WCG's early-access customers cut more than 20 hours of manual review from each study activation, the company said on Oct. 12. Its new tool, Activation Intelligence, extracts study information once, checks it against source documents, and reuses it across activation workflows. An AI assistant answers congruency questions so site teams can catch problems before they delay a study.
Twenty hours is a modest number next to the promises attached to clinical-trial AI. It is also the most useful figure in the sector this quarter. It measures work removed from a system that already exists.
The constraint on clinical-trial AI has moved. Model quality is no longer the limiting factor for most operational use cases. The limit is whether an agent can reach the systems of record where trials actually run: CTMS, EDC, eTMF, safety databases, and the site's own workflow logs.
Five vendor moves since June point the same way. Each one sells access to a workflow, not a model.
Traceability — which data fed the model and which human reviewed the output — is becoming the product.
The EU's ICH E6(R3) Annex 2 takes effect on Jan. 15, 2027, which turns audit trails into a procurement requirement.
What changed in 2026
WCG is the clearest case because it shipped a narrow tool with a measurable result. Activation Intelligence sits inside a broader Study Intelligence Hub. The company bought The Contract Network in April 2026 to spot inconsistencies in study documents before they cause amendments. Its ClinSphere platform already handles trial design, site feasibility and enrollment.
Medidata went wider. At a Seoul press conference in June, Jeff Ventimiglia, its senior vice president of offerings and portfolio, described a shift from product-based AI to agentic flows embedded in the software trial teams already use. Its orchestrator, Dot, is meant to keep clinical data inside the workflow rather than export it to a separate AI environment. Medidata says it processes about 70 billion data points a year and runs more than 7,600 active trials on its platform. Medidata Plus, launched in March, bundles the AI layer as a subscription.
Veristat attacked a workflow that had barely changed in decades: biostatistics. Its InStat platform automates parts of statistical analysis and reporting while capturing metadata, transformations and analytical decisions as results are generated. One study, according to the company, required more than 21,000 lines of code to produce submission-ready outputs.
Florence Healthcare pointed at the layer none of the platform vendors own. Sponsor systems capture outcomes, milestones and documents. They miss the day-to-day work at research sites. Florence estimates that adding site-level workflow data can raise the usable operational signal by 40–80%, depending on the model.
As we wrote in October, Rivercell's €22M round was framed as building the lab before the model. The same ordering now applies to trials: the data layer comes first, and the agent is the last mile. The Rivercell analysis made the case for infrastructure over model ambition.
The stack has four layers, and one is missing
Strip away the marketing and the clinical-trial AI market separates into four layers. Three are crowded. The fourth is where most deployments stall.
| Layer | Job in the trial | 2026 examples | Where value sits |
|---|---|---|---|
| Systems of record | Store the data regulators inspect | Medidata Rave, Veeva Vault, Oracle Clinical One | ✔ Incumbents; high switching cost |
| Orchestration | Route AI output into operational action | Medidata Dot, WCG Study Intelligence Hub, Oracle Life Sciences AI Data Platform | ◐ Contested; the current battleground |
| Analytical tools | Do one job well | Veristat InStat, Deep 6 AI, Quibim | ◐ Defensible, easily absorbed |
| Site data layer | Capture how work actually happens | Florence, CRIO | ✗ Thin; the missing layer |
Layer map assembled from vendor disclosures and trade coverage, 2026
The orchestration layer is where the 2026 deals cluster. Oracle and AstraZeneca announced a non-binding collaboration on Oct. 20 to use real-world data and AI for trial matching, safety monitoring and evidence generation. The first program targets hypertension care and research in Abu Dhabi with M42. Oracle had already launched its Life Sciences AI Data Platform, which it markets around agentic reasoning and customizable agents.
None of these products is a better model. They are connectors, permission systems and audit logs with a language model attached.
Traceability is the product
The strongest argument for that reading comes from Veristat, because its constraint is regulatory rather than technical. Statistical programming has always produced a submission package: code, outputs, and a path a reviewer can follow. Generative AI can write the code faster. It cannot, on its own, prove what changed between drafts.
InStat captures that path while the analysis runs. Instead of reconstructing the logic after the fact, the platform records metadata, transformations and decisions as they happen. The result is a live audit trail.
AI should not be determining whether a drug is safe or effective. It's not responsible to let AI decide whether the primary endpoint of a pivotal trial has succeeded.— Kyle McBride, Vice President of AI and Innovation, Veristat
McBride's framing is deliberate. AI works as an analytical assistant for exploration, pattern recognition and insight generation. Core statistical calculations and regulatory decisions stay with humans. The commercial pitch is defensibility.
“What regulators really want is traceability,” he told Clinical Trials Arena. That sentence explains why the orchestration layer is being built now, and why pure model vendors are being folded into larger platforms.
The data shadow at the site
Traceability has a precondition that no sponsor system satisfies: the data has to exist in a usable form. Florence's 2026 research puts the gap in numbers. About 24% of sites qualify as technology leaders and another 24% remain laggards. Only 12% of respondents say they use AI consistently. The company reports 90% adoption for its electronic investigator site file, which suggests sites will use tools that fit their workflow and ignore tools that do not.
The site layer matters because trials lose time in places the sponsor dashboard never sees: a coordinator chasing a signature, a protocol question answered by email, a screening log maintained in a spreadsheet. Florence estimates that site workflow data adds roughly 60% more signal to site performance models and as much as 80% to activation-readiness models.
A CRIO panel reported by Clinical Trial Vanguard in September reached a similar conclusion from the opposite direction. About three-quarters of sites now use eSource in active studies, according to a 2024 RealTime Reports survey, yet the panel's advice was to sequence the work: move off paper, set an AI governance policy, give staff enterprise LLM accounts, then attempt custom workflows. The marker they proposed is blunt. For any AI-assisted output, a site should be able to name the data that fed the model and the human who reviewed the result.
The regulatory clock is now a product deadline
Regulators did not wait for the market to settle. The joint FDA and EMA principles of good AI practice, published on Jan. 14, 2026, ask for proportionate validation and oversight based on the context of use. The FDA's draft guidance on AI for regulatory decision-making, issued in January 2025, is still a draft as of November 2026.
The binding date is in Europe. ICH E6(R3) Annex 2 was adopted on June 3, 2026 by ICH and on June 25 by the CHMP. It takes effect in the EU on Jan. 15, 2027. E6(R3) contains no explicit AI governance requirements, but its risk-based framework puts the documentation burden on the organization using the tool. A site or sponsor that cannot show its data lineage has a problem that a better model will not fix.
The FDA's early-phase AI pilot, opened for comment under docket FDA-2026-N-4390, made the same point from the industry side. One comment filed by ServiceNow argued that standalone models have no persistent connection to CTMS, EDC or safety databases, and urged the agency to require bidirectional data flows, automated action routing and complete audit-trail capture. That is a description of an integration layer, written as a regulatory request.
Jan. 14, 2026 — FDA and EMA publish 10 guiding principles for AI in drug development
June 11, 2026 — ICH M11 structured protocol template takes effect in the EU
Jan. 15, 2027 — ICH E6(R3) Annex 2 takes effect in the EU
Still pending — FDA draft guidance on AI for regulatory decision-making (docket FDA-2024-D-4689)
Who captures the value
The value does not accrue to the model. It accrues to whoever owns the workflow, the permissions and the audit log at the point where a decision is made. That is why Medidata can sell an orchestration subscription on top of 7,600 active trials, and why Oracle can trade distribution through health systems for AstraZeneca's clinical expertise.
It is also why the site layer is the most interesting weak spot. Florence, CRIO and similar vendors are small relative to the EDC incumbents, but they sit closest to the work that generates the missing data. If the EU deadline pushes sponsors to demand inspection-ready lineage, the site layer becomes a acquisition target rather than a niche tool.
One caution belongs here. The AstraZeneca–Oracle agreement is non-binding and exploratory. Each project must clear legal, privacy, regulatory and ethical review before it proceeds. The direction is clear; the revenue is not.
Where the stack settles
The next twelve months will not produce a model that makes trials autonomous. They will produce a smaller set of platforms that can prove what their agents did, on which data, with which human in the loop. WCG's 20 hours is a useful early benchmark because it counts removed work rather than claimed intelligence.
Probability: 65% — the EU Annex 2 deadline and the FDA pilot both reward traceability over raw accuracy.
✅ Arguments for
The EU deadline is fixed, and ICH E6(R3) shifts documentation duty onto the tool user.
Confirmation criteria: a major CRO discloses audit-trail revenue or bundles its AI layer into a compliance offering before Q4 2027.
❌ Arguments against
A model breakthrough in trial design could reset the value toward accuracy and away from workflow.
Disconfirmation criteria: the 2027 procurement cycle continues to reward raw model benchmarks over audit evidence.
The trial-AI stack is being rebuilt around the least glamorous part of the problem. That is usually where durable infrastructure gets built.