Clinical trial site feasibility and documents

• Feasibility answers are read and turned into a structured comparison, whatever format they arrive in, with every answer linked to the page it came from.
• Each site's answers sit beside its own past performance on the sponsor's earlier studies: recruitment against its estimate, set-up time, data query rates and protocol deviations.
• Estimates far above a site's own track record are flagged, so over-promising is discussed before selection, not discovered at month six.
• Eligibility fit uses aggregate counts of potentially eligible patients, produced by the site or an NHS secure data environment. Patient records never leave those environments.
• The TMF is checked every day against the expected document list for each site and country. Missing, expired or wrong-version documents go to their owner straight away.
• When a protocol modification is approved, every site document that needs updating is listed and tracked. The study team selects sites and the sponsor stays responsible; the AI never selects a site or contacts one.

• MHRA and HRAMedicines for Human Use (Clinical Trials) (Amendment) Regulations 2025, in force 28 April 2026
  →

  Combined MHRA and ethics review, a notifiable route for lower-risk trials, Route B for eligible substantial modifications, and for the first time legal duties to register trials and publish a results summary within 12 months of the end of the trial. The sponsor classifies each modification. The AI can list what a modification touches, but a person decides its category and route.

• MHRAICH E6(R3) Good Clinical Practice, implemented in the UK on 28 April 2026
  →

  The sponsor stays responsible for work it transfers to service providers, and must select investigators and sites that are qualified and adequately resourced. Quality management is risk-proportionate. Computerised systems must be fit for purpose, validated in proportion to risk, and keep audit trails. Feasibility scoring and TMF checks are such systems.

• MHRAArchiving and retention of clinical trial records (updated 28 April 2026)
  →

  For trials applied for from 28 April 2026, the TMF is kept for at least 25 years after the trial ends and must be readily available for inspection. A person confirms every filing the AI suggests, originals or certified copies stay the record, and any migration is validated and documented.

• HRAConfidentiality Advisory Group (section 251) and the national data opt-out
  →

  Feasibility uses aggregate counts, produced inside the site or an NHS secure data environment. If anyone outside the care team needs identifiable records without consent, that needs section 251 support through the Confidentiality Advisory Group, and in England the national data opt-out then applies. Properly anonymised and aggregate counts are outside the opt-out.

• ICOUK GDPR and the Data Protection Act 2018
  →

  Site documents carry personal data about investigators and staff: CVs, training records and signatures. That means a DPIA before the pilot, processing in UK regions, retention set by the TMF rules, and a contract that stops providers training their models on your data. If performance is scored per investigator, tell them, and let them comment on their record.

• HRAPlan for safe AI in health and social care research (2026–28)
  →

  The HRA is clarifying expectations for AI used in writing protocols, participant materials and lay summaries. Any AI-drafted participant-facing text is reviewed by the study team and goes through the usual approval route.

On a phone, the table shows the platform you pick. Sovereign / on-prem means open, portable components you can run in your own data centre or with a UK or European provider; IONOS is shown as one example, and others exist. What sovereign cloud means →

• Agreed feasibility measures, shared with sites, so they know how their past performance is used and can comment on it.
• An exception queue for documents the AI can't classify confidently, owned by the TMF team.
• A weekly sample check of AI-suggested filing, and a monthly review of the feasibility comparisons against actual recruitment.
• A named owner for the expected document list, prompts and models, with change control like any other validated system.
• Study managers and TMF staff trained before go-live, including what the AI can't see: a site's staffing changes and local knowledge.
• Service provider oversight written into contracts, so a CRO's use of AI is visible to the sponsor that stays responsible.

1. 3–4 weeks

   Discovery

   Map today's feasibility and TMF process, measure the baseline, confirm CTMS and eTMF API access, agree the feasibility measures, and confirm with data providers how aggregate counts are produced.

2. Gate · must be true to passBaseline agreed, supplier access confirmed, DPIA and risk assessment started, QA involved.

3. 8–10 weeks

   Pilot

   One study in shadow mode: the AI compares sites and checks the TMF, the team works as normal, and the two are compared.

4. Gate · must be true to passExtraction and filing accuracy within agreed limits, no patient-identifiable data in the platform, QA sign-off.

5. 4–6 weeks

   Production

   Feasibility comparisons and TMF checks go live for new studies, with monitoring, support and a rollback plan.

6. Gate · must be true to passValidation evidence, monitoring, support model and team training in place.

7. Ongoing

   Scale

   Add studies and therapy areas, then the document tracking for protocol modifications and results deadlines.

8. Gate · must be true to passEach new study passes the same pilot gate.

R ResponsibleA AccountableC ConsultedI Informed· not involved

Where I'd fit

1. 1

   Goes wrong

   The demo used tidy questionnaires. Real ones are scanned, half-completed and in a dozen formats.

   How the design avoids it

   Pilot on last year's real responses for one study, every format.

2. 2

   Goes wrong

   Past performance punishes sites that had a hard study, or new sites with no history.

   How the design avoids it

   Show the numbers and their context, let sites comment, and keep a route in for new sites.

3. 3

   Goes wrong

   Feasibility quietly starts using patient-level data.

   How the design avoids it

   Aggregate counts only, produced inside the site or a secure environment, and guardrails that block identifiers.

4. 4

   Goes wrong

   TMF staff accept every filing suggestion, so errors are filed faster.

   How the design avoids it

   A filing correction measure, a weekly sample check, and an exception queue for low-confidence documents.

5. 5

   Goes wrong

   A CRO uses AI on the sponsor's study and the sponsor doesn't know.

   How the design avoids it

   Service provider oversight in the contract, and the same logs available to the sponsor.

1. From about April 2027
   For trials in scope of the new transparency rules, results summaries are due within 12 months of the end of the trial, so the first deadlines fall about a year after 28 April 2026.

   Track end-of-trial dates and results deadlines per trial alongside the TMF checks.

2. 15 January 2027 (EU)
   ICH E6(R3) Annex 2, on decentralised elements, pragmatic trials and real-world data, was adopted by ICH on 3 June 2026 and takes effect in the EU on 15 January 2027. No MHRA date found yet.

   If feasibility uses real-world data, design for Annex 2's expectations now.

3. No date yet
   The MHRA's Strategy 2030 (14 October 2026) plans an AI-enabled Clinical Trial Protocol Checker for sponsors, and initial review of first-in-human trials within 14 days.

   Keep protocols and modification history structured and versioned, so they can be checked quickly by your tools and the MHRA's.

4. 2026–2028
   The HRA's AI plan: clarifying when AI work counts as research, when health data can be accessed with AI to find and contact people about research, and expectations for AI in protocols and participant materials.

   Keep feasibility on aggregate counts and record AI use in study documents, so the design fits the guidance as it lands.

5. From end of 2026 (phased)
   The Health Data Research Service: first services due by the end of 2026, giving approved researchers access to NHS data through secure environments.

   Design the eligibility count as a query run inside a secure environment, so a national service can replace site-by-site requests.