Drugmakers moved AI into enterprise systems and data exchanges while AI rules stayed in draft

Facing patent expiries and the cost of R&D, large drugmakers spent the second quarter buying late-stage assets for cash and buying AI as enterprise software and milestone-weighted options. Regulators built their own AI and data programs, and the FDA lost several of the officials who launched them.

Editorial snapshot: 31 July 2026

The central finding

AI in pharma became an enterprise and data-access commitment in Q2 2026, ahead of clinical evidence and binding rules. Until the FDA’s draft guidance and the EU’s draft Annex 22 are final, existing GxP rules on validation and change control govern how those deployments can be used.

01Key findings

What changed for AI in pharma and biotech in the second quarter

Drugmakers, regulators and investors each made larger AI commitments in the quarter. The clinical evidence and the rules behind those commitments moved more slowly.

Capability

AI moved into data access and real-time trial data

The FDA received and validated real-time signals from an AstraZeneca trial, and Lilly offered about 100 smaller biotechs models trained on decades of its research data in exchange for their own.

Capital

Isomorphic’s round dominated U.K. biotech venture funding as deal counts stayed low

Isomorphic Labs’ $2.1 billion Series B was about three-quarters of U.K. biotech venture funding in the quarter, while the 366 disclosed biotech venture rounds in the first half were near a decade low.

Adoption

Drugmakers bought AI at enterprise scale, without reported returns

Bristol Myers Squibb agreed to give more than 30,000 employees access to Claude, and Lilly had signed about 20 AI licensing deals by May. None of the quarter’s deals or deployments reported an effect on margin, price or capacity.

Accountability

FDA lost the officials behind its AI programs, and no AI rule became binding

The commissioner, the chief AI officer and the acting head of the drug center left in May, weeks after the FDA launched real-time clinical trials. The FDA’s AI guidance and the EU’s Annex 22 for manufacturing remained drafts.

Operating question

Which of your AI uses affect product quality, data integrity or patient safety, and can each be validated, version-locked and owned by a named person under the GxP rules that apply today?

02Capability

What AI made practical in drug development this quarter

The quarter’s capability changes were organizational. Regulators began receiving trial data continuously, a drugmaker turned decades of research data into a shared model service, and another committed to one vendor’s models across research, manufacturing and commercial work. These announcements concerned data access and deployment; none reported a measured gain in molecular-design accuracy.

Clinical trials · Proof of concept

Trial data can reach the regulator as it is generated

The FDA said it received and validated real-time signals from AstraZeneca’s Phase 2 TRAVERSE trial in mantle cell lymphoma, carried by Paradigm Health’s platform. Amgen’s Phase 1b STREAM-SCLC trial in small-cell lung cancer is the second proof of concept, and a pilot for early-phase trials is planned.

Discovery · Operating program

Proprietary research data traded for model access

Lilly’s TuneLab gives about 100 smaller biotechs access to models trained on decades of Lilly research data, using federated learning, in exchange for their own data. Incyte supplies proprietary experimental data to Genesis’ GEMS platform for structure and property prediction under their expanded collaboration.

Enterprise · Announced May 20

One vendor’s models across the value chain

Bristol Myers Squibb plans to use Claude for target identification, clinical study reports, regulatory submissions, manufacturing root-cause analysis, CAPA (corrective and preventive action) documentation, data-driven batch release decisions and engagement with health care professionals, connected to institutional knowledge under governance and audit controls.

Safety science · Sandbox announced June 9

Predictive safety models under regulatory supervision

The MHRA’s AI sandbox will test up to five approaches that predict pharmacokinetics, toxicity, adverse reactions and effects across patient subgroups, partly to reduce reliance on animal testing. Partner work was scheduled to begin in summer 2026.

Frontier models · Monthly releases

Frontier models arrived monthly, and OpenAI’s flagship doubled in price

OpenAI released GPT-5.5 in April at twice its predecessor’s list price, and OpenRouter measured 49% to 92% higher effective cost for the same users. Anthropic released Claude Opus 4.8 on May 28 and Claude Fable 5 on June 9, and OpenAI previewed GPT-5.6 on June 26.

Cost and consequence

No drugmaker disclosed what its model contracts cost. For validated work, the cost that recurs is change: draft EU Annex 22 makes retesting the default after configuration changes, and the FDA’s draft guidance expects manufacturing models to be maintained through change management, so each model replacement repeats part of the validation work.

03Capital

AI capital concentrated in one company while typical biotech rounds stayed small

Biotech venture dollars rose in the first half while the number of rounds stayed near a decade low, and one AI company shaped the quarterly totals it entered. Pharma’s own cash went to late-stage and China-originated assets, and discovery platforms were bought through milestone-weighted deals.

Disclosed transactions, Q2 2026

Isomorphic’s round and Incyte’s Genesis payment beside the quarter’s two largest acquisitions

Disclosed amounts in USD millions, April to June 2026. Acquisition amounts are upfront cash and equity; the instruments differ and are labeled.

AbbVie/Apogee

Acquisition — $10.9B

Announced within 14 days of GSK/Nuvalent.

GSK/Nuvalent

Acquisition — $10.6B

Isomorphic Labs

Series B — $2.1B

Led by Thrive Capital; no clinical program disclosed.

Incyte/Genesis

Collaboration upfront — $120M

$80 million cash and $40 million equity; up to $232 million in milestones per program, plus royalties.

J.P. Morgan and DealForma Q2 2026 report; Isomorphic Labs; Pharmaceutical Executive.

366

Disclosed biotech venture rounds, first half of 2026

Against 371 a year earlier and 663 in the first half of 2021; median round $23.7 million. BioCentury data.

7 of 14

Rounds of $200 million or more raised by biotechs that describe AI as core to their strategy

Those biotechs were about 24% of companies disclosing round sizes in the half. BioCentury data.

Median venture round by development stage, first half of 2026

Disclosed biopharma venture rounds compiled by DealForma for J.P. Morgan: $16.3 billion across 235 rounds in the half.

Median venture round by development stage, first half of 2026
Stage Median round
Platform or discovery Median round $22 million
Preclinical Median round $29 million
Phase 1 Median round $40 million
Phase 2 Median round $63 million
Phase 3 Median round $73 million

Concentration

Isomorphic dominated the quarterly totals it entered

Its $2.1 billion round was £1.55 billion of the U.K.’s £2.05 billion in second-quarter biotech venture funding and 37% of CB Insights’ quarterly digital-health total. Without it, U.K. venture funding was £498 million, against £279 million a year earlier.

Measurement

Totals depend on what a tracker counts

DealForma counted $9.2 billion of biopharma venture funding in the quarter. TD Cowen’s tracker counted $34.3 billion raised by biotech companies, its best quarter in almost two years, in a total not broken out by financing type. DealForma’s figure covers venture rounds only.

Patent cliff

Pharma’s cash went to late-stage and China-originated assets

J.P. Morgan tied oncology licensing to “large pharma’s need to replace revenue caused by upcoming patent cliffs.” First-half licensing totaled $166.7 billion in deal value, about 6% of it upfront. China-originated assets made up 42% of large-cap licensing deals with at least $50 million upfront and 68% of those upfront dollars.

Deal structure

Discovery platforms are bought as options

Big pharma signed 41 platform and discovery in-licensing deals in the first half, at a median upfront of $62 million. Incyte’s expansion with Genesis follows the pattern: $120 million upfront against potential value of more than $1 billion across the first five targets, most of it contingent on milestones.

What it signals

Money for AI discovery arrived as a few very large rounds and milestone-heavy partnerships, while cash for clinical assets moved at far larger scale. For a typical discovery biotech, an AI strategy did not lift round size: median rounds rose with clinical stage, and AI-describing biotechs’ median trailed their peers’.

04Products and delivery

How drugmakers and regulators are sourcing AI work

Delivery moved in three directions. Pharma bought horizontal model capacity and pushed it into regulated workflows, a drugmaker became a platform supplier to smaller companies, and regulators built their own data pipes and review tools. Each puts integration with existing trial, laboratory and quality systems at the center of the work.

Build or buy

Three ways companies are sourcing AI

Bristol Myers Squibb bought enterprise access to one horizontal vendor, Anthropic. Lilly built infrastructure, including its LillyPod system of 1,016 NVIDIA Blackwell GPUs, and had signed about 20 AI technology licensing deals by May, the most of any drugmaker on UBS’s count. Recursion and Insilico keep discovery through clinical development in-house.

Platform model

A drugmaker as supplier to smaller biotechs

Lilly’s TuneLab trades model access for data, and Dave Ricks likened it to an app store for biotech scientists. Lilly announced on May 20 that TuneLab would be integrated into Collaborative Drug Discovery’s CDD Vault, and Charles River joined to provide nonclinical testing to participating companies.

Regulators

Regulators building their own systems

The FDA said it had consolidated 40 application intake systems, three data-monitoring systems and seven adverse-event reporting systems into one platform each, saving at least $120 million a year. It also said new models reach its staff in about a week and that it uses Gemini and Claude.

Programs launched or contracted in the quarter

Programs launched or contracted in the quarter
Program DateStatus at June 30Disclosed terms or results
FDA real-time clinical trials Date April 28 Status at June 30 Two proof-of-concept trials; pilot planned Disclosed terms or results Signals from one trial received and validated
FDA one-day inspection pilot Date May 6 Status at June 30 Runs through fiscal 2026 Disclosed terms or results 46 one-day assessments; AI method undisclosed
Bristol Myers Squibb and Anthropic Date May 20 Status at June 30 Enterprise agreement announced Disclosed terms or results No terms or outcomes disclosed
Lilly TuneLab in CDD Vault Date May 20 Status at June 30 Integration announced Disclosed terms or results No terms disclosed
Incyte and Genesis Date May 20 Status at June 30 Expanded collaboration Disclosed terms or results $120 million upfront Up to $232 million in milestones per program
MHRA AI sandbox Date June 9 Status at June 30 Announced; partner work from summer Disclosed terms or results Up to five approaches; funding not stated

Dates are 2026 announcement dates.

Integration

In each case the engineering work is integration: trial-site data capture, laboratory and quality systems, and regulatory document management. Bristol Myers Squibb described connecting Claude to knowledge held in disconnected systems but named no integrations, and none of these announcements reported an outcome.

05Adoption and economics

Adoption by segment shows in contracts more than in surveys

Most of what can be observed about adoption in this industry is transactional: who signed what with whom, and how many people a deployment covers. Each row below uses its publisher’s own measure, so together they show the shape of adoption by segment rather than a single rate.

AI adoption signals by segment

Each row uses its publisher’s own measure and population; the rows are not comparable with one another.

AI adoption signals by segment
Signal SegmentValuePopulation and period
Biotechs describing AI as core to strategy Segment Discovery-stage biotech Value About 24% Population and period First-half 2026 financings; BioCentury
Biotechs using Lilly’s TuneLab Segment Discovery-stage biotech Value About 100 Population and period June 2026; Lilly via the Financial Times
Venture rounds to companies with a drug in human testing Segment Clinical-stage companies Value 42, about two-thirds Population and period First half of 2026; BioPharma Dive tracker Companies backed by 26 tracked investors
AI technology licensing deals signed by Lilly Segment Established pharma Value About 20 Population and period As of May 2026; Lilly and UBS via the Financial Times
Bristol Myers Squibb employees to receive Claude Segment Established pharma Value More than 30,000 Population and period Announced May 20; deployment plan
Platform and discovery in-licensing deals by big pharma Segment Established pharma Value 41 Population and period First half of 2026; DealForma Median upfront $62 million
FDA staff using generative AI Segment U.S. regulator Value More than 80% Population and period April 2026; FDA officials Some centers above 90%; about 1% in early 2025
AI use cases across the EU medicines regulatory network Segment EU regulator Value 61 Population and period 2025; EMA AI Observatory

Discovery-stage biotech

An AI strategy no longer sets a discovery biotech apart

AI has become common among financed biotechs without lifting the typical round. Incumbents now offer models trained on their own data in exchange for a biotech’s data, which makes data-sharing terms part of the price of model access.

Clinical-stage organizations

Clinical-stage companies drew most tracked venture rounds

About two-thirds of tracked first-half venture rounds went to companies with a drug in human testing; the count does not separate AI-derived programs. Sponsors entering the FDA’s real-time trial pilot would need endpoints and signals defined in advance and validated, near-real-time data flows from sites, and the program’s durability after the May departures is uncertain.

Established pharma

Large drugmakers moved from pilots to enterprise governance

Bristol Myers Squibb’s agreement spans GxP and non-GxP work, and Lilly expects 2026 R&D spending in the high teens of billions of dollars, up from $13 billion in 2025. The question has moved to governing one vendor’s models across regulated and unregulated functions, a line draft Annex 22 would sharpen.

Service and manufacturing partners

CROs, CDMOs and platforms as the integration layer

Charles River in TuneLab, Paradigm Health in the FDA’s real-time trials and CDD as a TuneLab host sit between AI platforms and wet-lab or clinical validation. Contract manufacturers running machine-learning models in critical GMP steps for sponsors would carry draft Annex 22’s obligations if it is adopted as drafted.

Economics

Contract values stayed private, and no deal or deployment reported an effect on margin, price or capacity. The public case for the spending rested on speed claims, such as the FDA’s 40% trial-time estimate and Recursion’s discovery timelines, rather than on reported returns.

06Outcomes

Results from AI-derived drug programs remained early-stage

The quarter’s clinical results for AI-derived molecules came from company releases and sat early in development: a Phase 2 readout, early safety data, a first patient dosed and an investigational new drug (IND) clearance. The FDA’s process results were reported by officials without published methods.

Phase 2 · Company-reported

Recursion’s polyp-burden data for REC-4881

In familial adenomatous polyposis, 40% of patients improved in Spigelman stage, and 15.8% had Grade 3 treatment-related adverse events, with none higher. Recursion is discussing a registrational design with the FDA and expects to give an update in the second half of 2026.

Limitation: Measured against baseline; no control arm was stated.

−43%Median change in polyp burden, week 13

−53%Week 25, after a treatment break

Phase 1/2 · Company-reported

Early safety data for REC-1245

In the DAHLIA trial of Recursion’s RBM39 degrader, no patient had a dose-limiting toxicity and the maximum tolerated dose was not reached. Grade 3 treatment-related adverse events occurred in 6.2% of patients.

16Patients treated

62.5%Any treatment-related adverse event

Preclinical · Company comparison

Recursion’s discovery-speed claim

Recursion says REC-1245 reached development-candidate status more than twice as fast as the industry average, without disclosing the benchmark, and reports about 20 months for REC-4539, whose first trial patient was dosed in April.

18 monthsDiscovery to development candidate, REC-1245

IND · Company-reported

Insilico’s inhaled rentosertib cleared in China

Insilico announced on April 28 that China’s drug regulator had cleared its investigational new drug application for an inhaled form of rentosertib, a TNIK inhibitor for idiopathic pulmonary fibrosis. The Phase 1 trial will test safety, tolerability and pharmacokinetics.

13thIND from Insilico’s AI-driven pipeline

About 80Planned Phase 1 participants

Process · Agency-reported

FDA’s administrative time claim

FDA officials said early generative AI pilots cut the time for some administrative tasks. No method or list of tasks was published with the claim.

10 days to 20 minutesSome administrative tasks in early pilots

Stage of evidence

Among these programs, the most advanced result was a company-reported Phase 2 readout measured against baseline. The results show AI-derived molecules producing ordinary early clinical data; they do not yet bear on whether AI improves success rates.

07Corrections

Four claims from the quarter and what the record supports

Each claim below circulated during the quarter or about it. The corrections change what a sponsor can plan on.

Trial duration · Official estimate

The 40% figure was an official’s estimate at a pilot launch, with no published analysis.

Claim under review: The FDA says real-time clinical trials will cut trial time by up to 40%.

The FDA’s chief AI officer, Jeremy Walsh, said real-time trials could cut up to 40% of trial time, and Commissioner Marty Makary said about 45% of the time from Phase 1 to submission is dead time. Neither figure came with a published analysis, and both men left the agency in May.

Regulation · Still in draft

No AI-specific rule for drug development became binding in the quarter.

Claim under review: Regulators have finalized rules for AI in drug development.

The FDA’s January 2025 draft guidance on AI to support regulatory decisions was still listed as a draft on May 1, 2026. The EU’s draft Annex 22 on AI in manufacturing had no finalization date at the Irish regulator’s May 12 briefing, and the FDA-EMA guiding principles of January 2026 are principles, not formal guidance.

Trial approvals · Misattributed

The measured gain came from 2023 process reforms, before AI tools were introduced.

Claim under review: The MHRA halved clinical trial approval times with AI.

The MHRA’s October 2025 release reported approval times falling from 91 to 41 days across 4,617 applications assessed between September 2023 and August 2024, under its 2023 risk-proportionate approach. Its text described AI tools as “now being introduced,” so the measured period predates them.

AI drug programs · Not disclosed

Isomorphic named no clinical program or trial timeline with its round.

Claim under review: Isomorphic Labs’ $2.1 billion round funds drugs already in human trials.

The May 12 announcement described advancing programs toward the clinic. A first-in-human target of the end of 2025 was revised in January to a Phase 1 start before the end of 2026, according to reports of Demis Hassabis’ remarks at Davos.

08Accountability

Where AI programs proved fragile in the quarter

The quarter’s visible failures were institutional: programs that depended on officials who left, a risk model whose workings were not disclosed, a model withdrawn days after release and a security weakness in agent tooling. In each case responsibility stays with the sponsor or manufacturer whose records the AI touches.

Incident · April and May

FDA’s AI programs lost the officials who launched them

Vinay Prasad, director of the biologics center, left in April. Commissioner Makary resigned May 12, and Tracy Beth Hoeg, acting director of the drug center, and chief AI officer Jeremy Walsh left in May, weeks after real-time trials launched. Mintz said the AI officer role “now seems rudderless,” and Skadden warned that informal policies could be reversed or challenged.

Risk · Pilot from May 6

An unexplained AI risk model now sets inspection depth

The FDA’s one-day inspection pilot uses AI to classify facilities as low-risk and had produced 46 one-day assessments; it runs through fiscal 2026. Skadden noted that the FDA has given little detail on how the classification works, so manufacturers cannot see what drives it.

Incident · June 12

A generally available model was withdrawn after three days

Anthropic disabled Claude Fable 5, generally available since June 9, and Claude Mythos 5 for all customers after a U.S. Commerce Department export-control directive barring foreign-national access. Greenberg Traurig advised checking model aliases, routing, fallback settings and deemed-export exposure. Under GMP-style change control, a forced switch to another model needs impact assessment and retesting.

Risk · Demonstrated attacks

Agents act on the content they can reach

Researchers demonstrated a design flaw in Model Context Protocol software development kits affecting an estimated 200,000 instances, and an injection technique that hijacked major coding agents with an 85% success rate in testing. Deployments that connect agents to institutional knowledge and run Claude Code, as Bristol Myers Squibb described, make text in tickets, logs and documents a possible instruction channel into systems holding GxP records.

Reading results

Recursion tied lower spending to data timing and portfolio reprioritization

Recursion’s first-quarter R&D expense fell to $87.9 million from $129.6 million, which it attributed mainly to the timing of Tempus data purchases and operating efficiencies; revenue fell to $6.5 million from $14.7 million. It tied lower operating cash use to efficiencies and a strategic reprioritization of its clinical portfolio, and lower administrative costs to lower salaries.

Who is accountable

Responsibility for validated outcomes stays with the sponsor or manufacturer, whatever happens to the FDA’s programs or a vendor’s model list. Model versions, fallbacks and agent permissions belong in the validation record alongside the data.

09Regulation

AI rules for drug development stayed in draft in the U.S., EU and U.K.

The FDA’s draft guidance, the FDA-EMA principles and draft Annex 22 share a risk-based logic built on context of use and human accountability, and none became binding in the quarter. The FDA ran its policy through press releases and pilots, the EMA documented practice, and the MHRA built a supervised test space. The enforceable constraint remained the existing GxP framework of validation, data integrity and change control.

United States · Draft since January 2025

FDA’s AI credibility framework is still a draft

The guidance covers AI that informs decisions on safety, effectiveness or quality from nonclinical work to manufacturing. It excludes drug discovery and operational uses, such as drafting submissions, that do not affect patient safety, drug quality or study reliability. Its seven steps run from the question of interest and context of use through model risk and a credibility plan to an adequacy determination.

United States · Comments closed June 29

Real-time trials are a pilot, not guidance

The FDA’s request for information on the pilot closed June 29 after an extension, with final selection criteria scheduled for July and selections for August. Skadden said the FDA may revert to requiring two pivotal trials in most cases, which would reverse the single-trial default announced in February.

European Union · Consultation closed Oct. 7, 2025

Draft Annex 22 would keep LLMs out of critical GMP use

In QbD Group’s reading of the July 2025 draft, machine-learning models used where product quality, data integrity or patient safety is at stake must be static and deterministic, and large language models are unsuitable. Test data must be independent of training data, acceptance criteria must be set before testing, and retesting is the default after changes.

European Union · Report posted June 4

The EMA documented AI use and pointed to existing GMP controls

The EMA’s 2025 AI Observatory said existing GMP guardrails already mitigate many general AI risks and listed what applicants are bringing next, including generative AI for drafting submissions, clinical-outcome prediction, digital endpoints and predictive stability modeling. It noted that the proposed European Biotech Act would require further AI guidance.

European Union · Pending at June 30

AI Act high-risk dates moved, pending final adoption

A provisional agreement on May 7 would move high-risk obligations for Annex III uses to Dec. 2, 2027, and for AI in products under sectoral laws such as the EU medical device regulations to Aug. 2, 2028. Parliament adopted it June 16; with Council adoption and publication pending, Aug. 2, 2026, remained the legal baseline at quarter end.

Scope: Most AI used in drug research is not high-risk under the Act; AI-enabled devices and diagnostics used in trials can be.

United Kingdom · Announced June 9

The MHRA opened a sandbox for predictive safety models

Developers of AI that predicts pharmacokinetics, toxicity and adverse reactions will work directly with the regulator from the summer. The industry body ABPI called it a possible step toward the safe harbor it has sought and asked for alignment with the U.K.’s Centres of Excellence in Regulatory Science and Innovation.

Dated and pending items, as scheduled at the end of June

Dated and pending items, as scheduled at the end of June
Item DateJurisdictionStatus at June 30
FDA real-time trial RFI comments Date June 29, 2026 Jurisdiction U.S. Status at June 30 Closed after extension from May 29
FDA real-time trial pilot selection criteria Date July 2026 Jurisdiction U.S. Status at June 30 Scheduled in April
EU AI Act remaining obligations Date Aug. 2, 2026 Jurisdiction EU Status at June 30 Legal baseline while the Omnibus is pending
FDA real-time trial pilot selections Date August 2026 Jurisdiction U.S. Status at June 30 Scheduled in April
MHRA sandbox partner work begins Date Summer 2026 Jurisdiction U.K. Status at June 30 Scheduled
FDA one-day inspection pilot ends Date Sept. 30, 2026 Jurisdiction U.S. Status at June 30 Runs through fiscal 2026
Final EU GMP Annex 22, Annex 11 and Chapter 4 Date No date Jurisdiction EU and PIC/S Status at June 30 Not final at May 12
Final FDA guidance on AI for drugs and biologics Date No date Jurisdiction U.S. Status at June 30 Draft at May 1
AI Act Annex III and Annex I high-risk obligations Date Dec. 2, 2027, and Aug. 2, 2028 Jurisdiction EU Status at June 30 If the Omnibus is adopted
New EU pharmaceutical legislation applies Date 2028 Jurisdiction EU Status at June 30 Agreed December 2025

What applies now

Until these drafts are final, AI in regulated work answers to existing GxP obligations for computerized-system validation, data integrity and change control. Draft Annex 22 shows where manufacturing expectations are heading: fixed, deterministic models, independent and access-logged test data, and defined human roles.

10Voices

Where regulators, drugmakers, vendors and analysts agree and differ

Three positions recur across constituencies. Access to proprietary experimental data is the binding constraint, a view shared by Lilly’s Dave Ricks, Bristol Myers Squibb’s Greg Meyers and CDD’s Barry Bunin, and expressed in Incyte’s deal terms. Evidence standards must be set before AI predictions replace experiments, as the MHRA’s sandbox advisers and the draft rules hold. The disagreements mostly describe different segments: AI-native vendors describe platform-scale change, while Ricks, whose company owns wet labs and clinical capacity, says AI-made drugs are not yet a reality. On trial time, the FDA’s departed leaders located the delay in data latency between phases; Skadden and PitchBook located the near-term risk in regulatory instability.

Regulators

The FDA’s departed leaders framed AI as a fix for trial latency; U.K. and EU voices framed it around evidence standards.

Marty Makary

Commissioner, FDA (resigned May 12)

“For 60 years, we’ve been conducting clinical trials in the same way”

Said at the April 28 launch of real-time clinical trials; he resigned two weeks later.

Regulator · launch remarks

Kyle Diamantas

Acting Commissioner, FDA

Pledged to minimize political interference and to listen to career scientists.

A lawyer and former deputy commissioner for human foods, he took over after Makary’s resignation.

Regulator · as reported by Skadden

Lord Vallance

U.K. Science Minister

“Too many promising medicines fail late in development”

His case for the MHRA sandbox on predictive safety and pharmacokinetic models.

Government · MHRA release

Alastair Denniston

Chair, National Commission into the Regulation of AI in Healthcare

The sandbox should establish what good evidence looks like for AI that predicts drug safety and pharmacokinetics.

An NHS doctor; his position anchors the call for evidence standards before predictions replace experiments.

Independent commission · MHRA release

European Medicines Agency

EU regulator

“AI was moving from concept to practice”

The same report said existing GMP guardrails mitigate many general AI risks.

Regulator · 2025 AI Observatory

Drugmakers and buyers

The most active buyers describe data access as the constraint and are cautious about claims of autonomous drug discovery.

Dave Ricks

Chair and CEO, Eli Lilly

Doubts that AI companies are yet making drugs with computers and says AI drug discovery is in its infancy, while promoting TuneLab, which he likened to an app store for biotech scientists.

The quarter’s clearest reservation came from the most active pharma buyer of AI technology.

Incumbent · Financial Times interview

Greg Meyers

Executive Vice President and Chief Digital and Technology Officer, Bristol Myers Squibb

Says most enterprise AI stops at the chatbot; Bristol Myers Squibb’s agreement is meant to put Claude into workflows across functions, with governance and audit controls.

Contract terms were not disclosed.

Buyer · trade press reports

Amy McKee

Senior Vice President, Oncology, AstraZeneca

Wants trial data to move across the ecosystem as quickly as possible.

AstraZeneca’s TRAVERSE trial is one of the FDA’s two real-time proofs of concept.

Sponsor · reported remarks

AI vendors and platforms

Vendors describe scale and breadth; their statements are commercial positioning.

Demis Hassabis

CEO, Isomorphic Labs

“a massive vote of confidence from a diverse group of top-tier international investors”

His framing of the $2.1 billion round rested on investor breadth; the release named no clinical timeline.

Vendor · company release

Eric Kauderer-Abrams

Head of Life Sciences, Anthropic

Described Claude as a single intelligence layer for pharma companies.

The positioning behind the Bristol Myers Squibb agreement.

Vendor · trade press report

Barry Bunin

CEO, Collaborative Drug Discovery

Says the TuneLab model depends on collaboration that gives broad access to research data.

Lilly and CDD announced on May 20 that TuneLab would be integrated into CDD Vault.

Platform partner · trade press report

Analysts and advisers

Analysts locate the near-term risks in capital concentration and the durability of FDA policy.

Lindsay Martin

Biopharma Analyst, BioCentury

“three times as likely to top $200M”

Describing AI biotechs’ financings in the first half; her analysis also found their median round trailed non-AI peers’.

Independent analyst

Ben Zercher

Senior Analyst, PitchBook

Said venture investment stayed strong in the first half, but that turnover at the FDA created uncertainty that could lead some investors to wait before sizing their bets.

Speaking to BioPharma Dive about first-half venture funding, which held up despite a threatened U.S. crackdown on investment in Chinese drug assets.

Independent analyst · reported remarks

Rachel Turow

Skadden, Arps, Slate, Meagher & Flom

“current program materials are best read as indicative rather than authoritative”

Written after the May departures, about the FDA’s real-time trials and related programs; law firms advise sponsors on FDA strategy.

Law firm · client alert

11Decisions

Three decisions for pharma and biotech leaders

Chief scientific officer

Decide which proprietary data to trade for model access

TuneLab and the Incyte-Genesis deal both price model access in experimental data. The terms decide who learns from your results and whether a platform’s models improve on your data for others.

Which datasets are we willing to share, under what controls, and what would we gain that we cannot build ourselves?

Head of quality

Decide where probabilistic models may sit in GxP processes

Bristol Myers Squibb lists batch release and CAPA documentation among its Claude uses, while draft Annex 22 would keep large language models out of critical GMP applications. Classifying uses now limits rework if the draft is adopted as written.

Which planned AI uses affect product quality, data integrity or patient safety, and can each run on a fixed, validated model with a defined human decision-maker?

Head of clinical development

Decide whether to build real-time trial data flows without a stable FDA program

Taking part in the FDA’s pilot would require signals defined in advance and validated data flows from sites, and the officials behind it left in May. The same capability serves a sponsor’s own trial monitoring, whatever happens to the pilot.

Which of our trials would justify pre-specified real-time signals on their own merits?

12Actions

Four actions for the second half of 2026

  1. 1

    Map AI uses against draft Annex 22’s critical-use test

    List every current and planned AI use in GMP work with its model type and the decision it informs. The result is a register that separates critical uses from support uses before the final text arrives.

    Head of quality

  2. 2

    Make model versions and fallbacks part of the validation record

    For each validated workflow that calls an external model, record the pinned version, the approved fallback and the retest a switch triggers, and limit agents to the systems the task needs. The result is a documented response ready for the next withdrawal or injection finding.

    Head of technology

  3. 3

    Keep a dated AI regulatory calendar tied to submission plans

    Track the FDA’s real-time trial criteria and selections, the AI Act Omnibus, Annex 22 and the FDA’s AI guidance as scheduled or pending items, and treat program materials as provisional until they become guidance. The result is a calendar showing which submissions depend on which drafts.

    Regulatory affairs

  4. 4

    Benchmark AI platform terms against the first half’s deals

    Compare proposed discovery-platform deals with the 41 big-pharma deals at a $62 million median upfront and with Incyte’s milestone-heavy Genesis terms. The result is a term-sheet reference that separates upfront cash from contingent value.

    Business development

13References

References

This briefing covers April 1 through June 30, 2026. Sources published after June 30 are cited only for what they report about that period.

Regulators and agencies

Clinical and research results

Models and security

Thoughtful engineering for what’s next.

Have a question about what this means for your team? Let’s talk it through.

Start a conversation

Have a project in mind, or just exploring an idea? Let’s talk.
Fill out the form or give us a call.

413-570-3299

What can we help you with?