Sunday, 6 September 2026
Artificial intelligence has attracted enormous investment across the pharmaceutical industry, yet the promised productivity revolution has not arrived at the same speed. Drug development remains expensive, clinical programmes remain lengthy and many AI initiatives are still struggling to move beyond demonstrations and pilot projects. That contradiction formed the central argument of a presentation by Kris Kaneta, Chief Product & Innovation Officer at Norstella, during AI4 2026. Kaneta’s message was that the pharmaceutical industry’s biggest AI problem may no longer be access to powerful models, but the inability to convert those models into trusted systems capable of supporting complex decisions involving billions of dollars, years of development and ultimately patients waiting for new treatments.
The timing of the discussion is significant. Pharmaceutical companies have spent heavily on digital technologies, data platforms and artificial intelligence, while the underlying economics of drug development remain difficult. Industry research continues to show rising development costs and long clinical timelines. AI may eventually help address those pressures, but deploying more tools does not automatically make pharmaceutical research more productive. Kaneta characterised the problem as a gap between AI demonstrations and actual decision-making. Pharmaceutical organisations have accumulated pilots, prototypes and proof-of-concept projects that can generate impressive outputs but frequently struggle when they encounter real business processes. The reasons are not necessarily failures of the underlying AI models. A system may work technically while lacking the data, context or workflow integration required to make it useful inside a pharmaceutical company.
That distinction matters because many decisions in life sciences are unusually consequential. Companies must decide which molecules deserve further investment, whether compounds should be developed for additional diseases, how clinical trials should be designed, where studies should be conducted, which investigators should participate and how medicines should ultimately reach patients. A mistake at an early stage can influence years of subsequent investment. Unlike editing a document or generating marketing material, many pharmaceutical decisions cannot simply be reversed once significant capital has been committed. This raises the standard that AI must meet before executives and scientists are willing to rely on it. A plausible answer is not enough. Users need to understand where information came from, whether the underlying evidence is current and whether the system has enough understanding of the specific pharmaceutical problem to produce something useful.
Kaneta argued that trust in these systems rests on several fundamental qualities. Outputs need to be traceable to credible information, they need to remain sufficiently reliable when similar questions are repeated, they must reflect the context of the job being performed, and they ultimately need to influence a real decision or action rather than simply produce an interesting response. That is a much higher threshold than the one required for a successful AI demonstration. A generic language model, for example, can produce a convincing summary of a therapeutic market, but pharmaceutical competitive intelligence requires more than assembling publicly available information. Analysts need to understand whether drugs remain in development, whether trials have been terminated, where regulatory approvals are pending, how competitors’ programmes are changing and how those developments affect a particular company’s strategy. An answer that includes a discontinued product or misses a recently approved medicine may still read perfectly well, yet it could lead to the wrong commercial conclusion.
This is one reason domain-specific data is becoming increasingly valuable in the AI economy. The competitive advantage may gradually shift away from simply possessing access to a large model and towards controlling well-structured, reliable and continuously updated information that gives the model meaningful context. In pharmaceutical markets, that information is spread across numerous areas. Drug pipelines, clinical trials, investigators, regulatory decisions, payer policies, reimbursement, market access, real-world patient data and commercial forecasts may all influence a single strategic decision. Historically, much of this information has existed in separate systems and organisational departments. Business development may maintain different intelligence from market access teams, while clinical development, forecasting and commercial teams use their own databases and analytical processes. Artificial intelligence creates an opportunity to connect these information environments, but only if the underlying data infrastructure is built to support that connection.
Norstella’s response has been to develop Atlas, an agent-based platform designed to combine its pharmaceutical datasets and intelligence across different parts of the drug-development lifecycle. Rather than creating one general AI assistant for every pharmaceutical task, the approach involves developing systems around particular professional roles and decisions. That distinction could become important across enterprise AI more broadly. A competitive intelligence analyst, clinical-trial specialist and market-access professional may all use artificial intelligence, but they do not need the same information or reasoning process. The system therefore needs to understand the objective of the user, not simply retrieve documents containing related words.
Kaneta used the analogy of giving an executive a thousand interns. The additional manpower could theoretically produce enormous output, but only if those workers were given precise instructions, reliable information and a clear definition of what a successful result should look like. Artificial intelligence creates a similar management challenge at vastly greater scale. An organisation can generate huge quantities of analysis, reports and recommendations, but additional output does not automatically mean additional productivity. Without reliable context, businesses risk producing more information without improving the quality or speed of the decisions that matter.
This may help explain part of the apparent productivity paradox surrounding generative AI. Companies can automate individual tasks while the total process remains slow because the underlying decision structure has not changed. A pharmaceutical company might reduce the time required to prepare an analysis from several days to several hours, for example, yet still spend weeks validating the information, moving it between departments and securing approval before acting. The real productivity opportunity therefore lies in redesigning the complete workflow rather than accelerating isolated components. This also changes how companies should evaluate AI investments. The number of employees using an AI platform or the quantity of content generated may say relatively little about its economic value. More relevant measurements include whether clinical trials can be designed more effectively, whether investment decisions are made earlier, whether unsuitable programmes are stopped sooner and whether promising medicines reach patients faster.
The financial implications can be enormous. Drug development requires billions of dollars of capital across research, clinical testing, manufacturing preparation and commercialisation. Improving one major portfolio decision could potentially create greater value than automating thousands of routine administrative tasks. Conversely, an unreliable AI recommendation used in a high-value investment decision could destroy far more value than the technology saves elsewhere. The quality of the data underneath the system therefore becomes an investment issue in its own right. Pharmaceutical companies have accumulated decades of clinical, regulatory and commercial information, but much of it remains fragmented across departments, legacy systems and external providers. Preparing that information for AI can involve substantial expenditure on integration, data governance, metadata, cybersecurity and cloud infrastructure.
The AI transformation of pharmaceutical companies may consequently produce considerable demand for technology infrastructure without immediately appearing as productivity in drug-development statistics. Companies are effectively building a new information architecture while continuing to operate extremely complex existing businesses. The benefits may only become visible once those systems begin influencing decisions across entire development programmes rather than isolated tasks. Another concern is the increasing volume of AI-generated material entering the information environment. As generative systems produce more articles, summaries and analysis, future models may encounter material that was itself generated by other models. Without strong links to original evidence, this creates the possibility of information circulating repeatedly while becoming progressively detached from authoritative sources. That risk is particularly problematic in pharmaceuticals, where outdated or inaccurate information can materially affect investment decisions.
Traceability therefore becomes more than a technical feature. It becomes part of the organisation’s risk-management structure. Executives need to be able to understand why a system reached a conclusion and verify the underlying evidence before committing capital or changing a development strategy. The same principle applies when AI moves towards more autonomous agents. An assistant that produces information remains relatively easy for a human to review, while an agent capable of triggering subsequent actions introduces an additional level of responsibility. The more autonomy these systems receive, the more important it becomes to establish precisely what information they can access, what decisions they can influence and when human approval remains necessary.
Pharmaceutical AI is therefore likely to evolve differently from consumer generative AI. Speed and ease of use remain valuable, but trust, provenance and specialist context can be considerably more important than producing an immediate answer. This favours organisations capable of combining technology with proprietary information and deep sector expertise. It may also reshape competition among pharmaceutical information providers. Companies that historically sold databases, market intelligence and research tools increasingly have the opportunity to transform those assets into the contextual layer supporting AI agents. In this model, the underlying database becomes more valuable because artificial intelligence can interrogate it continuously and connect previously separate areas of information.
The implications extend into pharmaceutical corporate strategy. As AI systems become embedded across research, development and commercial operations, businesses may need to reconsider how departments share information. Traditional organisational silos can directly reduce the effectiveness of AI. A system attempting to evaluate a drug’s commercial opportunity, for example, becomes considerably more useful if it can connect clinical evidence with competitor activity, payer behaviour, regulatory developments and real-world patient information. Breaking down those information barriers can be as difficult as developing the technology itself.
The challenge also affects corporate investment priorities. Pharmaceutical companies may be tempted to fund numerous AI experiments because the cost of creating individual prototypes has fallen dramatically. Yet the more initiatives an organisation launches, the harder it becomes to integrate, govern and evaluate them. A smaller number of applications connected to strategically important decisions may ultimately create more value than hundreds of disconnected AI pilots. The pharmaceutical industry does not have a shortage of artificial intelligence experiments. It has a shortage of AI applications trusted enough to influence decisions that genuinely matter.
Closing that gap could determine whether the industry’s technology investment eventually translates into higher research productivity. The stakes extend beyond corporate efficiency. Longer development timelines mean patients wait longer for medicines, while escalating R&D costs affect which therapies companies are willing to pursue in the first place. If artificial intelligence can help companies identify stronger drug candidates earlier, design better clinical programmes, select more appropriate patients and make faster portfolio decisions, its economic value could ultimately be measured in both capital efficiency and time.
The next phase of pharmaceutical AI may therefore be considerably less visible than the generative AI boom that preceded it. The important innovation will not necessarily be another chatbot or increasingly powerful general-purpose model. It will be the information and decision infrastructure sitting underneath those technologies. For pharmaceutical companies, the central question is shifting from how much AI they are using to whether that AI can be trusted to help make better decisions when billions of dollars and years of drug development are at stake. Until that happens consistently, the industry may continue experiencing the same paradox: increasingly sophisticated artificial intelligence operating inside a drug-development system that remains expensive, slow and difficult to change.
Source: CIJ.World Research & Analysis Team
