Pharma's AI Retrofit Ceiling

The scene is the same at every company I walk into, and only the slide template changes. Four pilots get presented, three of them work, and one works well enough that the room gets briefly and genuinely excited: the next-best-action model beats its control cell, the content tool takes a step out of the brief-to-asset path, and a line in the cost base has moved in the right direction, though nobody can say how much of that was AI and how much was the reorg running alongside it. Then the conversation turns to what it would take for any of this to become the way the company actually sells, and the temperature drops, because that question was never about the model.

It is about whether the brand director gives up the campaign calendar. Whether the MLR head accepts a review process he did not design and cannot fully see into. Whether the regional lead gives up the right to override targeting when a district is short. Whether somebody’s headcount, which still determines their pay-grade, comes out of the plan. Every commercial AI use case I have watched stall has been stopped by one of these hurdles, not on the technology, and every postmortem I have read afterward came back labeled change management.

If you have watched this happen four or five times, then you start to see the full shape. The shape is a ceiling and the name is the Retrofit Ceiling: the limit on how much AI an established pharma’s existing commercial model can absorb before the political capital required to absorb more exceeds the productivity it produces. And it is not the argument it routinely gets mistaken for, because AI works, the vendors mostly ship what they sold, and the pilot that looked good in the demo usually was good. What does not work is applying it at the margins of an operating model designed for a world where it did not exist.

The room was designed around walls that are no longer holding anything up

Your commercial model is a room, and it is a well-built room, because every wall in it was holding real weight the day it went up. Account teams cover the territories they cover because human reach is finite and the reach-to-density math was done by people who were right. Brand strategy runs annually because the data that would tell you to change course arrived annually. Segmentation is broad because personalizing across millions of decisions was not possible, so the segment was the honest unit of work. Review takes as long as it takes because the binding constraint is human attention on a reviewer bench sized for a slower world.

AI dissolves some of those walls. Not all of them, not evenly, and the compliance ones are structural in a way the others are not. But the ones it does dissolve, it dissolves completely, and that leaves you with a choice nobody frames as a choice in a budget meeting.

The retrofit move keeps the room intact and tries to move the wall a foot to the left with a chatbot, and the wall does not move, the chatbot eventually gets unplugged, and somewhere in a steering committee deck the failure gets attributed to change management. The tell is not sophistication, because retrofit pilots are frequently the more sophisticated ones - they spend their intelligence compensating for the room rather than changing it. The strongest commercial data scientist I have worked alongside spent most of a year building a feature store that existed for one reason: the CRM could not answer what a physician had done in the last eleven days. No amount of model quality was ever going to touch that.

Every win you have been shown is a win at the margin

The most-cited evidence that AI is already working in pharma commercial is a CEO on an earnings call, and it is also the cleanest description of the ceiling in print.

Read the last clause again, because Bourla is describing the ceiling while he believes he is describing a triumph. Reducing the cost without that being seen in the activity is cost substitution: same activity, same operating model, fewer people and dollars producing it. That is real money and I would take it in any year, but nothing about how Pfizer decides what to promote, to whom, on what cadence, through which review path, was rebuilt, and he credits simplification in the same breath, mid-restructuring. Cost-out is the cheapest and least contested class of AI use case in a commercial organization because it takes no decision right away from anyone who holds one, which is exactly why it is the one that scaled.

The rest of the evidence reads the same way. ZS asked 115 pharma and biotech technology executives in July 2025 and 47% said they were already consistently demonstrating measurable value in commercial sales, marketing and HCP engagement, which reads as a refutation until you notice the question asked about value “financial or otherwise” and was answered by the executive who owns the AI budget. In the same study, among firms that pursue innovation only when the value story is clear, just 40% of pilots reach scaled deployment. Both are true and they are one finding: marginal application is easy, it works, and it stays marginal. IQVIA announced in March 2026 that nineteen of the top 20 pharmaceutical companies “have begun incorporating” its agents into their workflows, and “have begun” carries that whole sentence, because beginning is free - a procurement decision, a pilot budget and a slide.

Deloitte put a version of the same question to 150 life sciences leaders in April 2026, and the gap in the answers is, in my judgment rather than Deloitte’s, the ceiling with a number on it - self-reported, by a sample with every incentive to report otherwise.

Watch where the headline capital goes and the same pattern surfaces in public. NVIDIA and Lilly announced a co-innovation AI lab at JPM in January 2026, up to $1 billion over five years, to reinvent drug discovery, with commercial operations sharing a single subordinate clause at the end alongside clinical development and manufacturing. NVIDIA sells compute and a lab with NVIDIA lands on molecular modeling by construction, so I went looking for the offsetting commercial announcement and only found more drug discovery efforts. Lilly licensed Insilico Medicine’s AI-discovered medicines in March 2026 for up to $2.75 billion, $115 million of it upfront and the rest contingent on milestones, following a $100 million agreement the previous November. One company, three named AI commitments inside fifteen months, all of them discovery - which is the answer to the objection that this is just what happens when your partner sells GPUs. And Sanofi, the company most often called the industry’s AI frontrunner, describes the destination through its Chief Digital Officer as “reinvention, not automation,” with its people “amplified by it.” Amplified is augmentation: the room staying where it is, described by the executive who owns the technology rather than the commercial P&L.

Check who published the number before you check the number

Nearly every quantity in this market is published by a company selling the cure, and you are being asked to make a rebuild decision on vendor arithmetic. Veeva’s founder-CEO, Peter Gassner, told investors on 3 June 2026 that commercial content “will be a key area of AI investment as we look to solve the MLR content review bottleneck for the industry.” He is right, and it is the best-sourced statement anywhere on the review-capacity constraint: a public-company CEO naming it at industry level in prepared remarks. Twenty days later, on 23 June, Veeva acquired Copli and launched an agentic MLR product. The dominant vendor named the industry’s bottleneck and bought its way deeper into it inside three weeks. Both things hold - the bottleneck is real, and the party naming it is positioned to sell you the relief - much as IQVIA’s agent-penetration number was announced at a GPU conference with no definition of an agent, no usage volume and no retention attached.

The one pressure here that is not vendor-authored is the regulator, and it is moving the wrong way for anyone planning to push more content through the existing gate. FDA’s Office of Prescription Drug Promotion posted nine untitled letters and one warning letter in the first quarter of 2026, against one such letter in all of 2023 and none in 2024, by Covington’s count. FDA has separately said it is “already implementing AI and other tech-enabled tools to proactively surveil and review drug ads.” Those are two facts and not one, since Covington’s review mentions AI nowhere, and the directional read is enough on its own: the gate you were planning to run ten times the volume through is getting more expensive to be wrong at.

Who feels the ceiling first, and how it reads three layers up

None of this reaches your desk in the form I have just written it. It arrives as four unrelated complaints from four functions across a year, each already translated, by the time you hear it, into a personnel problem.

This is the part most consulting work will not put in writing. The Retrofit Ceiling makes a small number of senior operators newly accountable for things they never signed up for. The Chief Data Officer, hired to govern and protect data, becomes accountable for pipeline velocity. The MLR head, hired to keep the company out of a warning letter, becomes accountable for commercial throughput in the exact quarter the regulator is raising the price of a mistake. The brand director, hired and compensated on a campaign, becomes accountable for a system that runs continuously and has no launch date to be judged against. Not one of them was consulted about that shift, and not one of their scorecards moved to reflect it.

Put an AI agenda in front of those three people without explicit air cover and they will read it as a setup, and they will be reading it right. That is not resistance. That is correctly reading the incentive structure they were hired into, which is the same judgment you promoted them for.

The standard institutional answer is a cross-functional steering committee, which is the failure wearing the costume of the fix. Shared accountability is no accountability the moment the work hits friction, and a rebuild is nothing but friction from the second week onward. Three executives sharing a rebuild means it belongs to the seams between them, and when the baton drops into a seam, none of them believes it is their job to pick it up, and all three are right. If that sounded like your last steering committee, the problem is the structure and not your team.

The scoreboard is still empty, and it is still live

That bar is not a high one - a function, a name, a date, which is what any company commits to when it opens a plant. A scoreboard beats a forecast here, because a forecast is a way of postponing the decision until somebody else has made it.

The decision, and the calendar it has to fit

The ceiling has a calendar attached, which is why I am writing this in July rather than October. Budgets go in during the autumn, pilots launch in Q1, results are expected by Q2, revisions land in Q3, and somewhere in that loop the sponsor rotates and a successor inherits commitments they did not author and has no reason to defend. If the only thing you put into the next budget is a larger pilot portfolio, you have chosen retrofit, and you will spend another year inside the operating model that ate the last fifteen pilots.

The commitment that breaks the loop is smaller than a transformation program, and it fits in three lines you can write before the budget closes.

Name the function. Not “commercial.” Medical information operations, or MLR and content, or HCP engagement orchestration, chosen because you can already describe how it would run if you designed it today knowing what the models can do.

Name the executive. One name, not a committee, with at least twenty-four months of expected runway and written air cover to take decision rights off the people who hold them today. That second half is the clause that gets dropped, and dropping it is how the commitment dies quietly in March.

Name the date it has to stand on its own. Not succeed, just stand: running under its own power with the old process switched off. Twelve months is the right shape, because anything shorter is a pilot wearing a new word and anything longer is a steering committee with a budget line.

The cost of waiting is not theoretical and it is not measured in forgone AI value. It is measured in quarters of compounding operating-model debt, carried by the same three operators already holding the ceiling up without the authority to name it, while whoever writes those three lines first spends 2027 building the thing the rest of the industry spends 2028 copying.

So: which function, which name, which date. And who on your leadership team is carrying this right now without the air cover to say so out loud?

If that last question had an obvious answer, book a working session with me.

Sundar

Sources

1. Pfizer, Q4 2025 earnings call, 3 February 2026, remarks of Albert Bourla (LSEG transcript). URL pending.
2. ZS, Scaling AI in pharma and biotech: 2026 CDIO Research, published October 2025; fielded July 2025 by The Harris Poll, n=115 pharma and biotech technology executives. zs.com
3. IQVIA, IQVIA Unveils IQVIA.ai, a Unified Agentic AI Platform Powered by NVIDIA, 16 March 2026. iqvia.com
4. Deloitte, Confidence under pressure: How life sciences leaders are recalibrating for the rest of 2026, published 25 June 2026; fielded April 2026, n=150. deloitte.com
5. NVIDIA Investor Relations, NVIDIA and Lilly Announce Co-Innovation AI Lab to Reinvent Drug Discovery in the Age of AI, 12 January 2026. investor.nvidia.com
6. Sanofi, Sanofi at VivaTech 2026: Scaling AI in Healthcare, 16 June 2026, remarks of Emmanuel Frenehard, Chief Digital Officer. sanofi.com
7. Veeva Systems, Q1 FY2027 Earnings Prepared Remarks, Peter Gassner, 3 June 2026. veeva.com
8. Veeva Systems, acquisition of Copli and launch of Veeva Falcon MLR, 23 June 2026. URL pending.
9. Covington & Burling, FDA Advertising and Promotion Enforcement Activities: Update, 20 April 2026. cov.com
10. U.S. Food and Drug Administration, FDA Launches Crackdown on Deceptive Drug Advertising, 9 September 2025. fda.gov
11. BioSpace, Lilly Doubles Down on Insilico’s AI Medicines, Bets up to $2.75B, March 2026. biospace.com