What eight years of pharma AI announcements add up to

The AI in Pharma Observatory tracks public artificial intelligence initiatives in the pharmaceutical industry. Each entry records who announced what, at which stage of the pipeline, with how much money attached where a figure was disclosed, and what status the initiative has reached. Every entry carries a source: a press release, a regulatory document, a filing or a paper.

There are now 246 entries from 131 organisations, the earliest dated April 2018. Sixty-eight of them fall in the first eight months of 2026, more than in the whole of 2024.

Two of the fields, the disclosed amount and the status, do not point in the same direction when sorted by pipeline stage. That divergence is what this article is about.

Where the disclosed amounts concentrate

Fifty-five of the 246 initiatives came with a public figure. Together they represent 38.2 billion dollars.

Thirty-two of those billions, 84 percent, are attached to drug discovery. The four largest are all discovery collaborations: XtalPi with DoveTree at up to 6 billion, AstraZeneca with CSPC at up to 5.3 billion, Novo Nordisk with Valo Health at around 4.6 billion, and Eli Lilly with Insilico Medicine at up to 2.75 billion.

The rest of the pipeline is quieter. Clinical trials account for 262 million dollars across three deals. Patient and real-world evidence work accounts for 372 million across three. Commercial and medical affairs, the functions in daily contact with prescribers, account for 750 million in a single transaction. No manufacturing initiative in the dataset carries a disclosed figure at all.

One qualifier matters before reading anything into the 38.2 billion. Of the 55 disclosed figures, 25 are total potential deal value including milestones, 21 are funding rounds, and only six are money paid at signature. The headline number is largely a ceiling on what could be paid over a decade, not capital that has changed hands.

What the status field shows

Every entry also carries a status: announced, pilot, deployed, or cleared by a regulator. Sorted by stage, the ranking does not follow the money.

Discovery holds 112 initiatives, of which 31 percent are marked deployed or cleared. Clinical trials hold 46, at 33 percent. Regulatory work holds 39, at 72 percent. Manufacturing and CMC holds seven, five of them in production.

The regulatory figure needs unpacking, because it is easy to misread. Twenty-three of those 39 entries come from regulators themselves: FDA guidance documents, the HMA-EMA artificial intelligence workplan, the FDA list of AI-enabled devices passing a thousand clearances. A published guidance is complete by definition, which inflates the rate. Among the sixteen regulatory-stage entries that come from companies rather than agencies, eleven are deployed or cleared. Still high, on a much smaller base.

Manufacturing is a smaller and cleaner case. Seven entries, all from named pharmaceutical companies, Amgen, Sanofi, Pfizer and Merck KGaA, five of them running in production. Seven initiatives is too few to carry a general claim, and worth noting for exactly that reason.

Across the whole dataset, 129 initiatives are marked announced against 87 deployed, and 114 of the 246 are partnerships against 50 product launches.

Three ways to read the same distribution

The gap between where the money is announced and where things are running has more than one credible explanation, and they are not mutually exclusive.

The first is about how work gets contracted. Discovery is bought from outside, through partnerships with biotechs and AI companies, and an external partnership produces a signed deal with a number that both parties want to publish. A regulatory writing assistant or a medical information triage tool is bought as a software licence, or built internally by a data team. Neither generates a press release with a figure. On this reading, the 84 percent measures contracting practice rather than investment, and the downstream spend exists but is invisible.

The second is about the size of the prize. A molecule that reaches the market is worth billions over its patent life. A regulatory submission produced three weeks faster saves staff time. Even if AI works more reliably downstream, the arithmetic upside upstream is larger by orders of magnitude, so concentrating capital there can be a rational allocation rather than a misallocation.

The third is about problem difficulty. Deployment rates are high in regulatory and manufacturing partly because those tasks are narrow, well specified and repetitive, which is what current models handle well. Discovery is attacking a harder problem earlier in its maturity curve, and a 31 percent deployment rate on a hard problem is not obviously worse than a 70 percent rate on an easy one.

Which of the three carries the most weight is not something this dataset can settle.

What the data does not capture

The observatory tracks what organisations choose to say in public. That limitation has a direction: it makes internal, unannounced work invisible, and internal work is concentrated downstream. A medical affairs team using a language model to triage information requests will not issue a press release. A manufacturing site tuning a vision model on a packaging line rarely does either.

So the dataset is a reliable record of industry communication and an incomplete record of industry practice. The distribution above describes what the sector funds visibly and talks about, which is a real object worth measuring, but it is not a census of deployed artificial intelligence in pharma.

Two smaller caveats. Stage attribution is a judgement call for initiatives that span several functions, and 68 entries are classed as cross-cutting for that reason. And regional coverage leans towards English-language sources, with 132 entries in North America against 82 in Europe and 29 in Asia, which probably understates activity in markets we read less well.

What would move the picture

The discovery thesis becomes checkable when AI-originated molecules produce clinical readouts, and the first of those have started arriving. Insilico Medicine published Phase IIa results for rentosertib in Nature Medicine in June 2025. Absci began first-in-human dosing of an AI-designed antibody in March 2025. Generate Biomedicines reported early Phase 1 efficacy data in COPD in August 2026, and BigHat Biosciences dosed its first patient in September 2026.

Ten such readouts, spread across companies and targets, would say more about whether the 32 billion dollars was well placed than any number of new partnership announcements. The dataset currently holds ten entries classed as clinical results, against 114 partnerships.

The other thing worth watching is whether the downstream stages start producing disclosed figures at all. If commercial, medical and manufacturing work stays at one or two disclosed transactions a year while genuinely spreading inside companies, that tells us the announcement gap is a communication artefact. If it stays quiet because little is happening, that is a different finding. Today the observatory cannot distinguish between the two, and saying so seems more useful than picking one.