Agentic AI Speeds Drug Work While FDA Rules Stay Draft

McKinsey’s life-sciences team says 75 to 85 percent of pharma workflows have tasks that AI agents could take on, freeing 25 to 40 percent of staff time. Agentic AI is the name for systems that plan, call tools, and loop without a prompt at every step. 2026 lab case studies already turn weeks of early drug work into hours.

The leftover problem is not speed. It is whether anyone can prove the agent was right when the output feeds a trial, a filing, or a plant.

The IPF Program That Finished Before Lunch

A 2026 review in Drug Discovery Today, led by Dinh Long Huynh and Srijit Seal with the AIAgents4Science Consortium, walks through agents that already run pieces of discovery. The paper’s line is dry and useful. Agentic systems couple large language models to tools, memory, and data so they can think, act, observe, and reflect in loops.

Kiin Bio’s Virtual Scientists platform stitched literature review, omics analysis, structural modeling, and generative chemistry into one idiopathic pulmonary fibrosis hunt. The authors say that job usually takes two to three weeks. The agent stack finished it in under two hours and pulled in more than 100 tools.

Potato’s Tater agent took a messier chore. A scientist asked it to derive an automated qPCR protocol on an Opentrons robot for adeno-associated virus quantification, including a path toward a clinical release test. Tater searched the literature, compared methods, wrote a MIQE-aligned protocol, and emitted robot code.

DISCOVERY CLOCKS IN THE 2026 CASE STUDIES

The Tater authors call that greater than a 400-fold cut when they use the short end of the manual range. Empirical checks still had to follow. onepot.ai goes further and ties the same style of agent to a tube picker, decapper, liquid handler, plate sealer, and LC/MS deck.

Those clocks measure design, search, and protocol writing. They do not measure a molecule in a person. Max Jaderberg of Isomorphic Labs has said the next step is to treat agents as collaborators that search molecular space and come back in hours with work that used to take weeks. That is the pitch drugmakers are buying. It is also the part that still needs a wet-lab receipt.

A Digital Twin of the FDA Is Already in the Pitch

Clinical development eats nearly 70 percent of total R&D spend, McKinsey said in a December 11, 2025 note, and failure rates stay high. The firm’s operational claim is blunt. Agents could let a company run up to twice as many trials with the same resources and cut trial duration by as much as 12 months.

Study start-up is the first place the math gets concrete. Agents score sites on past performance and demographics, draft first-time-right contracts against fair market value, and handle outreach. McKinsey says that mix can double site activation rates and trim start-up staff by 30 to 50 percent. One large company already runs a multi-agent trial copilot off a clinical control tower, watching activation, enrollment, and data flow, then flagging weak sites. It wants those agents talking directly to investigators and monitors for routine chores.

The stranger idea is a virtual regulator of health authorities. McKinsey describes a digital twin that reads a draft against past agency feedback, worked examples, and ICH E3 and E6, then flags language, unsupported claims, and structural holes before a dossier leaves the building. The aim is fewer rework loops and up to 40 days off review time. The same note says this suite is still in development.

So the industry is training a simulator of the FDA and EMA on old reviews, while those agencies are still writing how they want AI evidence presented. The twin is a rehearsal, not a stamp.

The FDA is committed to supporting innovative approaches for the development of medical products by providing an agile, risk-based framework that promotes innovation and ensures the agency’s robust scientific and regulatory standards are met.

Robert M. Califf, M.D., FDA Commissioner, January 6, 2025 announcement

What the FDA and EMA Now Ask of Drug AI

On January 6, 2025, FDA issued its first guidance on AI used to support a regulatory decision about a drug or biologic’s safety, effectiveness, or quality. It is a draft AI credibility framework for drugs, not a final rule. CDER still lists it as draft guidance. The text offers a seven-step, risk-based way to show that a model is credible for a stated context of use.

The agency said the draft grew from a December 2022 Duke Margolis workshop, more than 800 comments on 2023 discussion papers, and more than 500 drug and biologic submissions with AI pieces since 2016. Context of use is the load-bearing phrase. How you use the model, and what happens if it is wrong, sets how hard the credibility work must be.

A year later the bar moved sideways, not down. On January 14, 2026, FDA’s drug and biologics centers and the European Medicines Agency put out 10 guiding principles with EMA for AI across the medicine life cycle. EMA said the list will underpin future AI guidance on its side of the Atlantic. Guideline work in the EU is already underway. Principle one is the tell.

THE 10 FDA-EMA PRINCIPLES, SHORT

• Human-centric by design: People keep ethical control, and AI is built so users can see limits and failure modes.
• Risk-based approach: Checks and oversight scale with context of use and model risk.
• Clear context of use: Each system needs a stated role and scope, not a vague claim that it “does R&D.”
• Data governance and documentation: Training data, lineage, and records have to stand up to GxP-style scrutiny.
• Life cycle management: Models drift, so monitoring after launch is part of the product, not a patch.

The other five cover standards, mixed expertise, model design, risk-based performance tests, and clear essential information. None of that reads like a green light for a closed loop that writes a protocol, runs a robot, and files the CSR. It reads like a demand that someone named can explain the loop.

Clinical Study Reports in Six Weeks, Not 12

If discovery agents win on search, development agents win on paperwork. Data managers still drown in queries. Statistical programmers still walk a rigid path from capture to SDTM to ADaM to static tables. McKinsey says agent platforms that mix language models with domain rules can cut queries by two to three times, shrink database build from 2 to 3 months to under 2 weeks, and lift programmer output by up to 60 percent.

Document agents are further along. McKinsey, working with a large pharmaceutical client, fielded a multi-agent setup that plans data pulls, runs analyses, and drafts a clinical study report. That deployment cut drafting errors by 50 percent and took the path from database lock to a finished report from about 12 weeks to 6 weeks. The same architecture is being pointed at protocols and consent forms. Separate McKinsey work puts documentation agents at 75 to 80 percent productivity gains on first drafts.

MCKINSEY’S CLINICAL TIME MAP

Those bars are projections across functions, not the CSR client result. They also bake in a tax. Someone has to supervise the agents. Delphine Zurkiya of McKinsey has said about 40 percent of what agents can do is work that is not getting done now, the long tail of papers, contracts, and small segments that never justified a human week. That is extra labor, not a straight swap for today’s headcount.

Thirty Specialized Agents Need One Foundry

McKinsey’s September 2025 study did not sample a few use cases. It analyzed 270 workflows and 1,200 tasks across 180 job families. Full take-up, the firm estimates, could add 5 to 13 percentage points of incremental growth in pharma and 3.4 to 5.4 percentage points of EBITDA over three to five years, on top of work already in train.

THE MCKINSEY WORKLOAD SNAPSHOT

• Scope: 270 workflows, 1,200 tasks, 180 job families in pharma and medtech.
• Custom build: Nearly 60 percent of workflows will need agents made for that job, not a generic chat box.
• Lab slice: Wet labs, data analytics, and regulatory support could free 21 to 30 percent of capacity, a narrower cut than the 25 to 40 percent organization-wide figure.
• Runtime: End-to-end development is described as 30-plus specialized agents around one enterprise foundry that designs, trains, and runs them.

That foundry is the unglamorous half of the story. Pharma data is still split across functions, master data is messy, and GxP records have to be traceable. An agent that drafts an SOP or a deviation report is only as good as the history it can see. Language models still invent plausible sentences. In this industry a wrong sentence can become a wrong dose, a wrong site, or a wrong safety call.

Human-in-the-loop is not a slogan here. It is the only way the January 2026 principles and the 2025 draft stay in the same sentence as a robot that pipettes on its own. New jobs follow that logic: people who orchestrate agents, people who audit them, people who own the decision when the model and the clinician disagree.

Boehringer’s License Puts K Pro Inside Discovery

The checkbooks moved in 2026 anyway. Owkin, which calls itself an agentic AI firm, licensed K Pro to Boehringer Ingelheim on September 2, 2026, with multimodal oncology data and a plan to generate new immunology data. Terms were not disclosed. The deal follows a 2025 pilot in which Owkin used its MOSAIC spatial atlas to read a gene target’s tumor microenvironment.

K Pro is sold as one place to query data and run repeatable analyses, with both human-led work and self-driven campaigns for hypotheses, tests, and ranking. AstraZeneca took a three-year K Pro license in May 2026. Sanofi signed a five-year pact on June 5, 2026, on top of a €90 million oncology partnership that started in 2021. Thomas Clozel, Owkin’s co-founder and CEO, said the next decade of drug development will be defined by deep multimodal patient data and AI scientists that can reason over it.

On the clinical-ops side, San Diego-based Faro AI raised $37.3 million in Series B financing on August 26, 2026, co-led by Merck Global Health Innovation Fund and S32. Faro said six of the world’s 10 largest drugmakers already use its platform. CEO Scott Chetham said customers want agents that grasp development intent and act across processes, not tools that only draft text. Mike Morgan, a principal at Merck’s fund, said the opening is to apply AI reliably to complex clinical workflows.

THE 2025-2026 CLOCK

1. January 6, 2025: FDA posts draft guidance on AI used to support drug and biologic regulatory decisions.
2. September 8, 2025: McKinsey publishes its life-sciences agentic AI workflow study.
3. December 11, 2025: McKinsey describes CSR compression, trial copilots, and a virtual regulator still being built.
4. January 14, 2026: FDA and EMA release 10 good-practice principles for AI in medicine development.
5. March 2026: The Drug Discovery Today agents paper logs Kiin Bio, Tater, onepot.ai, and related case studies.
6. May 2026: AstraZeneca licenses Owkin’s K Pro for three years.
7. June 5, 2026: Sanofi adds a five-year K Pro collaboration.
8. August 26, 2026: Faro closes a $37.3 million Series B for clinical-development agents.
9. September 2, 2026: Boehringer Ingelheim licenses K Pro and multimodal data from Owkin.

Big pharma is not waiting for final FDA text. It is renting named AI scientists and standing up ontology layers so agents can read a protocol the way a development team does.

WHAT WE KNOW

• Time cuts: Independent 2026 case studies document hour-scale discovery design cycles that used to take weeks or months.
• Live licenses: AstraZeneca, Sanofi, and Boehringer Ingelheim have all taken Owkin’s K Pro in 2026.
• Rules: FDA’s 2025 AI drug guidance remains draft, and the 2026 FDA-EMA list is principles, not a binding playbook.

WHAT IS UNCONFIRMED

• BI economics: Owkin and Boehringer Ingelheim have not disclosed the value of the September 2 license.
• Virtual regulator: McKinsey still describes the digital twin of FDA and EMA as in development, not as a filed product.
• Trial length: The 12-month duration cut and “twice as many trials” figures are McKinsey projections, not a published Phase 3 result.

Boehringer’s chemists will now query licensed patient data through K Pro, while CDER’s first AI drug guidance stays on the draft pile and EMA writes the next layer of rules. The hours are already here. The signature on the output is still a person’s.

Disclaimer: This article is news reporting and analysis of public company statements, regulator documents, and published research. It is informational only and is not medical, legal, or investment advice, and it is not a recommendation to start, stop, or change any medicine, trial, or AI system used in care or drug making. Readers who need advice on a treatment, a clinical protocol, or a regulatory filing should consult a licensed physician, a qualified regulatory counsel, or the relevant health authority before acting. Figures, product names, and guidance status reflect the cited sources on the dates those sources carry and can change as agencies finalize rules and companies update deals.