Thursday, 10 September 2026
Corporate enthusiasm for artificial intelligence has reached extraordinary levels, but the financial returns inside many established companies remain much less dramatic. That disconnect is becoming one of the most important questions confronting chief information officers, technology executives and investors as businesses move from AI demonstrations towards large-scale deployment. At AI4 2026, Shadman Zafar, CEO of Vibrant Capital and a technology executive whose career has included senior roles at major financial and telecommunications companies, argued that the biggest obstacle to AI value is increasingly not the capability of the models themselves. It is the ability of companies to redesign their technology architecture, governance and organisations around them.
The starting point for the presentation was a striking contradiction. Technology companies increasingly describe AI as a transformational force capable of changing entire industries, yet a large proportion of businesses have still not translated adoption into clear financial gains. PwC’s 2026 Global CEO Survey supports that concern. Among 4,454 CEOs across 95 countries and territories, 56% reported that AI had produced neither higher revenue nor lower costs during the previous year, while only 12% said it had delivered improvements on both sides. At the same time, companies that have embedded AI more deeply appear considerably more likely to report financial benefits, suggesting the issue may be less about whether the technology works and more about how effectively organisations are implementing it.
That distinction matters because corporate AI spending is entering a different phase. The first years of generative AI were dominated by experimentation, pilots and individual productivity tools. Boards and finance departments are increasingly asking what those investments have actually changed in the income statement. Faster document summarisation or better coding assistance can clearly save time, but unless those gains change operating costs, increase capacity, improve revenue or allow the organisation to work differently, productivity improvements may never become meaningful financial returns.
Zafar argued that much of this gap can be traced to the technology architecture companies assembled during the first wave of generative AI. Many organisations moved quickly, connecting applications to large language models, adding retrieval systems and introducing new agent tools as they became available. The resulting environment can work for individual demonstrations but becomes increasingly difficult to operate as models, applications and suppliers change.
The problem is particularly visible in the constant turnover of AI tools. Companies can spend months introducing one coding assistant, training employees and redesigning workflows only to discover that another product has suddenly become more capable. Repeating that cycle each time a new model or tool appears can prevent employees from becoming proficient with any of them and can turn AI adoption into a permanent technology migration programme.
The alternative is to design the architecture so that individual models can change without forcing the organisation to rebuild everything around them. Rather than selecting a single AI model for every task, businesses can operate a portfolio of models and route different workloads according to their requirements. Routine, high-volume tasks can be directed towards smaller and cheaper models, while complex problems are escalated to more capable systems. This could become particularly important as the cost of AI increases with usage. Using the most powerful available model for every request may be technically simple but economically inefficient, because a large proportion of enterprise tasks do not require maximum reasoning capability.
The wider implication is that enterprises may need to treat AI models increasingly as interchangeable computing resources rather than permanent strategic platforms. The competitive advantage would then move away from access to a particular model and towards the architecture that decides how models are selected, how data reaches them and how their output is governed.
Data represents another part of this problem. Generative AI is only as useful as the organisational context surrounding it. If information is outdated, inconsistent or disconnected across multiple systems, an AI application can produce an apparently convincing answer based on the wrong underlying facts. That creates a different type of risk from the hallucination problem usually associated with generative AI. Zafar argued during the presentation that poor enterprise context can be more dangerous than a model simply inventing information because the answer may appear entirely reasonable while being based on outdated or inconsistent company data. The numerical comparisons he presented on this point were based on his own analysis rather than an independently established industry benchmark, but the underlying problem is important: better models cannot compensate indefinitely for badly managed corporate information.
For large companies, this increases the importance of connecting operational and analytical data more effectively. Traditional enterprises often have information distributed across transaction systems, data warehouses, analytics platforms, document stores and more recently retrieval systems built specifically for AI. If those sources update independently, different versions of the organisation’s reality can emerge. An AI-ready data architecture therefore needs to keep business context current as underlying information changes.
Governance requires a similar redesign. Traditional technology governance relies heavily on policies, approvals and reviews performed before software enters production. Autonomous agents introduce a different challenge because they can make decisions and take actions continuously after deployment. As the number of agents grows, organisations may increasingly need to govern them in a way that resembles workforce management. Each agent needs an identity, a defined role, specific permissions, operating boundaries and an auditable record of what it has done. Policies therefore increasingly have to be enforced directly by the technology rather than existing only in compliance documents.
This principle connects closely with the emerging enterprise agent environment. If AI systems become responsible for performing recurring operational work, companies will need to know exactly which information they can access, what systems they can modify and when approval from a human is required. AI governance consequently becomes part of the operating architecture rather than a separate compliance function.
The organisational implications may be even more significant. Much of the debate around AI and employment assumes that companies will automate routine work and simply remove the employees who previously performed it. Zafar argued that such an approach risks weakening the organisation over time by removing the mechanism through which future expertise is created. Traditional companies frequently operate through a pyramid, with senior employees making the most complex decisions, middle management coordinating activity and larger numbers of junior employees performing much of the detailed work. AI could reduce the amount of routine execution required at the middle and lower levels, but eliminating junior recruitment entirely would create another problem: there would eventually be nobody developing the experience required to become the next generation of senior employees.
The alternative proposed during the presentation was closer to an hourglass. Experienced professionals at the top use AI to multiply their capabilities, many routine middle-layer activities become automated, but organisations continue bringing in younger employees at the bottom and train them for a different set of responsibilities. Those responsibilities could increasingly include evaluating models, supervising automated workflows, controlling data quality, testing AI systems and managing the lifecycle of intelligent applications rather than performing the repetitive work that traditionally occupied the first years of a corporate career.
This relates to a deeper risk from excessive dependence on external AI. Companies learn partly because employees perform work, make mistakes, receive feedback and gradually improve. If increasing amounts of cognitive work are simply sent to external models without companies capturing the lessons generated in the process, some institutional learning could disappear. An organisation may therefore become more productive in the short term while becoming less capable of developing its own expertise.
The challenge is to create internal feedback systems that capture outcomes, errors and corrections so that organisational knowledge continues to improve even as machines perform more of the execution. This is particularly relevant because many companies deliberately prevent external AI providers from using proprietary corporate information to train their general models. That is sensible from a data-security perspective, but it also means improvements generated through internal experience will not automatically return to the company unless it builds its own learning mechanisms around those systems.
The emerging concept of AI lifecycle management therefore extends beyond selecting a model and deploying an application. Companies need systems for monitoring performance, evaluating errors, improving prompts and workflows, maintaining data quality, controlling permissions and learning from how employees and agents perform tasks over time.
The financial argument behind this approach is becoming stronger because the companies producing the most visible AI returns appear to be those moving beyond isolated applications. PwC’s research found that organisations with stronger AI foundations were more likely to report financial gains. This suggests that the next divide in corporate AI may not be between adopters and non-adopters. Most large organisations are already adopting the technology in some form. The more important division could be between businesses that simply add AI tools to existing processes and those that redesign the underlying operating system of the company.
For investors, this also changes how AI exposure should be assessed. A company announcing dozens of pilots or purchasing licences for the latest models may provide little evidence that artificial intelligence is changing its economics. More important questions concern whether processes have actually been redesigned, whether data is accessible and reliable, whether AI usage can scale without uncontrolled costs and whether measurable productivity gains eventually reach margins or revenue.
The same distinction applies to technology vendors. Enterprises are becoming increasingly crowded with AI products competing for limited implementation capacity. Products that generate impressive demonstrations but require companies to rebuild architecture continuously may struggle as corporate buyers become more selective. Systems that integrate into durable operating frameworks and produce measurable outcomes are likely to become more valuable.
The most important phase of enterprise AI may therefore have little to do with which model wins the technology race. Foundation models are advancing rapidly and will increasingly be available to many competitors simultaneously. The harder problem is building companies capable of absorbing that intelligence efficiently. That requires architecture able to survive changing models, data systems that maintain accurate context, governance that follows agents into production, employees capable of supervising increasingly automated work and management systems that convert productivity improvements into genuine financial performance.
The AI revolution inside established companies will ultimately be judged less by how impressive the technology appears than by what reaches the bottom line. On that measure, the race is only beginning.
Source: CIJ.World Research & Analysis Team
