Pharma's AI money went to compute and preclinical licenses as FDA and EMA set joint principles

Large drug companies, with more than $330 billion of revenue facing loss of exclusivity over the next five years, spent the first quarter of 2026 buying late-stage assets. Their AI commitments arrived as compute, milestone-heavy discovery licenses and agents inside existing software, while U.S. and European regulators began defining what AI-supported evidence must show.

Editorial snapshot: 30 April 2026

The central finding

AI advanced molecule design and document work in the first quarter of 2026 but did not change what sets trial timelines; pharma paid for AI discovery on milestone-heavy preclinical terms, and regulators built their first joint principles around context of use and human accountability.

01Key findings

What changed in pharma and biotech AI in the first quarter of 2026

Capability, capital, vendor results and regulation pointed the same way in the first quarter. AI's demonstrated value sits early in the pipeline and in document-heavy work, and buyers priced it and regulators governed it on that basis.

Capability

Gains came in design and document work

Isomorphic Labs reported a design engine with more than double AlphaFold 3's accuracy on the hardest protein-ligand cases, and Anthropic connected Claude to trial and chemistry databases. Veeva's CEO said AI might speed trial start-up and close-out, but not by much, because the protocol and human biology set the pace.

Capital

AI arrived as compute and preclinical licenses

Eli Lilly and NVIDIA committed up to $1 billion over five years to a joint lab, and Lilly licensed Insilico candidates for $115 million upfront, with up to about $2.75 billion in total. The large cash went to acquisitions: $40.9 billion committed upfront by J.P. Morgan's count, about $50 billion in total value by William Blair's.

Adoption and economics

Agents are live; AI revenue is not yet material

IQVIA said it runs more than 150 agents across more than 30 use cases, and Veeva's first agents are live with several customers. Veeva prices agents by token and does not expect AI revenue to be material in fiscal 2027; IQVIA disclosed no AI revenue.

Regulation

Regulators set joint principles and a draft limit on generative AI in GMP

FDA and EMA published 10 joint, nonbinding principles on Jan. 14 for AI that generates evidence across a medicine's life cycle. The EU's draft good manufacturing practice (GMP) Annex 22 keeps generative and large language models out of critical manufacturing uses; the final text is expected by the end of 2026.

Operating question

Which of your AI uses produce evidence that a partner, regulator or inspector will rely on, and can you show each one's context of use, validation record and accountable reviewer?

02Capability

Capability moved in molecule design and document work, not in trial duration

The quarter's technical changes clustered at two ends of the value chain: predicting and designing molecules, and handling the documents and data around research and trials. Trials, where human biology and patient recruitment set the pace, gained planning tools. Nothing announced in the quarter changed what determines how long a trial runs.

Discovery · Internal use only

Isomorphic's closed design engine

Isomorphic Labs' Feb. 10 technical report says its IsoDDE engine "more than doubles the accuracy" of AlphaFold 3 on the hardest protein-ligand generalization cases. The model is used only inside Isomorphic and the report does not say it was peer reviewed; outside researchers who praised it cannot inspect it.

2.3xvs. AlphaFold 3, low-homology antibody-antigen cases

19.8xvs. Boltz-2 at DockQ above 0.8, 334 cases

Discovery · Open models

Open models for RNA structure and synthesizability

NVIDIA announced two Clara open models on Jan. 12: RNAPro for RNA structure, and ReaSyn v2, which assesses whether AI-designed molecules can be synthesized, a known weakness of AI design. NVIDIA says features are offered when and if available, and both models produce computational predictions.

Research and trial planning · Generally available

Trial and chemistry data inside a general assistant

Anthropic's Jan. 11 release connected Claude to ClinicalTrials.gov, ChEMBL, Open Targets, bioRxiv and medRxiv, and to Medidata for Medidata's Study Feasibility customers. New skills draft clinical-trial protocols, including endpoint recommendations. The public datasets were already accessible; they now sit inside a general assistant at no stated price change.

Frontier models

Agents became standard in flagship models

OpenAI's GPT-5.4 (March 5) shipped native computer use; Anthropic's Claude Opus 4.6 (Feb. 5) added agent teams in research preview and a 1 million-token context in beta. METR's measurements show long, well-specified tasks increasingly feasible, while Mercor's APEX-Agents benchmark found agents completing fewer than 25% of professional tasks in simulated banking, consulting and law workplaces.

Clinical operations

AI tools reached trial planning

Recursion says AI cut trial country and site selection from months to hours, and IQVIA said large-pharma clients chose it for AI-enabled trial planning and recruitment. These are planning tasks. Neither company reported a change in how long a trial takes.

Cost and consequence

Frontier list prices held or rose in the quarter: GPT-5.4's input price is $2.50 per million tokens, 43% above GPT-5.2's, and Claude Opus 4.6's input price stayed at $5. Agentic work consumes more tokens per task, so the useful budget unit for pharma is the cost of a completed, reviewed work product, including the validation and human review that a token price does not show.

03Capital

Biopharma money went to late-stage assets; AI came as compute and licenses

The AI funding boom of the first quarter largely passed drug developers by. Inside biopharma, money moved toward later-stage companies, and pharma's largest checks bought clinical or near-clinical assets to replace revenue at patent expiry. AI was bought as compute and as preclinical options.

$300B

Global venture funding, Q1 2026

Crunchbase count; AI took $242 billion, or 80%.

$5.1B to $6.9B

Biopharma venture funding, Q1 2026

William Blair (U.S. private financing) and J.P. Morgan/DealForma (biopharma venture); scopes differ.

Q1 2026 biopharma financing by dataset

Each publisher uses its own scope and inclusion rules, so totals are not comparable across rows.

Q1 2026 biopharma financing by dataset
Dataset ScopeTotalDealsTypical round
J.P. Morgan/DealForma Scope Venture, therapeutics and platform companies Total $6.9B (Q1 2025: $8.6B) Deals 101 Typical round Medians $34M-$108M by stage
William Blair Scope U.S. private financing Total $5.1B Deals 62 Typical round Mean $82M; no median given
BioWorld Scope All financings, public and private Total $25.14B (Q1 2025: $13.12B) Deals — No comparable figure Typical round — No comparable figure
BioIndustry Association Scope U.K. biotech equity Total £552M (Q1 2025: £924M) Deals 25 transactions Typical round — No comparable figure

J.P. Morgan counts disclosed figures only and treats tranches of one series as one round. BioWorld counts public and private financings together.

Disclosed M&A

Largest disclosed biopharma acquisitions, Q1 2026

Announced deal value in USD millions. For the quarter, William Blair counts about $50 billion across 20 deals; J.P. Morgan counts $40.9 billion committed upfront across 32 deals, with a median upfront of $1.4 billion.

Gilead/Arcellx

Acquisition — $8.1B

Up to; includes contingent value. Feb. 23.

Lilly/Centessa

Acquisition — $7.8B

March 31.

Merck/Terns

Acquisition — $6.7B

March 25.

Biogen/Apellis

Acquisition — $6.1B

March 31.

Novartis/Synnovation

Acquisition of private company — $3B

March 20.

Lilly/Orna

Acquisition of private company — $2.4B

Feb. 9.

William Blair Q1 2026 U.S. biopharma recap and J.P. Morgan Q1 2026 biopharma deck. Neither described any of these as an AI acquisition.

Early-stage biotech

Seed and Series A money thinned

In J.P. Morgan's DealForma data, later rounds took about twice the money of seed and Series A rounds, and first-time financings were on pace for their lowest annual count since before COVID. Median rounds were $34 million at Phase 1, $73 million at Phase 2 and $108 million at Phase 3; platform and discovery companies' median was $40 million and preclinical companies' $39 million.

$2.3BSeed and Series A, 50 rounds

$4.5BSeries B and later, 51 rounds

Pharma buyers

AI money came as compute and research deals

Lilly and NVIDIA committed up to $1 billion over five years, jointly, in talent, infrastructure and compute for a South San Francisco lab announced Jan. 12. Johnson & Johnson signed a multitarget research collaboration with Isomorphic Labs on Jan. 20, with terms undisclosed.

Licensing terms

AI-derived assets were licensed on the market's usual terms

Lilly licensed Insilico's preclinical oral candidates on March 29, with tiered royalties on top of milestones. The upfront is about 4% of the headline value, close to the share J.P. Morgan reports across all first-quarter licensing, so pharma paid for AI discovery on the same back-loaded terms as other licensing.

$115MUpfront

Up to about $2.75BIncluding milestones

6%Upfront share of all Q1 licensing value

Licensing flows

Asian licensors took about half of partnering upfronts

William Blair counts $81 billion of announced licensing value across 57 partnerships, with Asia-headquartered licensors taking $2.8 billion, about 49% of partnering upfronts. The largest deal, CSPC's with AstraZeneca, carries $18.5 billion in total value and $1.2 billion upfront.

What it signals

Against more than $330 billion of large-pharma revenue facing loss of exclusivity over five years, buyers spent their cash on late-stage assets and treated AI discovery as an option paid out over milestones. For AI-native companies that puts most deal value years away, and early-stage biotechs, AI-based or not, face a thinner seed and Series A market.

04Products and delivery

Agents went inside incumbent systems while general assistants connected from outside

Two delivery models competed for the same workflows. Veeva and IQVIA put agents inside the systems and services pharma already uses and price or bundle them there; Anthropic and OpenAI offered general agents that connect to those systems from outside. Lilly is building its own compute and offering model access, and AI-native companies sold their platforms through partnerships.

Commercial software · Generally available

Veeva's agents ship inside Vault CRM

Veeva released its first AI agents for customer relationship management (CRM) and commercial content in December 2025, and several customers were live by March 4. Veeva says most adoption uses its own agents, with custom agents mainly in lighter helper roles. Ten of the top 20 pharma companies had committed to Vault CRM by Jan. 31, and Veeva expects about 14 eventually. Support for its Salesforce-based CRM is now scheduled to end in December 2029, moved up from about September 2030.

CRO and data services · Deployed

IQVIA runs more than 150 agents

IQVIA said on Feb. 5 that it runs more than 150 agents across more than 30 use cases, choosing a model per task from OpenAI, Anthropic or its own, built with NVIDIA and with AWS as its preferred provider. Its CEO argues that AI investment pays only at IQVIA's scale of more than 10,000 clients.

General assistants

Frontier labs connect to pharma data from outside

Anthropic's life-sciences connectors and skills are generally available on its paid plans, including Benchling with single sign-on. Medidata data is limited to Medidata's own Study Feasibility customers, so the incumbent still controls access. OpenAI's Frontier agent platform launched Feb. 5 for a limited set of customers, with Thermo Fisher among the named early adopters.

Pharma builder

Lilly builds compute and opens models to biotechs

Lilly's lab with NVIDIA is to link automated wet labs with computational work, with scientists involved throughout, and may extend to clinical development, manufacturing and commercial operations; the announcement is in future tense. TuneLab, which gives biotechs access to selected Lilly models, is to add NVIDIA Clara models in a future offering.

AI-native partners

AI-native firms sell platforms through partnerships

Recursion has received $213 million in upfront and milestone payments to date from Roche and Genentech, mostly before 2026, and $134 million in payments from Sanofi, including a $4 million milestone in February, while running its own clinical programs. Insilico licenses assets regionally and in March signed a global collaboration with Lilly.

Selected Q1 2026 launches and their availability

Dates are 2026 unless stated. Availability is as each company described it at announcement.

Selected Q1 2026 launches and their availability
Product or program DateValue-chain stageAvailability
Anthropic life-sciences connectors and skills Date Jan. 11 Value-chain stage Research and trial planning Availability Generally available on paid plans
NVIDIA Clara RNAPro and ReaSyn v2 Date Jan. 12 Value-chain stage Discovery Availability Availability not stated
Lilly-NVIDIA co-innovation lab Date Jan. 12 Value-chain stage Discovery to manufacturing Availability Announced; work to begin early 2026
Lilly TuneLab with NVIDIA Clara models Date Jan. 12 Value-chain stage Discovery Availability Planned future offering
Thermo Fisher autonomous lab infrastructure Date Jan. 12 Value-chain stage Lab operations Availability Announced in future tense
OpenAI Frontier Date Feb. 5 Value-chain stage Enterprise agents Availability Limited customers
IQVIA agents Date Status at Feb. 5 Value-chain stage Clinical and commercial services Availability Deployed; more than 150 agents
Isomorphic IsoDDE Date Feb. 10 Value-chain stage Discovery Availability Internal use only
Veeva AI Agents for CRM and content Date Released December 2025 Value-chain stage Commercial Availability Generally available; several customers live

Where control sits

Access to trial and commercial data still runs through the incumbents that hold it, as Medidata's gating shows. A general agent's usefulness in pharma depends on the connectors and permissions those system owners grant, which makes data-access terms a procurement question alongside model choice.

05Adoption and economics

AI use shows up in deals and vendor counts, not yet in vendor revenue

The quarter's measures of AI use in pharma are structural: deal mix, vendor deployment counts and what vendors said about demand and revenue. None splits use by discovery biotech, clinical-stage company and large pharma, so the segment picture comes from where money and products went.

What the quarter measured about AI use in pharma

Each row is a separate measure from its publisher; none is a survey of AI use by drug developers.

What the quarter measured about AI use in pharma
Measure SegmentValueReported by
Preclinical collaborations involving AI-enabled platforms Segment Discovery biotech and pharma partners Value Nearly half Reported by William Blair, 57 partnerships tracked
Agents deployed Segment CRO and data services Value More than 150, in 30+ use cases Reported by IQVIA
Client wins citing AI Segment Pharma clients, mostly large Value Six, all unnamed Reported by IQVIA, Q4 2025
Customers live on first AI agents Segment Pharma commercial teams Value Several Reported by Veeva, as of March 4
AI revenue outlook Segment Pharma software Value Not material in fiscal 2027 Reported by Veeva
Named Claude users in life sciences Segment Pharma and biotech Value Novo Nordisk, Sanofi, Genmab, Schrödinger, Flatiron Reported by Anthropic, vendor-selected
Trade bodies seeking clarity on generative AI Segment EU industry Value Six associations, including EFPIA and EuropaBio Reported by EMA meeting notes, Feb. 4

Discovery-stage biotech

AI platforms feature in preclinical partnering

With seed and Series A money thin and nearly half of tracked preclinical collaborations involving AI platforms, an AI platform is now part of how discovery biotechs present themselves to partners. Deals run through milestone-heavy licenses, Lilly's TuneLab offers biotechs access to Lilly models, and diligence will test data provenance and whether computational claims can be checked.

Clinical-stage companies

Capital reopened for later-stage programs

Six or seven IPOs raised about $1.8 billion, already more than the $1.6 billion raised in all of 2025 by J.P. Morgan's count, and J.P. Morgan said the window "remains open primarily to a limited set of later-stage, well-positioned companies." AI tools on offer to these companies address feasibility, site selection, recruitment and protocol drafting; EMA's guidance on AI in clinical development will start with a concept paper.

Established pharma

Large companies build, buy and deploy vendor agents

Lilly is building a lab and compute with NVIDIA and licensed Insilico's candidates; J&J partnered with Isomorphic; Novo Nordisk, Sanofi and Genmab are named Claude users. Veeva's CEO described leadership pressure to adopt AI but said AI-driven demand is not yet a major driver of purchasing.

Service and software partners

Vendors pass gains through margin and add-ons

IQVIA guides its R&D Solutions unit to about 4% growth, says productivity gains will be shared with clients over time, and its CFO attributes the improvement in its selling, general and administrative margin largely to productivity. Veeva sells agents as a token-priced add-on and expects CRM to fall from about 20% to about 10% of its revenue by 2030, mainly because other products grow.

AI-native discovery companies' 2025 results, reported in Q1 2026

Fiscal 2025, as reported by the companies.

AI-native discovery companies' 2025 results, reported in Q1 2026
Company RevenueR&D expenseNet lossCash at Dec. 31, 2025
Recursion (reported Feb. 25)Runway into early 2028; 2025 cash operating expense of $399.2 million was about 10% below guidance. Revenue $74.7M (2024: $58.8M) R&D expense $475.3M Net loss $644.8M Cash at Dec. 31, 2025 $753.9M
Insilico (reported March 29)Cash was $125.9 million a year earlier; Insilico listed in Hong Kong on Dec. 30, 2025. Revenue $56.2M (2024: $85.8M) R&D expense $81.4M (2024: $91.9M) Net loss $352.3M Mainly a $296.7M fair-value loss; non-IFRS adjusted loss $43.8M Cash at Dec. 31, 2025 $393.3M

Where the gains land

AI-native companies capture efficiency only through milestones and royalties on assets pharma takes with small upfronts; contract research organizations (CROs) report it mainly as margin they say they will share with clients over time; software vendors sell it as an add-on that is not yet material. None of these mechanisms closes the patent-cliff gap, which large companies addressed in the quarter with acquisitions.

06Outcomes

Reported results: a 12-patient clinical readout and company-measured discovery speed

The quarter's evidence of AI working in drug development is early-clinical and commercial: one surrogate-endpoint signal in a small study, partner payments and Lilly's license. Speed claims are measured by the companies against baselines they chose; they describe discovery and say nothing about how the resulting candidates fare in trials.

Clinical endpoint · Company-reported

REC-4881 in familial adenomatous polyposis

Recursion presents the result as the first clinical validation of its AI platform. In the Phase 2 portion of its TUPELO study, at 4 mg once daily, 75% of evaluable patients had reduced polyp burden. Grade 3 treatment-related adverse events occurred in 15.8%, with none at Grade 4 or above. The study is single-arm with a surrogate endpoint; Recursion planned to discuss a registration path with FDA in the first half of 2026.

43%Median polyp-burden reduction at 12 weeks

12Evaluable patients

9 of 11Kept a durable reduction 12 weeks off therapy

Pipeline status

Insilico's candidates and clinical assets

Six of Insilico's preclinical candidates were nominated in 2025. Its lead program, rentosertib, a TNIK inhibitor for idiopathic pulmonary fibrosis, had completed a Phase 2a study whose results were published in Nature Medicine, and by March 29 Insilico had submitted an investigational new drug application for an inhaled formulation.

28Preclinical candidates nominated to date

10Assets in clinical development

Operations · Vendor-curated

Faster software work at Schrödinger

Schrödinger reports developing software with Claude Code "up to 10x faster in some cases," in a statement on Anthropic's launch page. The tasks are unspecified, and the result concerns software engineering at a computational chemistry company, not a drug outcome.

Company-reported discovery speed and the baselines behind it

The two companies measure different milestones against different baselines; neither figure covers clinical development.

Company-reported discovery speed and the baselines behind it
Company Milestone measuredCompany figureComparatorComparator source
Insilico Milestone measured Target to preclinical candidate Company figure 12-18 months on average Comparator 4.5 years, traditionally Comparator source Company-stated; source not given
Recursion Milestone measured Advanced candidate Company figure About 17 months; about 330 compounds Comparator 42 months; more than 2,500 compounds Comparator source Industry averages cited by company; source not given

How to read these results

Each result sits on a different rung: a computational benchmark (IsoDDE), a preclinical milestone (time to candidate), a clinical endpoint (REC-4881) or a commercial decision (Lilly's license). None of the quarter's results shows a medicine reaching patients faster.

07Corrections

Four claims about pharma AI and what the quarter supports

Each claim below compresses the quarter's announcements into a broader conclusion. The supported version is narrower and changes what a leader should plan for.

Development time · Overstated

Companies report faster discovery against their own baselines; none of them reported a shorter clinical program

Claim under review: AI is shortening drug development.

The speed figures from Insilico and Recursion cover discovery milestones and are measured against baselines the companies chose. Novartis's Fiona Marshall, as reported by Fortune, said AI might eventually cut about five years from discovery but cannot bypass human safety trials. FDA's January bulletin said AI is anticipated to help reduce time to market, an expectation rather than a measured result.

Regulation · Imprecise

They published 10 joint principles, which are voluntary and are not formal guidance

Claim under review: FDA and EMA issued AI guidance for drug development in January.

McGuireWoods described the Jan. 14 principles as "not formal industry guidance," and Jones Day noted that adoption is voluntary, though early alignment may shape future guidance. EMA says the principles will underpin EU guidance now in development. FDA's detailed framework for AI credibility in regulatory submissions was issued as draft guidance in January 2025.

Manufacturing · Incorrect

Draft Annex 22 keeps generative and large language models out of critical GMP applications

Claim under review: Validated generative AI can be used across GMP work under the EU's new AI annex.

The July 2025 consultation draft covers static, deterministic models in critical uses that affect patient safety, product quality or data integrity, and says dynamic, probabilistic and generative models "should not be used in critical GMP applications." Generative AI is allowed in noncritical uses with a qualified human in the loop. Trade associations asked EMA for more flexibility on Feb. 4; the final text is expected by the end of 2026.

Vendor displacement · Not shown in Q1 results

In the first quarter, Veeva and IQVIA added agents to their platforms and guided to continued growth

Claim under review: AI agents are displacing pharma's software and services vendors.

Investors sold information and analytics stocks in early February after Anthropic released Cowork plug-ins; Thomson Reuters fell nearly 18% on Feb. 3. Veeva guided fiscal 2027 revenue to $3,585 million to $3,600 million, up from $3,195.3 million, and IQVIA guided its R&D Solutions unit to about 4% growth. IQVIA's CEO conceded that some lower-level consulting and analytics work could be displaced.

08Accountability

Failure mechanisms: closed claims, untrusted content and unsupervised output

The quarter's cautionary evidence for pharma consists of mechanisms and reservations more than reported incidents at drug developers. Each points to work that stays with the sponsor: checking claims it cannot inspect, controlling what its agents can read and reviewing what they produce.

Incident · Cross-industry

Prompt injection reached agents with data access

On Jan. 15, researchers showed a hidden instruction in a document making Claude Cowork, then a research preview, upload a user's files to an attacker's account; Anthropic said it would ship a fix that day. On Jan. 20, flaws in Anthropic's Git MCP server were disclosed that allowed remote code execution when chained with its filesystem server. They had been fixed in December, with no known exploitation.

Risk · Life-sciences connectors

The same pattern now reaches lab and trial data

Life-sciences deployments now connect general agents to trial-performance data in Medidata, electronic lab notebooks in Benchling and preprint servers whose text no sponsor controls. Draft Annex 11's emphasis on identity and access management and on audit trails is the GMP-side answer, but it covers GMP systems only.

Risk · Verifiability

Closed computational claims

IsoDDE's benchmark results come from a vendor report on an internal model that outsiders cannot inspect, and a Takeda scientist noted that Isomorphic may have had access to partners' private structural data, with an unknown effect on the gains. Fortune reported that Isomorphic had not moved a drug into clinical trials as of January. A partner cannot reproduce such a claim before relying on it.

Limit · Vendor's own view

Outputs are drafts until reviewed

Anthropic's head of life sciences, Eric Kauderer-Abrams, said unsupervised operation is not where the technology is today and advised treating output as a draft, R&D World reported. The same report noted that Anthropic's medical benchmark scores rest on tasks with verifiable answers.

Enterprise evidence

Most firms report no measurable return yet

PwC's survey of 4,454 CEOs, published Jan. 19, found 56% reporting no financial benefit from AI to date. An NBER working paper revised in March found 89% of about 6,000 firms reporting no productivity impact over the past three years. Both are executive self-reports across industries, not studies of pharma deployments.

Who is accountable

EMA told industry on Feb. 4 that sponsors remain responsible for compliance, and the FDA-EMA principles call for assessing the complete system, including how people and the AI interact. A sponsor cannot hand that responsibility to a model vendor.

09Regulation

Joint FDA-EMA principles, the EU's AI guidance sequence and draft GMP rules

The quarter produced one jointly published text and an EU timetable. The principles are voluntary, and the text with legal force in prospect, Annex 22, is still a draft, so 2026 planning rests on drafts and stated intentions. The U.K. regulator aligned itself with the direction without signing.

U.S. and EU · Published Jan. 14; nonbinding

10 FDA-EMA principles for AI evidence

The principles cover AI that generates or analyzes evidence across nonclinical, clinical, post-marketing and manufacturing phases. They ask for human-centric design, a risk-based approach, adherence to standards including GxP, a clear context of use, data governance and documentation, assessment of the complete system including human-AI interaction, and life-cycle monitoring for drift. EMA says they will underpin future guidance.

EU · Set out Feb. 4

EMA's sequence of AI guidance

A concept paper will precede draft guidance on AI in clinical development, and the Pharmacovigilance Risk Assessment Committee (PRAC) is preparing a Q&A document. A proposed Biotech Act would make developing AI guidance a legal duty for EMA, from preclinical research to post-authorization monitoring. The network said it holds no position yet on industry's requests and that detailed rules may not be feasible.

EU GMP · Consultation draft, July 2025; final expected by end of 2026

Draft Annex 22 validation requirements

The draft drew about 1,300 consultation comments and is under revision. For models allowed in critical GMP uses, it requires an intended use approved by subject-matter experts that describes the input sample space, acceptance criteria at least as high as the process replaced, and test data kept independent of training with restricted, audited access. It also requires feature attribution recorded in testing, per-prediction confidence logging with thresholds, change control and drift monitoring.

EU GMP · Consultation draft, July 2025

Draft Annex 11 shifts effort to identity and audit

The revised computerized-systems annex grows from five to 19 pages in 17 chapters, with chapters on identity and access management, audit trails, security, electronic signatures, periodic review, backup and archiving. ECA Academy reads it as shifting emphasis from validation toward security, access management and audit trails, where the quarter's agent incidents exposed weaknesses.

U.S. · Draft guidance, January 2025

FDA's draft credibility framework

FDA's draft asks sponsors to assess AI credibility in seven risk-based steps, from defining the question and context of use to judging adequacy, with model risk set by the model's influence on a decision and the decision's consequence. It excludes drug discovery and operational uses, such as drafting a submission, that do not affect patient safety, drug quality or study reliability.

U.K. · Not a signatory

MHRA aligned with the direction

The Medicines and Healthcare products Regulatory Agency (MHRA) was not a party to the FDA-EMA principles; a spokesperson told The Pharmaceutical Journal on March 18 that its own principles are closely aligned with the direction EMA and FDA set. The National Commission into the Regulation of AI in Healthcare closed its call for evidence on Feb. 2 and plans recommendations to MHRA in 2026.

Dates that shape AI plans in medicines, as known at April 30, 2026

Dates published by April 30, 2026. Dates after March 31 are scheduled or expected, not completed.

Dates that shape AI plans in medicines, as known at April 30, 2026
Item DateJurisdictionStatus
FDA-EMA guiding principles Date Jan. 14, 2026 Jurisdiction U.S. and EU Status Published; voluntary
MHRA National Commission call for evidence Date Closed Feb. 2, 2026 Jurisdiction U.K. Status Recommendations planned in 2026
Commission guidelines on AI Act high-risk classification Date Due Feb. 2, 2026 Jurisdiction EU Status Deadline missed
HMA/EMA AI focus group with industry Date Feb. 4, 2026 Jurisdiction EU Status Held; notes published Feb. 11
Council and Parliament Digital Omnibus positions Date March 13 and 26, 2026 Jurisdiction EU Status Adopted; trilogue not concluded
Next HMA/EMA AI focus groups Date June 2 and Oct. 5, 2026 Jurisdiction EU Status Scheduled
AI Act high-risk obligationsCouncil and Parliament positions would set Dec. 2, 2027, for Annex III uses and Aug. 2, 2028, for Annex I products. Date Aug. 2, 2026, under current law Jurisdiction EU Status In law; Omnibus positions would move it
Final EU GMP Annex 22 Date By end of 2026 Jurisdiction EU Status Expected, as of February
EMA clinical-development concept paper and PRAC Q&A Date Not set Jurisdiction EU Status In development
End of support for Veeva's Salesforce-based CRM Date December 2029 Jurisdiction Commercial systems Status Vendor schedule, moved up from about September 2030

The Omnibus dates are proposals until Council and Parliament agree a final text.

Planning basis

Until the final Annex 22 text, expected by the end of 2026, the draft's line on generative AI in critical GMP uses is the safest planning assumption, and the voluntary principles are the clearest statement of what both agencies will look for.

10Voices

Where regulators, incumbents, AI developers and researchers agree and differ

The positions converge on three points: people stay accountable for AI output, AI's demonstrated leverage is early in the pipeline, and regulators are building risk-based frameworks tied to a context of use. They diverge on whether AI shortens development, which is mainly a question of stage: AI-native companies and FDA's bulletin point to faster discovery and time to market, while Novartis's research president and Veeva's CEO say trials remain gated by biology. They also differ on whether incumbents will be displaced, on generative AI in critical GMP uses and on how far to trust closed computational results.

Regulators and public institutions

Olivér Várhelyi

European Commissioner for Health and Animal Welfare

“a first step of a renewed EU-US cooperation”

He presented the joint principles as the start of renewed EU-U.S. cooperation on novel medical technologies, framed around competitiveness and patient safety.

Regulator · EMA announcement, Jan. 14

HMA/EMA AI focus group

Co-chaired by Peter Arlett (EMA) and Joerg Zinserling (BfArM)

The network is listening but holds no regulatory position yet on industry's requests; guidance can indicate approaches, but detailed rules may not be feasible because context matters, and sponsors remain responsible for compliance.

The answer leaves validation templates and generative-AI boundaries to sponsors for now.

Regulator · published meeting notes, Feb. 4

U.S. Food and Drug Administration

Office of Global Policy and Strategy

AI is anticipated to help promote innovation, reduce time to market, strengthen pharmacovigilance and reduce reliance on animal testing by better predicting toxicity and efficacy in humans.

Stated as an expected benefit, without supporting figures; neither agency's release quoted its head.

Regulator · agency bulletin, Jan. 14

Pharma and incumbent platform leaders

Dave Ricks

Chair and CEO, Eli Lilly

“could reinvent drug discovery as we know it”

He was describing Lilly's data combined with NVIDIA's compute. At the J.P. Morgan Healthcare Conference he described the aim as turning artisanal drug-making into an engineering problem.

Buyer and incumbent · company announcement, Jan. 12

Fiona Marshall

President of biomedical research, Novartis

AI might eventually cut about five years from average discovery timelines, but algorithms cannot bypass human safety trials.

A supporter of AI whose reservation concerns where time can be saved.

Pharma research leader · reported by Fortune

Peter Gassner

CEO, Veeva Systems

AI is not replacing software. It may speed trial start-up and close-out somewhat, but not much, because the protocol and human biology set the timeline and patient recruitment is not an AI problem; customer demand for AI is not yet a major driver.

His company's release the same day called agentic transformation a substantial opportunity; his dissent concerns pace and where gains appear.

Incumbent vendor · earnings call, March 4

Ari Bousbib

Chairman and CEO, IQVIA

AI is more likely to augment clients' teams than replace IQVIA, and demand is unchanged because pharma's AI focus is upstream, in discovery.

Disputes the thesis that AI displaces contract research organizations.

Incumbent CRO · earnings call, Feb. 5

Buyers and pharma scientists

Andrew Adams

Group vice president, molecule discovery, Eli Lilly

“accelerate the identification of promising therapeutic candidates”

The buyer's stated purpose for a license that pays most of its value on milestones.

Buyer · deal announcement, March 29

Diego del Alamo

Computational structural biologist, Takeda

Isomorphic may have had access to partners' private structural data, with an unknown effect on IsoDDE's reported gains.

A diligence question a partner or licensee would put to any closed model.

Pharma scientist · quoted by Nature

EU industry associations

AESGP, AnimalHealthEurope, EFPIA, EUCOPE, EuropaBio and Europharm SMC

They asked regulators for templates, checklists and examples on validation, proportionality for smaller companies, clarity on generative AI in submissions and more flexibility for generative AI and large language models in critical GMP uses.

Draft Annex 22 takes the opposite position on critical uses.

Industry · EMA meeting notes, Feb. 4

AI developers

Najat Khan

CEO, Recursion

“Recursion has reached an inflection point”

She described a move from proving AI can take part in discovery to clinical proof; the company's new clinical data come from a 12-patient surrogate readout.

AI-native developer · results release, Feb. 25

Dario Amodei

CEO, Anthropic

Much of today's AI capability goes unused, held back by security, legal and software-integration work.

Places the constraint in deployment work: security review, legal clearance and integration.

Vendor · reported by R&D World, Jan. 13

Independent and academic researchers

Mohammed AlQuraishi

Columbia University

IsoDDE is a major advance, on the scale of an unreleased AlphaFold 4, but because the model is closed its details are unknown.

Academic researcher · quoted by Nature

Gabriele Corso

Leads Boltz

Proprietary data was probably not essential to IsoDDE's gains, much can still be improved with public data, and the report sets a new baseline to match.

Boltz has launched Boltz Lab, a molecular-design software platform, so his organization offers a competing tool.

Researcher at a competing developer · quoted by Nature

11Decisions

Three decisions for R&D, quality and technology leaders

Head of R&D or business development

Set the evidence bar for AI-derived assets and platform claims

Pharma bought AI discovery on preclinical terms, closed models such as IsoDDE cannot be inspected, and company speed claims rest on their own baselines. Deciding what evidence a platform or candidate must carry protects both deal terms and portfolio choices.

What must an AI-derived candidate or computational result show, and who must be able to check it, before it changes a partnering or portfolio decision?

Head of quality

Classify AI uses against draft Annex 22 now

The draft keeps generative and large language models out of critical GMP uses and sets detailed validation duties for the static models it allows. Industry has asked for flexibility, and the final text is expected by the end of 2026.

Which current and planned AI uses would count as critical GMP applications under the draft, and what would each need to qualify or be redesigned?

CIO or head of R&D technology

Choose where agents run and who controls their access

Veeva and IQVIA embed agents in systems of record and price them there; general assistants reach the same data through connectors whose terms the data owner sets. January's prompt-injection disclosures show the risk when untrusted text reaches an agent with data access.

For each workflow, will agents run inside the system of record or connect from outside, and who owns identity, the audit trail and exposure to untrusted content on each path?

12Actions

Four actions for the second quarter of 2026

  1. 1

    Measure planning tasks, not promised trial speed

    Set baselines for feasibility, site selection and protocol drafting before adopting AI tools, and report time saved per task. The result is a defensible account of where AI saves time, without promising shorter trials.

    Clinical operations lead

  2. 2

    Map AI uses to their context of use

    Record each AI use that generates submission evidence against the FDA-EMA principles and FDA's draft credibility steps, separating discovery and drafting uses that fall outside FDA's draft scope. The register will be ready for EMA's clinical concept paper, the PRAC Q&A and the focus group scheduled for June 2.

    Regulatory affairs lead

  3. 3

    Run a gap check against draft Annexes 22 and 11

    Compare GMP-relevant AI and computerized systems with the drafts' requirements for test-data independence, confidence logging, identity and access management, and audit trails. A gap list now leaves time to fix systems before the final texts.

    Quality head

  4. 4

    Test connectors for prompt injection

    Treat preprints, partner documents and lab-notebook text that reach connected agents as untrusted input; restrict connector scopes and test for injection before agents get write or export access. The result is a documented risk assessment that draws on January's public exploits.

    Information security lead

13References

References

The briefing covers Jan. 1 to March 31, 2026, as known at April 30, 2026. Sources published in April report on the first quarter; documents from 2025 are background.

Regulation and policy

Capability and cost

Research and outcomes

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?