AI moved into pharma’s R&D and trial systems while the clinical evidence stayed early

Frontier AI labs and established software vendors sold agents into discovery, clinical operations and regulated content in the third quarter of 2026. They did so as US tariffs and most-favored-nation pricing deals pressed on revenue per drug, with more than $500 billion of worldwide sales at risk from loss of exclusivity through 2032.

Editorial snapshot: 31 October 2026

The central finding

The quarter’s demonstrated AI gains sat in early discovery and trial operations, while pharma’s economics increasingly depend on late-stage success; only 10% of the 50 US R&D technology executives Citi surveyed in June expect meaningful AI impact on Phase III trials.

01Key findings

What changed in pharma and biotech AI in Q3 2026

The quarter’s clinical milestone for an AI-derived drug was Insilico Medicine’s start of a Phase III trial of rentosertib at 47 sites in China, built on a 71-patient Phase IIa whose efficacy result was exploratory. Four other developments explain the quarter for pharma and biotech leaders.

Capability

AI moved into the systems that run R&D and trials

Anthropic signed collaborations with the contract research organization ICON on July 28 and with Novo Nordisk on Sept. 16, while Veeva, Medidata and IQVIA built AI into clinical, regulatory and commercial software. Claude Science was in beta, Veeva’s Falcon had five early adopters and ICON described its capabilities as planned.

Capital

Biopharma venture fell to $6.7 billion

J.P. Morgan counted $6.7 billion of biopharma venture funding in Q3, down 33% from Q2, with first-time financings the lowest since Q3 2025. Chai Discovery’s $400 million Series C, larger than any round the J.P. Morgan report names, does not appear in it.

Adoption and economics

Buyers expect early gains as US price pressure grows

Of 50 US R&D technology executives Citi surveyed in June, 72% are scaling or fully scaling AI and 10% expect meaningful Phase III impact. By Aug. 31, 26 manufacturers covering 89% of the US branded market, by the White House’s count, had signed most-favored-nation (MFN) pricing deals.

Accountability

AI rules stayed in draft; GMP duties already apply

FDA’s guidance on AI in drug submissions remained a 2025 draft, and the EU’s draft GMP annex on AI, Annex 22, was not adopted. Under existing rules, an April FDA warning letter held a manufacturer’s Quality Unit responsible for AI-generated production records.

Operating question

Where in the pipeline would AI have to reduce Phase II and Phase III failure, not only speed discovery, to justify its cost at lower US prices?

02Capability

AI capability moved into scientific and regulated workflows

The quarter’s technical change was in where AI sits. General-purpose labs packaged scientific workbenches, and R&D software vendors built AI that acts inside systems of record. No capability claim in the quarter rose above a company-reported wet-lab result or operational association.

Discovery research · Beta

A scientific workbench from a general-purpose lab

Anthropic’s Claude Science, in beta from July 1, links PubMed, Jupyter notebooks and R with more than 60 functions across genomics, structural biology, proteomics and cheminformatics, from CRISPR screen design to single-cell analysis. Anthropic published no performance metrics. Novo Nordisk said on Sept. 16 that it would test it in selected R&D workflows.

Clinical and regulatory operations

AI that works inside systems of record

Medidata Plus, launched July 24, automates electronic data capture (EDC) and electronic clinical outcome assessment (eCOA) study builds with synthetic test data, ingests trial, real-world and wearable data without code, and answers natural-language questions about audit logs. Veeva’s Falcon is an agentic platform for clinical, regulatory and safety work.

Discovery to trial strategy

Agent chains from target to enrollment

Recursion describes a Target Discovery Agent, Drug Design Agents and a Clinical Strategy Orchestration Agent working across biology, molecule design and clinical strategy, and it appointed Hoifung Poon chief AI officer. The company links the clinical agents to higher enrollment rates, covered under outcomes.

General-purpose models

Faster model turnover at lower token prices

OpenAI’s GPT-5.6 launched July 9; Google’s Gemini 3.7 Flash arrived Aug. 13 at an introductory price half that of 3.6 Flash; Anthropic’s Claude Opus 5.5 followed in September at $4 and $20 per million input and output tokens, 20% below Opus 5. Ramp’s business spending data showed the effective price per million tokens down 41% from a March peak, to $0.68, with token spend flat in August and September.

Regulatory review · Background, May 2026

FDA reviewers work with AI tools

FDA’s Elsa 4.0, with custom agents, document generation and data analysis, runs on HALO, a platform consolidating more than 40 submission data sources across centers. The HHS secretary said more than 90% of reviewers were using AI tools. Officials said Elsa does not train on submitted data and AI does not make regulatory decisions.

Cost and consequence

Token prices fell, but agents that write into EDC, content, CRM and safety systems make the integration the thing to validate, and a model release every few weeks makes each update a change-control decision. Pharma Now’s guidance on draft Annex 22 names drift monitoring and revalidation triggers as the costliest preparation.

03Capital

Biopharma venture contracted as AI model companies raised separately

Two capital markets ran side by side. Conventional biopharma venture shrank and moved to later rounds, while AI model companies such as Chai Discovery raised from generalist and AI investors, apparently outside biopharma tallies. Large pharma spent on acquisitions and Chinese molecules and bought AI discovery largely through contingent payments.

Disclosed venture rounds, Q3 2026

Q3 rounds named by J.P. Morgan, with Chai Discovery alongside

USD millions. Biopharma rounds as named in J.P. Morgan’s Q3 report, which cites large, later-stage and European rounds separately rather than ranking them; Chai Discovery’s round, from its own announcement, is not mentioned in the report.

Chai Discovery

Series C — $400M

AI model company; $3.8 billion valuation; no clinical assets

Encoded

Series F — $275M

Solstice Oncology

Series A — $225M

LifeMine

Series E — $188M

Vaderis

Series B — $152M

Claris

Series B — $118M

Kynexis

Series A — $110M

R1

Series A — $81M

J.P. Morgan Q3 2026 Biopharma Licensing and Venture Report (DealForma data); Chai Discovery announcement, July 14.

$6.7B

Biopharma venture funding in Q3, globally

J.P. Morgan count; down 33% from $9.9 billion in Q2.

$159B

All-sector global venture funding in Q3

Crunchbase count; 64% went to AI companies, a share that does not isolate drug developers.

Venture structure

Later rounds took most of the money

Sixty-eight rounds of $100 million or more were raised in the first nine months. J.P. Morgan still projects about $33 billion of biopharma venture for 2026, above 2025’s $30.5 billion, despite the third-quarter decline.

72%Of year-to-date capital to Series B and later rounds, in 56% of rounds

$1.7BFirst-time financings in Q3, lowest since Q3 2025

AI model companies

Chai’s round is absent from J.P. Morgan’s report

Chai Discovery’s $400 million Series C, at a $3.8 billion valuation, was led by Index Ventures with Kleiner Perkins, Sequoia and Dimension; OpenAI and Thrive are existing investors. The company is preclinical, with Eli Lilly and Pfizer as named partners. Model companies without pipelines appear to fall outside biopharma counts, so AI’s share of biotech funding cannot be read from them.

Deal structure

AI discovery is bought as options

Upfront cash was 6% of announced biopharma licensing value year to date, and 35 deals carried $100 million or more upfront, against 46 in all of 2025. AbbVie’s September discovery deal with Iambic Therapeutics includes an upfront payment, milestones and tiered royalties, all undisclosed; Anthropic’s deals with ICON and Novo Nordisk disclosed no terms. In May, Incyte paid Genesis Molecular AI $120 million upfront against up to $232 million in milestones per program.

Licensing

China supplies a growing share of licensed molecules

China-origin upfronts this year total roughly $5.7 billion to $5.8 billion across 18 deals, against 19 in all of 2025. The largest Q3 collaboration was the CSPC and AstraZeneca joint venture: $1.2 billion upfront on $18.5 billion in total value.

51%Of big-pharma licensing deals with at least $50M upfront were China-origin, year to date (39% in 2025)

65%Of those deals’ upfront dollars (36% in 2025)

M&A

M&A remains the main answer to the patent cliff

Biopharma M&A upfront value was $31.1 billion across 32 deals in Q3, down from $66.8 billion in Q2, but the year-to-date $138.9 billion already exceeds 2025’s $122.1 billion. Announced Q3 deals included Vertex–Crinetics (about $10 billion), Lilly–Merida Biosciences ($2.8 billion) and Novartis–Myricx Bio (about $1.5 billion). Since the start of 2024, Merck has announced six deals worth $30.9 billion ahead of Keytruda’s 2028 patent expiry.

UK and Europe

Two counts of European biotech funding

J.P. Morgan counts $1.1 billion across 31 European rounds in Q3, against $584 million a year earlier, with $380 million in the UK. The BioIndustry Association counts £270 million for UK biotech, up 44% from £187 million, led by Basecamp Research’s £104 million Series C, and notes 17 quarters without a UK biotech IPO. The two differ in currency, company definitions and rounds included.

What it signals

For AI-native biotechs, revenue depends on partners advancing programs, because pharma pays mainly through milestones. For large pharma, AI discovery partnerships are a small, contingent line beside M&A and Chinese licensing, which carry the near-term response to the patent cliff.

04Products

Three routes for supplying AI to pharma and CROs

Frontier labs contracted directly with sponsors and contract research organizations (CROs). Established R&D and commercial platforms added AI to software their customers already run. AI drug developers sold their own stacks and design services. The route determines who validates the AI and who controls its versions.

Direct lab contracts

Frontier labs sign sponsors and CROs

ICON’s multi-year collaboration puts Anthropic’s Claude across the trial lifecycle on its Orbis platform, for site planning, enrollment-risk prediction and protocol optimization, all described as future capabilities. Novo Nordisk’s collaboration covers selected R&D workflows and software development. In May, Bristol Myers Squibb agreed to deploy Claude across clinical development, manufacturing, commercial and corporate functions, with a rollout to more than 30,000 employees. Pharmaceutical Technology lists Novo Nordisk, Eli Lilly, Moderna and Sanofi as OpenAI partners, without dates or terms.

Installed platforms

Incumbents sell agents into their installed base

Veeva counts more than 180 Vault CRM customers live, five of them top-20 biopharma companies, and 12 of the top 20 committed; its Agentic Call Report runs across one top-20 company’s full US field force. IQVIA reports more than 400 AI agents deployed alongside its CRO and data services. Medidata places its AI layer inside the EDC system where trial data already sit.

AI drug developers

Developers sell their own stacks and designs

Recursion and Insilico sell their own agent stacks; Genentech optioned a first neuroscience target from Recursion into an early discovery program, and Sanofi named Recursion and Owkin as AI partners. Chai sells antibody design as a service. Nimbus Therapeutics’ chief executive, Abbas Kazimi, separates tool builders, drug-discovery companies and development companies, placing Nimbus in the middle as an integrator.

Q3 launches and deals by status

Status as announced by each company or reported in launch coverage. Veeva items are from its Aug. 26 results for the quarter ended July 31; IQVIA’s count is from its Oct. 27 third-quarter results.

Q3 launches and deals by status
Product or deal DateStatusWhere it works
Claude Science (Anthropic) Date July 1 Status Beta Where it works Literature, notebooks and analysis for scientists
Chai-3 models (Chai Discovery) Date July 14 Status Platform offered to partners Where it works Antibody design for pharma partners
Medidata Plus (Dassault Systèmes) Date July 24 Status Launched; availability not stated Where it works Study build and data management in EDC
ICON and Anthropic Date July 28 Status Announced; capabilities planned Where it works Trial planning, enrollment risk, protocols
Veeva Vault AI agents Date August Status Released Where it works Content, CRM and clinical apps; custom agents
Veeva Falcon Date August Status Five early adopters; go-lives planned for 2026 Where it works Clinical, regulatory and safety work
Veeva Agentic Call Report Date August Status Deployed at one top-20 biopharma Where it works US field-force call reporting
Novo Nordisk and Anthropic Date Sept. 16 Status Announced Where it works Selected R&D workflows; software development
IQVIA AI agents Date Through Q3 Status More than 400 deployed Where it works Sold with CRO and data services

Who validates

Embedded in a platform such as Veeva or Medidata, the AI changes with the vendor’s release cycle; delivered through a CRO, it arrives inside a service contract; licensed directly from a lab, as at Novo Nordisk and Bristol Myers Squibb, validation, change control, data-protection terms and model-version management fall to the sponsor’s IT and quality teams.

05Adoption

AI adoption by R&D stage, and where the economics land

Adoption is broad among large US R&D organizations, concentrated in structural biology and early discovery, and expectations fall away by phase. US price policy reduces revenue per approved drug, so AI’s economic case rests on more assets per R&D dollar or fewer late-stage failures. No survey in the quarter split adoption by company segment.

What 50 US R&D technology executives reported to Citi

Citi Research survey of US executives responsible for R&D technology strategy, fielded in June 2026 and reported Sept. 9; company segments not specified. None reported having no adoption plans, and a majority expect double-digit AI budget increases.

What 50 US R&D technology executives reported to Citi
Measure Share of respondents
Use protein-structure prediction or structural biology Share of respondents 78%
Scaling or fully scaling AI Share of respondents 72%
Have engaged AI vendors Share of respondents 84%
Spent more than $50 million on AI drug discovery in the past year Share of respondents 26%
Expect the greatest success-rate gains in target identification Share of respondents 70%
Expect meaningful impact on Phase III Share of respondents 10%

Digital leaders reporting measurable AI outcomes, by function

ZS survey of life-sciences chief digital and information officers, described in a PharmTech interview published Oct. 23; field dates and sample size not given. ZS sells consulting and technology services.

Digital leaders reporting measurable AI outcomes, by function
Function 20252026
Supply chain and manufacturing 2025 29%
2026 53%
Drug discovery 2025 17%
2026 35%
Clinical and medicalAbout flat; no figures given 2025 — No comparable figure 2026 — No comparable figure

38% reported measurable outcomes in multiple functions, and 17% said AI governance was fully operational across the enterprise.

>$500B

Worldwide prescription sales at risk from loss of exclusivity, 2026–2032

Evaluate estimate published in October, peaking at $158 billion in 2031; sales at risk are not sales lost.

89%

Share of the US branded market under MFN pricing deals, by the White House’s count

26 manufacturers after nine more signed on Aug. 31.

All branded manufacturers

US price policy took effect

Section 232 tariffs of up to 100% on patented drugs applied from July 31 to 17 large manufacturers and from Sept. 29 to other branded makers without both MFN and onshoring agreements; companies with both pay 0% through Jan. 20, 2029. The nine that signed on Aug. 31 committed to MFN prices in Medicaid and on new launches and to at least $19.6 billion of US manufacturing investment. Gibson Dunn notes that MFN terms referencing ex-US net prices may affect royalties and transfer pricing in licensing deals.

Established pharma

Large companies frame AI as capacity

Lilly’s US price fell 7% in Q3, or 10% excluding a rebate and discount adjustment, and it expects net price to be a headwind in 2027, which Lilly USA’s president attributed to a payer mix shifting toward Medicare and Medicaid. Sanofi’s R&D head said R&D spend should not change much while the company targets upper-quartile speed and cost and 12 to 15 first-in-human studies a year. He gave no quantified AI outcome.

Clinical-stage companies

AI-native developers face financing limits

Recursion reported Q2 revenue of $7.7 million and a net loss of $131.0 million, with cash down from $753.9 million at the end of 2025. It cut 2026 cash operating expense guidance to under $375 million from under $390 million and reduced headcount.

$556.8MRecursion cash on June 30

Early 2028Stated cash runway

Discovery-stage biotech

Chai raised $400 million as first-time biotech financings fell to a one-year low

Discovery-stage companies sell into the stage where buyers report the most use, led by structural biology at 78% of Citi’s respondents. In Q3, Chai Discovery raised $400 million at a $3.8 billion valuation while first-time biopharma financings totaled $1.7 billion, the lowest since Q3 2025. As general-purpose workbenches reach pharma directly, Nimbus’s Abbas Kazimi argues that the lasting advantage is turning tools into validated learning.

CROs and software platforms

Service providers report growth

IQVIA’s R&D Solutions revenue rose 8.7% in Q3 and net new bookings 15%, a book-to-bill of 1.2, with backlog of $32.6 billion after a roughly $2.1 billion adjustment for inactive trials. Veeva’s revenue for the quarter ended July 31 rose 18% to $928.0 million, with GAAP operating income up 40%. Both raised full-year guidance; neither result measures what sponsors save.

Manufacturers and CDMOs

Manufacturing faces the first AI-specific GMP text

Manufacturing and supply chain show the steepest rise in ZS’s outcome measure. Manufacturers and contract development and manufacturing organizations (CDMOs) also face draft Annex 22 on AI, a draft revised Annex 11 that makes cybersecurity a GMP requirement, and the April US warning letter. US supply-chain rules add pressure: Section 805 of the FY2024 defense authorization act reaches contractors’ sub-tier suppliers on June 30, 2027, and biologics tech transfers take 12 to 24 months if BIOSECURE’s list forces a supplier change.

Economics

Measured gains sat in early discovery and manufacturing, and growth showed in CRO and platform results, while US policy reduced revenue at launch. For sponsors, AI’s economic case rests on more approvable assets per R&D dollar or lower late-stage attrition, yet only 10% of Citi’s respondents expect meaningful impact on Phase III.

06Outcomes

What the quarter’s AI results show, by evidence level

The quarter’s named results sit on different rungs: computational design with vendor-reported lab testing, company-reported operational associations, internal tooling and one Phase III start built on early clinical data. None reached approved use, and none measured a shorter overall development timeline.

Clinical endpoint · Company-reported

Rentosertib’s Phase III rests on an exploratory Phase IIa

On July 7, Insilico Medicine started a randomized, double-blind, placebo-controlled Phase III of its TNIK inhibitor for idiopathic pulmonary fibrosis: 320 planned patients at 47 sites in China, with annual decline in forced vital capacity (FVC) as the primary endpoint over 52 weeks. The 71-patient, 12-week Phase IIa met its safety objective; its FVC result was exploratory, and the release gives no p-value.

+98.4 mLMean FVC change at 60 mg, Phase IIa

−20.3 mLMean FVC change on placebo, Phase IIa

Clinical transition rates · Background, May 2026

AI-enabled programs lead at Phase I, not Phase II

The IQVIA Institute studied programs with vetted AI involvement at emerging biopharma over the latest three-year window. Phase III numbers were too few to assess, the cohort is small, industry-wide success rates were unchanged and end-to-end timelines were at a decade high. IQVIA sells R&D services.

75%Phase I success, AI-enabled programs

On parPhase II success versus non-AI peers

Trial operations · Company-reported association

Agent-supported enrollment at Recursion

Recursion associated enrollment strategies supported by its clinical-strategy agents with higher enrollment rates than historical benchmarks. It did not describe the benchmarks, the trials or how much of the difference it attributes to the agents.

1.3–1.6xEnrollment rate versus historical benchmarks

Computational design, lab-tested · Vendor-reported

Antibody design hit rates from Chai Discovery

Chai describes its 2025 Chai-2 model as reaching double-digit experimental success rates in zero-shot de novo antibody design, and Chai-3 as improving target success rates and binding affinity, without figures. Targets, methods and denominators were not published, and the company has no clinical assets.

Internal tooling · Self-reported

Faster internal tooling at Nimbus Therapeutics

Chief executive Abbas Kazimi also reports using coding agents to rewrite a pharmacophore and shape-matching codebase in about a week, work he says might previously have taken months. It is one company’s account, without comparator data.

9 daysTo build and deploy a custom pKa model

Where the results hold

The strongest independent signal, from the IQVIA Institute, shows an advantage at Phase I that is gone by Phase II, the stage a Nature Reviews Drug Discovery paper cited by Axios identifies as the bottleneck for AI-derived drugs. That matches what Citi’s respondents expect and does not yet support claims of shorter development overall.

07Corrections

Claims about AI in drug development that need correcting

Four claims that recur in announcements and coverage would change how a leader sizes AI’s effect on development. Each has a narrower supported version.

Clinical success rates · Overstated

Recent data show a Phase I edge for AI-enabled programs and parity by Phase II

Claim under review: AI-discovered drugs succeed in Phase I 80% to 90% of the time.

The figure is widely attributed to a 2024 Boston Consulting Group analysis. The IQVIA Institute’s May 2026 analysis of AI-enabled programs at emerging biopharma found 75% Phase I success over the latest three-year window, above comparable programs, and Phase II success roughly on par with non-AI peers. The cohort is small, and IQVIA sells R&D services.

UK trial approvals · Misattributed

Risk-based reforms cut UK approval times; AI tools came afterward

Claim under review: AI halved UK clinical trial approval times.

The October 2025 release from the UK Medicines and Healthcare products Regulatory Agency (MHRA) carries a headline crediting AI and reforms, but its body attributes the fall from 91 to 41 days to risk-based reforms, with AI tools then being introduced.

Development timelines · Premature

Rentosertib’s Phase III has only started, and no announcement measured a shorter timeline

Claim under review: An AI-derived drug in Phase III shows that AI shortens drug development.

Insilico’s Phase III, initiated July 7, runs at 47 sites in China and builds on a 71-patient Phase IIa whose efficacy result was exploratory. None of the quarter’s announcements measured an overall development timeline, and in May the IQVIA Institute found end-to-end timelines at a decade high.

Launch statistics · Unsourced

The ICON–Anthropic release gives no source for the 80% figure

Claim under review: Enrollment delays affect up to 80% of clinical trials.

ICON’s July 28 release attributes the figure to Anthropic without citing a study. Coverage of the Medidata Plus launch likewise cited 92% of organizations planning to raise AI spend and 29% saying AI meets expectations without naming a survey, and claimed study builds in hours instead of weeks without a stated basis. Neither set of figures can support a business case until the underlying data are published.

08Accountability

Where AI deployments failed or exposed risk

The documented failures concern human review, translation to efficacy, the security of agent integrations and insurance cover more than model error. In each, responsibility stayed with the organization using the AI.

Enforcement · Background, April 2026

FDA holds the Quality Unit responsible for AI output

An April FDA warning letter to Purolea Cosmetics Lab cited 21 CFR 211.22(c) and 211.100: AI-generated specifications, procedures and master production records had no Quality Unit review, and process validation was skipped, which the firm said its AI agent had not flagged. An authorized Quality Unit representative must review and approve AI output; the duty cannot be delegated to the tool.

Translation · Sept. 16

AI-derived drugs meet the same Phase II bottleneck

Axios, citing a Nature Reviews Drug Discovery paper, reported that AI’s clinical impact has been disappointingly limited: AI-derived drugs meet the same Phase II efficacy bottleneck, and few have progressed beyond Phase II. Faster discovery does not by itself change that stage.

Security · August

Agent connectors became an attack surface

The Cloud Security Alliance’s Aug. 30 briefing described malicious Model Context Protocol (MCP) servers that stayed dormant for three calls before harvesting credentials, a Langflow vulnerability (CVE-2026-9198) on CISA’s known-exploited list, and an injection technique against a coding agent’s auto mode. These were cross-industry incidents; the draft revised EU Annex 11 makes cybersecurity a formal GMP requirement.

Insurance

Insurers moved to exclude generative AI from general-liability and products cover

Insurance Journal reported in July that carriers were increasingly filing to use three generative-AI exclusions from the Insurance Services Office (ISO): two for commercial general liability and one, CG 35 08, for products and completed-operations cover, the part of a general-liability policy that answers product claims. Verisk, which runs ISO, did not know how many insurers had adopted them. The report did not address clinical-trial insurance.

Who answers for it

The Quality Unit answers for GMP records, the sponsor for efficacy, the deployer for its integrations and, where policies exclude AI, the company’s own balance sheet for losses. Contracts with labs and platforms should be read against those duties.

09Regulation

AI rules for drug development stayed in draft

Regulators described AI more than they ruled on it in the quarter. FDA’s news concerned leadership, staffing and pilot design; the EU’s AI-specific GMP text stayed in draft; and the UK worked through sandboxes. For AI in GxP work, the operative obligations at quarter end were existing CGMP and computerized-system rules plus the April enforcement precedent.

FDA · Drugs

FDA’s AI guidance stayed in draft under acting leadership

The AI page of FDA’s Center for Drug Evaluation and Research, last updated May 1, still lists the January 2025 draft guidance, a risk-based credibility framework for AI that produces data for safety, effectiveness or quality decisions, beside nonbinding principles agreed with the European Medicines Agency (EMA) on Jan. 14, 2026. Commissioner Marty Makary resigned in May, and by Aug. 20 the White House had picked Heidi Overton as nominee. FDA was reported to have lost more than 3,000 staff in a year.

FDA · Devices, discussion paper, August

A discussion paper on generative-AI devices

FDA’s device center issued a discussion paper, not draft guidance, proposing a foundation-model master file, competency-based performance evaluation, a distinction between informational and action-taking functions, and periodic re-benchmarking. Comments were due Oct. 19. It bears on companion software, digital endpoints and patient-facing tools.

EU · GMP

EU GMP Annex 22 on AI remains a draft

After a 2025 consultation that drew about 1,300 comments and a follow-up workshop on June 30 and July 1, EudraLex Volume 4 listed no Annex 22 on Aug. 12, and Annex 11 and Chapter 4 were still the 2011 versions. The draft covers model selection, training, validation, intended use, performance metrics, data quality, change control, monitoring and human review. Final text to the Commission in Q4 2026 is an unconfirmed target.

EU · Medicines network

EU regulators plan guidance for trials and pharmacovigilance

The 2025 AI Observatory report of the EMA and Heads of Medicines Agencies (HMA) Network Data Steering Group, published in June, counted 61 AI use cases at national authorities and found that tools for assessing AI, including validation, explainability and audit, lag adoption. Its 2026–2028 workplan foresees guidance on AI in clinical development and pharmacovigilance.

UK

The UK works through sandboxes and pathway reforms

The MHRA’s medicines AI sandbox, launched June 9 with partner work from the summer, will test up to five AI approaches that predict how medicines behave in the body and their safety, including in children, older people and diverse groups, aiming to strengthen safety evidence and reduce animal testing. The AI Airlock for AI medical devices is funded to 2029, and an advance, nonbinding review of nonclinical (Module 4) data for animal-free products is planned by the end of 2026.

US · Supply chain

BIOSECURE deadlines continue despite the WuXi ruling

On Aug. 7 a federal court in Washington preliminarily enjoined the Defense Department from giving effect to WuXi AppTec’s Section 1260H designation, finding it had misread its evidence. The order does not bind the Office of Management and Budget, whose initial list of biotechnology companies of concern is due Dec. 18 under BIOSECURE, enacted in December 2025.

Compliance and policy dates known at Oct. 31, 2026

Dates as set by statute, proclamation, court order or published work plan by the end of the quarter, unless noted.

Compliance and policy dates known at Oct. 31, 2026
Item DateStatus
Section 232 tariffs for 17 large manufacturers Date July 31, 2026 Status In effect
EU AI Act general-purpose AI enforcement and transparency duties Date Aug. 2, 2026 Status In effect
Section 232 tariffs for other branded manufacturers without MFN and onshoring agreements Date Sept. 29, 2026 Status In effect
Comments on FDA’s generative-AI device discussion paper Date Oct. 19, 2026 Status Deadline set in August
Final Annex 22, Annex 11 and Chapter 4 text to the European Commission Date Q4 2026 Status Work-plan target, unconfirmed
Vacancies Act limit for FDA’s acting commissioner, Kyle Diamantas Date Early December 2026 Status As reported in August
OMB initial BIOSECURE list of biotechnology companies of concern Date Dec. 18, 2026 Status Scheduled
MHRA advance nonclinical review for animal-free products Date By the end of 2026 Status Planned
Sub-tier supply-chain flow-down under Section 805 of the FY2024 National Defense Authorization Act Date June 30, 2027 Status Scheduled
EU AI Act high-risk obligations for Annex III uses Date Dec. 2, 2027 Status Scheduled
EU AI Act high-risk obligations for AI in regulated products, including medical and in vitro diagnostic devices Date Aug. 2, 2028 Status Scheduled
End of 0% tariff rate for companies with MFN and onshoring agreements Date Jan. 20, 2029 Status Scheduled
10Voices

Where pharma leaders, vendors and skeptics agree and differ

The people below largely agree that AI’s demonstrated contribution is early, in structure prediction, design and target work, and that human oversight is a condition of use. They differ on three questions. On whether AI will shorten development overall, Sanofi’s R&D head expects shorter cycles and higher clinical success, while investor Bijan Salehizadeh expects effects to stay at the surface; the first speaks for a large company’s portfolio, the second for late-stage trial economics across the field. On where advantage lies, Lilly committed up to $1 billion over five years to a co-innovation lab with NVIDIA in January and Novo Nordisk contracted with Anthropic, while Nimbus’s Abbas Kazimi argues that infrastructure enables but does not differentiate. On whether regulators will speed early trials, FDA’s chief AI officer, Jeremy Walsh, said in April that real-time trial data could help companies advance to the next phase faster; BIO and PhRMA doubt the design of FDA’s expedited investigational new drug (IND) pilot. Vendor voices, and most incumbent ones, carry commercial interests.

Large pharmaceutical companies

Paulo Fontoura

Head of R&D, Sanofi

“AI will be a key lever across the R&D value chain.”

He aims to deploy integrated AI agents company-wide and expects faster decisions, higher clinical success and shorter cycles, with R&D spend roughly level. He gave no quantified AI outcome.

Incumbent · Q3 earnings call, Oct. 30

Mike Doustdar

President and CEO, Novo Nordisk

AI can raise R&D productivity and compress the path from research to a marketed product.

Novo named data protection, compliance alignment and human oversight as central to the collaboration.

Buyer · Anthropic collaboration, Sept. 16 (as reported)

CROs and biotech buyers

Barry Balfe

CEO, ICON

“We are determined to lead AI acceleration of clinical trials.”

A top CRO is building trial services on a frontier lab’s models instead of its own; the announced capabilities were all described as planned.

CRO buyer · July 28 announcement

Abbas Kazimi

CEO, Nimbus Therapeutics

“the advantage isn’t simply speed. It’s the speed of validated learning”

He argues that as AI tools become broadly available, infrastructure enables the work but does not set a company apart.

Biotech buyer and integrator · LifeSciVC essay, July 28

AI drug developers

Alex Zhavoronkov

Founder and CEO, Insilico Medicine

“this is no longer only a speed story”

Said as rentosertib entered Phase III; the asset’s efficacy evidence so far is an exploratory Phase IIa result.

Vendor · July 7 Phase III announcement

Joshua Meier

CEO, Chai Discovery

“AI drug discovery has moved from promise to deployment”

Chai has no clinical assets; its performance claims rest on its own lab testing.

Vendor · July 14 funding announcement

Investors and skeptics

Bijan Salehizadeh

Investor, NaviMed Capital

“I'm afraid the AI applications in development will be surface level”

He expects AI to leave costly late-stage trial work untouched, the stage where IQVIA Institute data show parity at Phase II.

Investor · Axios, Sept. 16

Trade associations and regulators

BIO and PhRMA

US biotechnology and pharmaceutical trade associations

Questioned whether the qualified research institutions in FDA’s proposed expedited-IND pilot would speed trial-start timelines.

FDA’s request for information drew more than 200 comments.

Trade associations · comments to FDA, August–September

EMA–HMA Network Data Steering Group

EU medicines regulators’ data steering group

Regulators’ tools for assessing AI, including validation, explainability and audit, lag adoption; guidance on AI in clinical development and pharmacovigilance is planned for 2026–2028.

The EU network described use and planned guidance; it issued no binding AI instrument in the quarter.

Regulator · AI Observatory report and workplan

11Decisions

Three decisions for R&D, quality and technology leaders

Chief R&D officer

Decide which development stage AI is expected to change

The quarter’s evidence places AI’s gains in structure prediction, design, target work and trial operations, and IQVIA Institute data show a Phase I advantage that is gone by Phase II. With MFN launch pricing and tariffs lowering revenue per approved drug, AI’s economic case rests on more assets per R&D dollar or lower late-stage attrition.

Which programs will test whether AI-selected or AI-designed candidates fail less often in Phase II than our own historical baseline, and when will we know?

Head of quality

Decide how AI enters change control

Annex 22 is a draft and FDA’s AI guidance is not final, but existing CGMP and computerized-system rules already apply, and the April warning letter put AI-generated records under the Quality Unit. Major model vendors released new versions every few weeks in the quarter.

Which AI uses touch GxP records, who in the Quality Unit approves their output, and which vendor changes trigger revalidation?

Chief information officer

Decide where validation of the AI layer sits

Sponsors can take AI embedded in a vendor’s system of record, such as Veeva or Medidata, through a CRO’s services, such as ICON or IQVIA, or by contracting a lab directly, as Novo Nordisk and Bristol Myers Squibb did. Each route places validation evidence, data-protection terms and model-version control with a different party.

For each AI capability we use, who holds the validation record, model version history and audit trail, and can our auditors inspect them?

12Actions

Four actions for the next quarter

  1. 1

    Build an inventory of AI that touches GxP records

    List each model and agent that drafts, checks or changes a regulated record, with a named owner, risk classification, human-review step and revalidation trigger. The inventory serves inspections under current rules and maps onto draft Annex 22 if it is adopted.

    Quality

  2. 2

    Prepare submissions for AI-assisted review

    Use consistent terminology and cross-references and make conclusions traceable to tables in key modules, as Hogan Lovells Cadwalader advises, because FDA reviewers work with AI tools. Record where AI drafted content and who reviewed it. The result is a dossier whose AI use and human review are documented.

    Regulatory affairs

  3. 3

    Ask vendors for the basis of operational claims

    Before enrollment or study-build claims enter a business case, request the trials, comparators and periods behind figures such as Recursion’s 1.3 to 1.6 times enrollment rates, the hours-instead-of-weeks study-build claim in Medidata Plus coverage and the up-to-80% enrollment-delay statistic. The result is a case built on measured baselines, including your own.

    Clinical operations

  4. 4

    Treat agent connectors as regulated interfaces

    Review MCP servers, workflow tools and coding agents connected to quality, trial management, safety or CRM systems for credential handling and known exploited vulnerabilities, given the August incidents and the draft Annex 11’s cybersecurity requirement. The result is an integration inventory with owners and patch status.

    IT security

13References

References

The briefing covers July 1 to Sept. 30, 2026, using sources published through Oct. 31, 2026. Sources dated before July 1 appear as labeled background.

Regulation, courts and insurance

Voices and earnings calls

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?