From experimentation
to transformation.

The starting point for this article was a question about developments that appear difficult to reconcile. Pharmaceutical companies are announcing substantial AI investments, yet we also hear about pilots that do not scale, restrictions on data access and internal processes that are proving difficult to change. At the same time, global capability centres are expanding and initiatives such as the partnership involving AstraZeneca, Oracle and M42 are exploring closer connections between research and patient care.

I wanted to understand how these developments fit together and what they tell us about the next stage of adoption. From a PharmaLedger perspective, I am particularly interested in whether the right people can access information they can trust and use it to take an appropriate next step. That question applies throughout the industry, from the scientist deciding which experiment to conduct to the patient looking for current medicine information.

The research below suggests that progress and difficulty are both real, although they occur in different parts of the business and at different stages of implementation. Understanding those differences gives us a more useful basis for deciding where to invest and what we should expect in return.

Although conversational AI has brought the subject to a much wider audience, pharmaceutical companies were applying machine learning to molecular prediction and biological analysis before these tools became familiar in everyday work. FDA reports experience with more than 500 submissions containing AI components between 2016 and 2023, which demonstrates an established history of use without implying that 500 AI-discovered medicines had been approved.^[1]

The expansion of the AlphaFold database to more than 200 million predicted protein structures in July 2022 made the scientific possibilities especially visible. Researchers gained access to predictions on a scale that would have been difficult to imagine previously, while experimental and clinical work remained necessary to establish what those predictions meant for a treatment.^[2]

Generative AI subsequently allowed medical writers, quality specialists, commercial teams and administrative staff to interact with systems directly. As a result, the challenge extended beyond developing a useful model to deciding how a much larger workforce could apply it within existing processes and responsibilities.

What interests me about the direction of spending in this survey is the increasing emphasis on improving existing workflows, alongside a reduced emphasis on finding additional use cases. I interpret this as a sign that organisations are concentrating more attention on making selected applications work in practice. The survey provides a useful indication of that shift, although the eventual quality of the operating results remains the measure by which it should be judged.

To give the comparison a consistent basis, I used the top ten companies in Fierce Pharma’s ranking by 2025 company revenue. This includes J&J’s MedTech business and Roche’s diagnostics business, so it should not be read as a ranking of prescription-drug sales alone; Merck refers here to the US company known internationally as MSD.^[3]

I would be cautious about assigning any of these companies a single level of AI maturity, since an established application in one function may coexist with early experimentation elsewhere. Public disclosures also tend to provide more detail about selected successes and investments than about unsuccessful tools or their full operating costs, which limits the comparisons we can make.

The range of activity shows that companies are looking beyond molecule design to documentation, trial operations, commercial planning and management decisions. In assessing this activity, I find it useful to consider how far an application has moved from an announced intention to an operating process whose contribution can be measured.

MSD provides a concrete example: the company reported reducing the effort required for a fully human-reviewed clinical study report first draft from 180 to 80 hours, together with a 50% reduction in errors in specified categories. It also said that reports produced using the platform had already been submitted. Although these are company-reported results for one stage of work, they allow us to see more clearly what changed.^[6]

The next question is how that improvement affected the work that followed. A reduction in drafting effort could create capacity for additional programmes or help a team complete its work sooner, but we need to establish which outcome occurred before drawing conclusions about the wider business benefit.

A UK government-commissioned survey of engineering biology illustrates the difference between using AI and making it part of an established process. Among 368 respondents, 69% reported AI use, while only 7% reported fully embedded use. The survey covered a broader community than pharma and was non-representative, so I would use it to illustrate an implementation challenge rather than present its percentages as pharmaceutical industry adoption rates.^[16]

The difficulty also extends to measuring results. At Pistoia Alliance’s April 2026 conference, 30% of respondents reported an enterprise-wide AI platform, while 69% lacked clear metrics for assessing its effect on drug R&D time or cost. Because the release does not provide a response count for each question, the event’s 300 attendees should not be treated as the sample for every finding. Nevertheless, the polling gives us a reason to pay closer attention to how organisations assess value.^[14]

In the same polling, 54% identified regulatory submissions and reporting as the teams benefiting most, compared with 21% identifying research analysis teams. These assessments help explain the prominence of document and information workflows in the company examples, without providing an independent measurement of productivity.^[14]

Bessemer’s 2026 research provides another indication of the scaling difficulty: within its framework, pharma’s pilot-stage share increased from 20% to 31%, while production moved from 5% to 6%. These are figures from a survey of healthcare executives, rather than the percentage of pharmaceutical companies using any AI.^[15]

A pilot may be supported by selected data, committed users and immediate access to its developers, whereas an operating process has to accommodate incomplete records, unfamiliar users and exceptions. This helps explain why a model can perform its immediate task well while the organisation still has substantial work to do before the application can be used routinely.

The interest in discovery is understandable, because a better choice of candidate or biological target could avoid substantial work on an unsuccessful programme. However, the ability to generate candidates more quickly and the ability to improve clinical success require different evidence, and we should take care not to assume that progress in the first establishes the second.

IQVIA’s 2026 analysis found a 75% phase I success rate in a relatively small validated cohort of AI-enabled emerging-biopharma programmes, while phase II performance was comparable with non-AI peers and phase III transitions were too few for meaningful analysis. I regard this as encouraging evidence whose limits remain important.^[18]

A separate October 2026 study identified 129 AI-enabled therapeutic assets across 69 companies, of which only five had entered or completed phase III. The cohorts and methods differ between studies, so their success percentages should not be combined into one industry estimate.^[19]

I expect investment to continue, but I would like to see a clearer account of how the resulting predictions improve development decisions. Evidence about experiments avoided, candidate quality and performance in human studies would help us judge whether early advantages are carrying through into clinical value.

The country perspective helps explain why offshore expansion, investment in infrastructure and selective AI deployment can occur at the same time. India is developing global delivery capability, the UK is investing in research-data access and Abu Dhabi is connecting clinical and genomic information, while the Chinese manufacturing and Swedish computing examples illustrate other contributions. Since these figures measure different capabilities, they are more useful for understanding the overall development than for ranking countries against one another.

EY identified Indian capability centres belonging to 23 of the world’s top 50 life sciences companies, with responsibilities extending into discovery, clinical operations, pharmacovigilance and regulatory affairs. This indicates a development broader than the relocation of routine IT support, although it does not quantify jobs transferred from other countries or displaced by AI.^[20]

In the UK, the Health Data Research Service’s chief executive described £600 million in five-year funding, comprising £500 million from government and £100 million from Wellcome. This investment addresses access and infrastructure, and its eventual value will depend on what researchers can do with the resulting service.^[21]

At AstraZeneca’s Wuxi site, a programme of 34 industrial applications, including AI and algorithms, reported 55% higher output and a 44% reduction in manufacturing lead time. Because these were results from a combined transformation programme, they cannot be assigned to AI alone.^[22]

Taken together, these examples suggest that AI will develop alongside changes in data infrastructure, engineering capability and operating processes. From a management perspective, the outcome of that wider change matters even when it is difficult to isolate the contribution of each technology.

The AstraZeneca–Oracle–M42 initiative is relevant because it explores how these capabilities could connect within care delivery and research. Announced on 20 October 2026, its first Abu Dhabi programme concerns uncontrolled hypertension and clinical research, under a non-binding agreement in which projects require separate review and AstraZeneca will not receive patient-identifiable information.^[13]

The setting provides substantial foundations for this work: Malaffi reported 3.5 billion clinical records across 3,072 facilities, while a separate Emirati Genome Program preprint described 504,000 sequenced individuals, including 426,382 with linked clinical phenotype data in its analysis. These figures help explain the opportunity, but do not establish results for the new hypertension initiative.^[23][24]

What interests me is how the organisations’ different contributions could support the work of clinical and research teams. The practical test will be whether the combination helps people identify an appropriate next step, improves recruitment and creates benefit without adding an excessive burden to those delivering care.

At PharmaLedger, I would approach the wider discussion by asking what information must be trusted, who should be allowed to use it and what needs to happen as a result. Medicine information and supply visibility are relevant because a patient pathway depends on people being able to act on information and access the required medicine. This is the perspective I bring to these partnerships.

As companies gain experience, some are reconsidering earlier choices. J&J’s CIO described reviewing nearly 900 generative AI use cases and concentrating on high-impact, scalable applications, supported by secure environments and curated data. I see this as a decision about where the organisation should focus its resources.^[4]

Sanofi’s digital chief said a small Copilot pilot did not lead to broad employee deployment.^[29] This is useful evidence about one product in one setting, although it does not settle the wider question of internal AI adoption.

A project may need to stop because its output is unreliable, checking consumes the expected savings or integration costs outweigh its benefit. Security can also be a genuine constraint, but I would want to understand the particular circumstances before drawing a conclusion about the technology or the organisation.

For that reason, I do not regard the wider transition as a failure simply because some pilots are delayed or discontinued. We are still establishing how a new generation of tools fits into a demanding industry, and part of that work involves recognising the applications that do not justify further investment. The learning has value when it improves what we do next.

Research provides some support for taking the organisational dimension seriously. An August 2026 peer-reviewed study found associations between business success and organisational capability maturity, particularly project maturity, while model type and data source showed no significant associations on their own. This does not establish causation, but it reinforces the importance of preparing an organisation to use the technology.^[25]

Change management therefore needs to begin with the process itself. If AI prepares a regulatory draft, the organisation has to decide how the writer’s responsibilities change, what the reviewer must verify and how the next team can rely on the approved result. These decisions affect people whose existing experience and professional accountability remain important.

Training in how to use a tool will only address part of the requirement, because employees also need time to learn the revised process and confidence that raising a problem will lead to an improvement. A reviewer who repeatedly corrects the same output may be identifying a design issue, just as an employee who avoids an additional interface may already have a more effective way of completing the task.

Data protection and governance need the same attention to practical responsibilities. A controlled hosting environment is useful, but access permissions, source quality, evaluation and monitoring must also be understood across the teams involved if the application is to operate reliably.

FDA and EMA’s January 2026 principles address intended use, data governance and lifecycle management, while FDA’s qualification of AIM-NASH for a defined trial use provides an example of controlled adoption in which pathologists retain final responsibility.^[26][27]

Over the next twelve months, I expect companies to expand AI selectively within existing workflows, particularly where results can be assessed over a shorter period than is possible in drug discovery. Documentation, knowledge retrieval and defined analytical tasks should therefore provide opportunities to establish whether an application delivers a sustained improvement.

I also expect agents to take on more work within clearly defined permissions and escalation arrangements, with the degree of autonomy depending on the evidence and the consequences of an error. Preparing a recommendation and executing a decision create different responsibilities, which should be reflected in how each application is assessed.

Discovery investment is likely to continue, although a year is a short period in which to demonstrate a durable improvement in industry-wide approval rates. For healthcare partnerships, I will be looking for operating programmes that report what changed for recruitment, clinical teams and patients.

My starting point would be to choose a workflow that matters and understand how it performs before introducing a new approach. Someone then needs the authority and responsibility to improve it, with review, support, training and maintenance included in the assessment of cost. Agreeing in advance what would justify expansion, revision or closure would make subsequent decisions easier to explain.

Some pilots will close during this period, and that can be a sensible outcome when resources move to applications with a stronger case. What matters is whether the organisation becomes better at selecting, implementing and maintaining approaches that improve its work.

By October 2027, I would like to see more repeatable evidence of fewer errors, less rework, better experiments and faster decisions supported by reliable information. From a PharmaLedger perspective, the patient remains the reason to connect these developments, because the ultimate value of our industry’s investment will depend on whether people can work together more effectively along the path from scientific insight to care.