Loading
Skip to content
Miracle Minds AI
  • Try “HIPAA”, “model risk”, “FinOps” or “agentic”.

Clinical trials and development

AI in clinical development pays off where it removes a delay you can measure, such as a screening backlog, and where you can show a regulator the model is credible for that exact use. The proof required rises with how much the output affects participants or trial results.

Part of AI for Life Sciences & Healthcare. 8 sources, checked .

of phase I to IV protocols had at least one amendment, in a 2022 Tufts CSDD study of 950 protocols from 16 drug companies and contract research organizations
76%
of phase I to IV protocols had at least one amendment, in a 2022 Tufts CSDD study of 950 protocols from 16 drug companies and contract research organizations
Therapeutic Innovation & Regulatory Science (Springer), 2024(opens in a new tab)
of randomized patients were found eligible by the end of the trial with AI-assisted prescreening, against 12.7% with manual screening, among 4,476 patients at one US health system (May to September 2024)
20.4%
of randomized patients were found eligible by the end of the trial with AI-assisted prescreening, against 12.7% with manual screening, among 4,476 patients at one US health system (May to September 2024)
JAMA, 2025(opens in a new tab)
submissions with AI components seen by FDA's Center for Drug Evaluation and Research from 2016 to 2023
Over 500
submissions with AI components seen by FDA's Center for Drug Evaluation and Research from 2016 to 2023
U.S. Food and Drug Administration, Center for Drug Evaluation and Research, 2026(opens in a new tab)

The problem

Trial delay often starts in the protocol. A 2022 Tufts Center for the Study of Drug Development study of 950 protocols from 16 drug companies and contract research organizations found that 76% had at least one amendment, up from 57% in 2015, and the mean number of amendments per protocol rose to 3.3.1 Moving an amendment from identified need to last oversight approval took an average of 260 days.1 The same study deemed 77% of amendments unavoidable, with regulatory agency requests and changes to study strategy as the top reasons, so a protocol design tool can only work on the rest.1

The strongest trial evidence is in patient matching. In a randomized trial at Mass General Brigham from May to September 2024, 4,476 patients were assigned to manual chart review or AI-assisted prescreening for a heart failure trial.2 The AI-assisted arm reached eligibility determinations faster, and at the end of the trial it had 37 patients left to screen against 887 in the manual arm.2 Among patients whose screening was completed, the share found eligible was similar in both arms (20.8% and 21.1%), so the gain was speed through the backlog. Enrollment was 35 patients against 19. The authors list a single center and a single trial as limitations.2

Good clinical practice was rewritten while these tools arrived. ICH endorsed the E6(R3) principles and Annex 1 in January 2025, endorsed Annex 2 on June 3, 2026, and adopted the consolidated guideline on June 16, 2026. Annex 2 covers trials with decentralized elements, pragmatic elements and real-world data.3 The guideline does not mention AI by name. It asks sponsors to build quality into the design, identify critical to quality factors and manage risk in proportion, and it describes centralized monitoring as a way to find systemic or site-specific issues, including protocol noncompliance and potentially unreliable data.3 Computerized systems are to be validated on the basis of risk before use.3

AI-specific guidance is still settling. FDA's guidance on AI to support regulatory decision-making for drugs and biologics was issued as a draft in January 2025 and was still listed as a draft in November 2026.4 It sets out a seven-step framework for establishing the credibility of a model for a specific context of use, with effort scaled to model risk.4 It does not cover AI used for operational efficiency, such as internal workflows or drafting a submission, where that use does not affect patient safety, drug quality or the reliability of study results.4 EMA's 2024 reflection paper is strict about models used to analyze pivotal trials: they should be frozen and documented before database lock, and incremental learning is not accepted.6 In January 2026 the two agencies published 10 guiding principles of good AI practice in drug development.7 FDA says its drug center has seen over 500 submissions with AI components from 2016 to 2023.8

What we build

  1. 01

    Protocol feasibility and amendment risk review

    We test draft eligibility criteria and visit schedules against your past protocols and patient data, flag the criteria most likely to force an amendment or starve enrollment, and leave the design decision with the study team.

  2. 02

    Patient prescreening and eligibility matching

    Assistants that read structured and unstructured records against each criterion, show the passage behind every answer, and queue candidates for study staff to confirm before anyone is contacted.

  3. 03

    Site selection and enrollment forecasting

    Models that rank sites on past enrollment, startup time and data quality, with every forecast tracked against actuals so the ranking is corrected while the trial runs.

  4. 04

    Risk-based quality management and central monitoring

    Centralized analytics tied to your critical to quality factors and quality tolerance limits. Site and data anomalies go to monitors for review, and each signal and its disposition is logged.

  5. 05

    Data cleaning, query drafting and trial documents

    Assistants that propose data queries, check trial master file completeness and draft clinical study report sections from locked tables, with each output traced to its source and approved by a named reviewer.

Where it goes wrong

The failure modes we design against from the first week, and the control for each.

  • Calling an evidence-producing model an operational tool

    FDA's draft leaves operational uses out of scope only where they do not affect patient safety, drug quality or the reliability of study results. Its own example, a model that decides which participants skip inpatient monitoring, is rated high risk.4 Control: classify every use by question of interest, context of use and model risk before build, and engage the agency early on anything rated high.

  • Models that change during a pivotal trial

    For models used to analyze pivotal trials, EMA does not accept incremental learning, and non-prespecified changes after a dataset has been opened make the results post hoc.6 Control: freeze and version the model and the data pipeline, record both in the statistical analysis plan before database lock, and route any change through change control.

  • Faster screening that only moves the bottleneck

    In the randomized prescreening trial, AI cleared the screening backlog, yet the AI-assisted arm produced 35 enrollments (1.6%).2 Control: measure enrollment and screen failures as well as screening speed, and staff the consent and enrollment steps for the larger flow of candidates.

  • Vendor AI outside sponsor oversight

    Under ICH E6(R3) the sponsor keeps ultimate responsibility for activities transferred to service providers, and systems must be validated as fit for purpose, including those developed by other parties.3 Performance can also change over time as input data drift.4 Control: keep a per-trial inventory of AI systems with validation status, audit trails that stay on and are reviewed, scheduled re-evaluation, and contract rights to vendor validation records.

Rules that apply

  • ICH E6(R3) Guideline for Good Clinical Practice (consolidated with Annex 2, June 2026)

    Primary text on database.ich.org(opens in a new tab)

    Sponsors are expected to apply proportionate, risk-based quality management focused on critical to quality factors, validate computerized systems on the basis of risk, and keep ultimate responsibility for activities transferred to service providers.

  • FDA draft guidance on AI to support regulatory decision-making for drug and biological products (January 2025)

    Primary text on fda.gov(opens in a new tab)

    Recommends a seven-step, risk-based framework to establish and document the credibility of an AI model for its context of use, with life cycle maintenance and early engagement with FDA. It is a draft and not for implementation.

  • 21 CFR Part 11, electronic records and electronic signatures

    Primary text on ecfr.gov(opens in a new tab)

    Closed systems that hold required electronic records need validation to ensure accuracy and reliability, access limited to authorized individuals, and secure, computer-generated, time-stamped audit trails that do not obscure earlier entries.

  • EMA reflection paper on the use of AI in the medicinal product lifecycle (September 2024)

    Primary text on ema.europa.eu(opens in a new tab)

    AI used in trials should meet ICH E6. For uses with high regulatory impact or high patient risk that EMA has not qualified, model architecture, development logs and training data may be requested at assessment or GCP inspection.

How to measure it

  • Amendments per protocol and the share judged avoidable, against your own baseline
  • Days from candidate identification to eligibility determination, and patients still waiting to be screened
  • Enrollment against plan and screen failure rate, by site
  • Central monitoring signals raised, confirmed on review and closed, by site
  • Model performance against its acceptance criteria at each scheduled re-evaluation

Questions buyers ask

Has the FDA finalized its guidance on AI in drug development?

Not as of November 2026. FDA issued the draft guidance on using AI to support regulatory decision-making for drug and biological products in January 2025, and its guidance page still lists it as a draft. It describes a seven-step, risk-based credibility framework tied to a model's context of use. In January 2026 FDA and EMA also published 10 guiding principles of good AI practice. Plan against the draft and watch for the final version.

Does ICH E6(R3) allow AI in clinical trial monitoring?

E6(R3) does not mention AI by name, but it sets the conditions such a tool has to meet. It supports centralized monitoring and data analytics to find site and data issues, expects computerized systems to be validated according to risk, and expects audit trails to stay on and be reviewed. The sponsor stays responsible when a vendor runs the tool. Treat an AI monitoring model as a computerized system under those rules.

Can AI speed up patient recruitment for clinical trials?

It can speed up screening. In a randomized trial of 4,476 patients at one US health system, published in 2025, AI-assisted prescreening reached eligibility decisions faster than manual chart review and led to 35 enrollments against 19. Among patients fully screened, the share found eligible was about the same in both arms, so the gain came from clearing the backlog. The study covered one center and one heart failure trial, so test it on your own protocol.

Sources

Every figure on this page comes from the sources below. We checked each link on .

  1. 1.New Benchmarks on Protocol Amendment Practices, Trends and their Impact on Clinical Trial Performance(opens in a new tab)Therapeutic Innovation & Regulatory Science (Springer), March 4, 2024 · link.springer.com
  2. 2.Manual vs AI-Assisted Prescreening for Trial Eligibility Using Large Language Models: A Randomized Clinical Trial(opens in a new tab)JAMA, February 17, 2025 · pmc.ncbi.nlm.nih.gov
  3. 3.ICH Harmonised Guideline: Guideline for Good Clinical Practice E6(R3) (final version, adopted 16 June 2026)(opens in a new tab)International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use (ICH), June 16, 2026 · database.ich.org
  4. 4.Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products: Draft Guidance for Industry and Other Interested Parties(opens in a new tab)U.S. Food and Drug Administration, January 2025 · fda.gov
  5. 5.21 CFR Part 11: Electronic Records; Electronic Signatures(opens in a new tab)Electronic Code of Federal Regulations (FDA, HHS), March 20, 1997 · ecfr.gov
  6. 6.Reflection paper on the use of Artificial Intelligence (AI) in the medicinal product lifecycle (EMA/CHMP/CVMP/83833/2023)(opens in a new tab)European Medicines Agency, September 9, 2024 · ema.europa.eu
  7. 7.Guiding principles of good AI practice in drug development(opens in a new tab)European Medicines Agency and U.S. Food and Drug Administration, January 2026 · ema.europa.eu
  8. 8.Artificial Intelligence for Drug Development(opens in a new tab)U.S. Food and Drug Administration, Center for Drug Evaluation and Research, May 1, 2026 · fda.gov