriters, allowing them to identify patterns that were previously difficult to extract. Better use of historical data could lead to improved risk models and more precise pricing. Even a relatively short-lived advantage in actuarial understanding can have significant consequences because underwriting decisions compound through future renewal cycles and claims experience.

However, the same characteristics that make large language models powerful also create problems when they are used in deterministic processes. AI models are probabilistic. Their outputs can vary, and performance can change when the underlying model is updated. An application that produces acceptable results using one model version cannot automatically be assumed to behave identically after an upgrade. This creates a reliability problem for insurers. A consumer application can often tolerate occasional variation or an imperfect response. Insurance systems supporting policy decisions, premium calculations or claims require considerably greater consistency.

Carriers therefore need strong testing, monitoring and control environments whenever AI influences consequential processes. Model changes need to be tested against previous behaviour, while systems require clear rules determining when an AI output can proceed automatically and when human review is necessary. Auditability creates another challenge. Conventional software can generally provide a clear record of which rules were executed and which inputs produced an outcome. Generative AI does not naturally provide the same form of deterministic explanation.

This does not mean AI systems cannot be monitored or audited at all. Organisations can record prompts, inputs, model versions, outputs, retrieval sources, intermediate actions and other telemetry. But achieving the level of traceability required for regulated financial processes generally requires additional architecture and governance rather than emerging automatically from the model. That infrastructure also costs money. Logging, evaluation systems, additional model calls, human reviews and testing all increase the real expense of operating AI. The cost of the model itself therefore represents only one component of the total economics.

This is why insurers need to evaluate AI at the level of the complete operating process rather than the price of an individual model call. A system that appears inexpensive during experimentation may require substantial additional controls before it is suitable for production. The third category identified in the presentation is work where human relationships remain central. Claims provide the clearest example. When a policyholder has lost a home, suffered a serious accident or experienced another major event, the interaction with the insurer is not merely an information-processing exercise.

Customers may need explanation, reassurance and someone capable of responding to circumstances that do not fit neatly into a predefined workflow. A technically correct automated answer does not necessarily provide the same experience as a skilled claims professional who can understand the emotional and practical circumstances surrounding the loss. This human element can have commercial value. Insurance is unusual because customers pay for a product they hope never to use. When a major claim finally occurs, the quality of that interaction can determine whether the customer renews, recommends the insurer or leaves.

Automating that relationship purely to reduce operating expenses can therefore destroy value elsewhere in the business. The lowest-cost claims process is not necessarily the most profitable one if it damages retention, reputation or distribution relationships. The more effective model may be to automate everything surrounding the human interaction. AI can gather policy information, summarise claims history, prepare documentation and recommend next steps while allowing the claims professional to concentrate on the policyholder.

The same principle applies to complex underwriting and broker relationships. Technology can provide better information and remove administrative work, but experienced professionals may remain responsible for interpreting unusual risks, negotiating coverage and understanding the commercial circumstances surrounding a client. This leads to a more practical framework for insurers considering AI investment. Instead of beginning with a technology and searching for places to deploy it, companies can begin with an existing business process and break it into individual tasks.

Each task can then be assessed according to its inputs, expected outcome and degree of judgement involved. If the task is deterministic and high-volume, conventional software may remain the best solution. If it involves understanding ambiguous or unstructured information, AI may be appropriate. If success depends heavily on empathy, negotiation or relationships, technology should probably support the employee rather than replace them.

The exercise also requires insurers to involve employees who actually perform the work. Operational processes frequently contain exceptions and informal practices that are poorly understood by senior management or technology teams. Frontline employees know which tasks consume time, which information is difficult to obtain and where judgement genuinely matters. This bottom-up analysis can reveal AI opportunities more effectively than starting with a list of fashionable technologies. It also reduces the risk of automating a process that should have been redesigned or eliminated altogether.

The unit economics become particularly important as agentic AI develops. Autonomous agents can theoretically execute increasingly complex sequences of tasks, but every additional model call consumes computing resources. At modest volume this may appear insignificant. Across millions of insurance transactions, the economics can look very different. Insurers therefore need to distinguish between technical possibility and economic scalability.

Some tasks will justify relatively expensive AI because the value of the decision is high. Spending significantly more computing resources to help an underwriter evaluate a multi-million-dollar commercial risk may be economically reasonable. Using the same architecture to perform elementary calculations millions of times is considerably harder to justify. This is also likely to influence how insurers evaluate technology vendors. Start-ups offering complete AI replacements for existing workflows need to demonstrate not only better performance but sustainable economics once usage reaches enterprise scale.

Carriers will increasingly examine model costs, contractual exposure, reliability, portability and the additional governance required before allowing vendors to become embedded in core processes. The insurance industry’s AI opportunity therefore appears less universal than the early hype suggested, but potentially more valuable. The objective is not to maximise the number of processes touched by AI. It is to identify the places where AI creates an advantage that another technology cannot deliver as effectively.

In some parts of the organisation, that means using AI to turn unstructured information into structured data. Elsewhere, it means helping professionals analyse complex material more quickly. In software development, AI can reduce the cost of building deterministic systems. And in customer-facing situations where trust and empathy matter, the best use of AI may be almost invisible to the policyholder.

This also changes the definition of return on investment. Cutting administrative costs remains useful, but it is only one possible outcome. Better underwriting decisions, more accurate actuarial models, additional capacity, new products and access to previously difficult risks can create considerably greater value. For insurers, the strategic opportunity is therefore not simply greater efficiency. It is the possibility of using technology to redirect skilled employees away from low-value administration and towards areas where judgement generates competitive advantage.

The companies that benefit most from AI may consequently be neither the fastest adopters nor those with the highest model usage. They will be the insurers that understand which parts of their operations should become software, which should become AI-assisted and which should remain fundamentally human. As the technology becomes more capable, making that distinction may become harder rather than easier. AI will increasingly be able to perform tasks that do not necessarily make economic or operational sense for it to perform. The discipline will lie in knowing when not to use it.

For an industry built around evaluating risk, that may ultimately become one of the most important AI decisions of all.

Source: CIJ.World Research & Analysis Team

front page info
LATEST NEWS