The Agentic AI-First Life Sciences Enterprise

Nine in ten large companies use AI. Thirty-seven percent can find it in their earnings. The gap is not a technology gap, and life sciences is where it costs the most.

Enterprise AI Series · CXO Intelligence Series · All writing

01The productivity mirage

The pharmaceutical industry spent 2025 telling itself a good story about research productivity. Deloitte's sixteenth annual analysis of the twenty largest global biopharma companies put the projected internal rate of return on late-stage R&D at 7.0 percent, up from 5.9 percent the year before and the third consecutive year of improvement.¹ After a decade of watching that number fall toward zero, three years of recovery reads like a turning point.

Read the same report one column across. The average cost to develop an asset rose from $2.23 billion to $2.67 billion in a single year, an increase of roughly twenty percent. Average forecast peak sales per asset rose from $510 million to $598 million. The return improved because the numerator grew faster than the denominator, not because the denominator was brought under control.

Read one column further. GLP-1 assets account for an estimated 38 percent of projected pipeline sales, and obesity has displaced oncology as the largest contributor to pipeline value for the first time in the report's sixteen-year history.

Then Deloitte does the arithmetic that settles the question. Exclude GLP-1 assets from the analysis and the rate of return is 2.9 percent.

Returns went up. Strip out one molecule class and the return is 2.9 percent. That is not a productivity recovery. It is a portfolio concentration.

This matters for what comes next, because the industry's answer to the cost side is now, almost universally, artificial intelligence. Sanofi has stated an ambition to become the first R&D-driven biopharmaceutical company powered by AI at scale, with more than seventy thousand employees inside that ambition and named platforms running across research, clinical, manufacturing and internal operations.² Its chief executive, Paul Hudson, has been unusually direct about the measurement problem, arguing that return on AI investment isn't about flashy demos or model size.³

He is right, and the aggregate evidence says the industry has not yet taken his point.

02The adoption-earnings gap

McKinsey's State of AI survey for 2026, fielded in May and June across 1,719 respondents in 97 countries, found that nearly nine in ten organizations now use AI regularly in at least one business function, and eighty percent of respondents report that AI has improved their individual productivity.⁴

Thirty-seven percent attribute any enterprise-level EBIT impact to it. That figure is unchanged from the previous year. Six percent qualify as high performers, meaning they can attribute at least five percent of EBIT to AI and describe the value as significant.

You will have seen the more dramatic version of this. An MIT research group reported in August 2025 that around 95 percent of corporate generative AI pilots produced no measurable return, and the number moved markets. It deserves a caveat most people repeating it do not give. The study rested on 153 surveyed leaders, 52 interviews and a review of some 300 public initiatives, and it measured formal pilot programs rather than enterprise AI value as a whole.⁵ I would not build a board argument on it alone.

The load-bearing evidence is the McKinsey series, because it is longitudinal. The 37 percent did not move between 2025 and 2026, across a year in which adoption, spend and model capability all rose sharply. A single alarming survey can be dismissed. A flat line through a year of that magnitude cannot.

Gartner's first Hype Cycle for Agentic AI, published in April 2026, adds the operational picture. Only 17 percent of organizations have actually deployed AI agents, while more than 60 percent expect to within two years, which Gartner describes as the most aggressive adoption intent among the emerging technologies it tracks. Agentic AI sits at the peak of inflated expectations, and Gartner's analysts note that fully autonomous agents are not ready for the majority of enterprise use cases.⁶ Gartner separately expects more than forty percent of agentic AI projects to be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls.⁷

03What absorbs the gains

Consider what actually happens when a medical writer's first draft of a clinical study report goes from 180 hours to 80, a result Merck and McKinsey reported from a generative AI pilot that also cut errors roughly in half.⁸ That is a genuine, large, verified improvement in a real pharmaceutical task.

Now follow the document. It enters a review cycle whose cadence was set when drafts took four weeks. It waits for a cross-functional review meeting that convenes on a fixed calendar. It queues behind quality control checks sized for the old arrival rate. It reaches a decision forum that meets monthly. The hundred hours you saved are still saved. They are now sitting in a queue.

This is the mechanism behind the funnel. When you accelerate a step inside a process whose sequence, handoffs, controls and review geometry were all designed around the constraint of human throughput, the released time pools in front of the next constraint. The work gets faster. The outcome does not.

McKinsey's own data isolates the variable. Among the six percent of organizations capturing real value, 73 percent report having fundamentally redesigned workflows. Among everyone else, 25 percent do.⁴ This is the single largest differentiator in the study, and it is not a technology variable at all.

04The retrofit trap

I want to name the failure mode precisely, because it is diagnosable and most enterprise AI programs are inside it right now.

A retrofit inserts AI into a process whose shape was determined by human capacity. The agent gets a task. The task sits where it always sat. The sequence, the batch cadence, the approval geometry and the evidence trail are all inherited. The program reports a percentage time saving on the task and cannot report a cycle-time saving on the process, because the process was never the unit of change.

Three things a retrofit structurally cannot fix.

Batch cadence. Work that arrives continuously is still processed in batches sized for meetings. A monthly review committee imposes a floor of roughly fifteen days of average latency on anything that passes through it, regardless of how fast the upstream work becomes.

Review geometry. Controls designed for human error are aimed at the wrong failure mode. A human reviewer checking whether a machine fabricated a citation is performing a different task from a human reviewer checking whether a colleague was careless, and the review process usually has not been told.

Evidence design. In a regulated enterprise, the evidence that a decision was sound is generated as a byproduct of the process. Change who makes the decision without changing how evidence is produced and you get a defensible outcome with an indefensible file.

Four questions that tell you which one you have

Programs rarely describe themselves as retrofits, so the label is not useful diagnostically. These four questions are, and they can be asked in a steering committee in under ten minutes.

Did any meeting get canceled? Genuine redesign removes coordination, because coordination existed to synchronize humans working at human speed. If the calendar is unchanged, the process is unchanged.

Did the unit of work change size? Batch sizes are fossils of throughput constraints. A submission assembled in one pass rather than in sequential functional packages is a different process. A submission assembled in the same packages, faster, is the same process.

Can you point to a role that no longer exists in the workflow, and to a new one that does? Not a job eliminated. A role in the flow. If nobody's position in the sequence moved, nothing structural happened.

Does the audit evidence look different? In a GxP environment this is the sharpest test of all. If the validation package for the new process is the old package plus an appendix about a model, the decision architecture was not touched.

A program that answers no to all four has bought capability and changed nothing. It will report impressive task-level metrics and will appear in next year's survey as one of the sixty-three percent.

05The two clocks

There is a timing argument underneath all of this that most boards have not made explicit, and it is the reason the redesign work cannot wait for the technology to settle.

At Davos in January 2026, Novartis chief executive Vas Narasimhan told CNBC-TV18 that the pharmaceutical sector is likely to see tangible benefits from artificial intelligence over the next seven to ten years.⁹ Coming from the chief executive of one of the industry's most committed AI adopters, that is a sober and probably correct assessment of when the science pays off.

Now set a second clock beside it. The EU AI Act's high-risk obligations, deferred under the omnibus agreement of May 2026, bite in December 2027 for standalone systems and August 2028 for AI embedded in regulated products such as medical devices.¹⁰ FDA and EMA published joint guiding principles in January 2026.¹¹ Gartner's cancellation forecast lands at the end of 2027.

The scientific payoff is seven to ten years away. The governance and design deadlines are eighteen to thirty months away. Those two clocks run at different speeds, and the organizations that treat the first as permission to defer the second will meet the second unprepared.

06Where the value is, with the arithmetic shown

Abstraction is cheap in this field, so here are two places in a pharmaceutical enterprise where the number is knowable.

Regulatory submissions

Leading companies now file eight to twelve weeks after database lock, cutting historical timelines by fifty to sixty-five percent. McKinsey values that acceleration at approximately $60 million of net present value per month for an asset with $1 billion of peak sales, and puts the total at around $180 million for moving from the 2020 median to the 2024 top quartile.⁸

Take a mid-cap company with six late-stage assets averaging $1 billion in peak sales. Two months of acceleration on each is 6 × 2 × $60 million, or roughly $720 million of net present value. That is not a productivity metric. It is a number a board recognizes.

Two details in the same analysis matter more than the headline. Only thirteen percent of companies automate table and figure formatting at scale, and health authority query workflows remain largely unautomated despite being among the largest workload drivers. The unexploited ground is not the glamorous part of the process.

And the method McKinsey names for capturing it is zero-based design, rebuilding the entire submission process starting from the last patient's last visit and running to filing, rather than improving the steps that exist. An outside firm advising the industry arrived at the same conclusion from the opposite direction.

Protocol amendments

Tufts Center for the Study of Drug Development published new benchmarks in 2024, drawn from 950 protocols and 2,188 amendments across 16 companies and CROs. Across phases I to IV, 76 percent of protocols now carry at least one amendment, up from 57 percent in the 2015 baseline, and the mean number of amendments per protocol has risen 60 percent to 3.3.¹²

Volume is not the interesting number, though. This is: the average time from identifying the need to amend to the last oversight approval is 260 days, and investigative sites operate with different versions of the protocol for an average of 215 days.

Sit with those two figures, because they are the entire argument of this article expressed in a single operational statistic. Eight and a half months from recognizing a problem to having the correction approved. Seven months during which the same study is running under different instructions at different sites.

Almost none of that is drafting time.

And here the honest reading cuts against the obvious conclusion. In the 2024 data, 77 percent of amendments were judged unavoidable. Only around a quarter are the product of a design decision a better-informed process would have avoided.

07The stack the work actually needs

Most enterprise AI programs buy two layers and skip five. They buy models and they buy an orchestration framework, which are layers three and four below, and then they discover that value lives in layers five, six and seven, which cannot be purchased from anyone.

A note on reading the diagram in both directions. You build bottom up, because an agent reasoning over data with no provenance produces conclusions you cannot file. You fund top down, because a program that has not answered the layer seven question, who owns the decision and whose P&L carries it, will produce very good pilots and no earnings impact.

A note on buying layers three and four

Layers one to four are real work and should be bought where buying is sensible. The caution is about what the market currently is. Gartner's assessment is that of the thousands of vendors marketing agentic products, roughly 130 are genuinely building agents; the rest are repackaging robotic process automation and chatbots under a new label.⁷ In a regulated environment this is more expensive than ordinary vendor disappointment, because a system that cannot explain its own reasoning cannot produce the evidence layer six requires.

Two procurement questions cut through most of it. Ask a vendor to show how the system decides to stop and escalate, and ask what artifact it produces when it does. A genuine agent has an answer to both because both are design decisions it had to make. A wrapped workflow engine has an answer to neither, because escalation in such a system is a branch in a script rather than a judgment, and the artifact is a log line.

The second question matters more than it sounds. In life sciences the escalation artifact is the audit trail. A system that escalates well and documents the escalation badly has moved your risk rather than reduced it.

08Autonomy is a consequence decision, not a capability decision

The most common error in agentic design is treating the autonomy level as a function of what the model can do. It is not. It is a function of what happens when the model is wrong and how quickly that can be undone.

A literature surveillance agent that misses a paper is caught by the next sweep. A batch release decision that is wrong reaches patients. Those two facts, not the benchmark scores, set the autonomy level.

09What the regulators already decided

This is the part of the argument that changed in January 2026, and a surprising number of AI programs have not caught up with it.

On 14 January 2026, FDA's Center for Drug Evaluation and Research and Center for Biologics Evaluation and Research published, jointly with the European Medicines Agency, ten Guiding Principles of Good AI Practice in Drug Development, spanning nonclinical research, clinical trials, manufacturing and post-market safety surveillance.¹¹ This sits alongside FDA's January 2025 draft guidance on using AI to support regulatory decision-making, which remains in draft and sets out a risk-based credibility assessment framework organized around a defined context of use.¹³

Read the ten principles as an architecture specification rather than a compliance checklist and something becomes obvious. They are not about models. Principle one is human-centric by design. Principle two is a risk-based approach. Principle four is a clear context of use. Principle eight is risk-based performance assessment. Principle nine is lifecycle management. Every one of those is a decision about who decides what, under which conditions, with what evidence.

Two regulators on two continents have converged on the same answer. Autonomy is a risk and context decision, not a capability decision.The transatlantic alignment is the news. The content is what many programs still have not built.

In Europe the timeline moved and the requirement did not. Under the digital omnibus agreement reached on 6 May 2026, high-risk obligations for standalone Annex III systems shift to 2 December 2027, and for AI embedded in regulated products under Annex I, including medical devices, to 2 August 2028. General-purpose AI obligations have been in force since 2 August 2025, prohibited practices since 2 February 2025, and Article 50 transparency obligations remain on the original schedule from 2 August 2026.¹⁰

The deadline moved. The design requirement did not. A company that treats the deferral as relief will build the wrong architecture twice.

10The formula

Every enterprise AI business case I have reviewed multiplies hours saved by a loaded rate. That calculation is why the funnel in Exhibit 2 collapses. It measures the thing that improved and ignores the thing that absorbed the improvement.

11The 100-day proof

The reasonable objection to everything above is that redesigning an enterprise is not a quarter's work. It is not. But proving whether your organization can do it is.

Two design rules make the difference between this being a real exercise and another pilot.

Select by decision density, not by AI feasibility. The instinct is to pick the workflow where the technology obviously fits. That instinct produces demonstrations. Pick instead the workflow where the most consequential decisions per unit of elapsed time are being made by people using incomplete information, because that is where redesign has somewhere to go.

Build layer six before layer four. Every organization builds the agents first and the assurance layer afterward, and then discovers that the assurance layer it needed would have required different agents. In a GxP environment this is not a sequencing preference. It is the difference between a system you can validate and a system you rebuild.

12The board's five questions

Nothing in this argument is a technology decision, which is inconvenient, because technology decisions can be delegated and these cannot.

The next article in this series takes the same argument to a place with an unusual advantage. Most health systems are trapped retrofitting AI onto twenty years of accumulated digital process debt. A few are building capacity now, at scale, on a compressed timeline, which is the one condition under which you can design the operating model assuming AI is a first-class workforce rather than bolting it on afterward.