Cleveland·Miami·Basel·Tel AvivForce Multipliers  ·  Elevate Humanity
LeverVenture
Home / Reference / AI Diligence in Life Sciences
Reference

AI Diligence in Life Sciences

AI diligence in life sciences is the evidence-based assessment of whether a company's artificial intelligence claims are supported by its data, validation, regulatory record, and quality system.

Reviewed by Peleg Chevion, Managing Partner

AI drug discovery is the use of machine-learned models in the discovery stage of the drug product life cycle, to predict the bioactivity and physicochemical properties of candidate molecules, screen compounds virtually, design molecules de novo and validate drug targets (Paul et al., Drug Discovery Today, 2021). FDA's draft drug guidance places AI used for drug discovery outside its scope, so no FDA credibility record exists to check. Diligence asks instead whether wet-lab experiments confirmed the model's predictions, whether any model-generated candidate has reached an investigational new drug application, and who holds rights to the training data.

Mechanism

The assessment follows the regulatory channel that governs the product, and its first finding is which FDA documents are final. As of October 2026, FDA's pages show the device lifecycle guidance and the drug development guidance as drafts, and the change control plan guidance as final.

For a device, the unit of analysis is the AI-enabled device software function: a software function that meets the device definition in section 201(h) of the Federal Food, Drug, and Cosmetic Act and implements one or more AI models. FDA's draft guidance Artificial Intelligence-Enabled Device Software Functions (January 7, 2025, docket FDA-2024-D-4488) organizes a submission around device description, labeling, risk assessment, data management, model development, validation, performance monitoring, and cybersecurity.

A predetermined change control plan authorizes planned model changes in advance. FDA's final guidance (issued December 4, 2024, reissued August 18, 2025, docket FDA-2022-D-2628) describes three components: a Description of Modifications, a Modification Protocol, and an Impact Assessment. Under section 515C of the FD&C Act (21 U.S.C. 360e-4), a change consistent with a plan approved or cleared by FDA does not require a supplemental application (for a premarket approval) or a new 510(k).

For drugs and biologics, FDA's draft guidance Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making (January 2025, docket FDA-2024-D-4689) sets a seven-step credibility framework: define the question of interest, define the context of use, assess model risk, plan credibility activities, execute the plan, document the results, and determine adequacy. Model risk combines model influence with decision consequence. The draft excludes drug discovery and operational efficiencies.

Three further rules govern the record around the model. The Quality Management System Regulation, 21 CFR Part 820 (89 FR 7496), took effect February 2, 2026 and requires a quality management system that complies with ISO 13485. Decision support interventions in certified health information technology fall under 45 CFR 170.315(b)(11), revised by the HTI-1 final rule (89 FR 1192, January 9, 2024).

A deregulatory proposed rule published December 29, 2025 (HTI-5, 90 FR 60970; comments closed February 27, 2026) would remove that criterion's source attribute and intervention risk management requirements; no final rule had published in the Federal Register as of October 10, 2026, so the HTI-1 text remains in force. Deceptive performance claims fall under section 5 of the FTC Act, 15 U.S.C. 45(a)(1), and the FTC expects a reasonable basis for a claim before it is disseminated.

Worked Example

FDA's own drug example asks which trial participants are low risk and need no inpatient monitoring after dosing. The model would be the sole determinant of monitoring, so model influence is high, and a misclassified participant could suffer a life-threatening reaction, so decision consequence is high. FDA rates the model risk as high.

A diagnostics company says its imaging algorithm improves continuously. Without a change control plan, a retraining that could significantly affect safety or effectiveness ordinarily requires a new 510(k) under 21 CFR 807.81(a)(3); an authorized plan covering the retraining permits the change without one under section 515C.

Diligence Questions

  1. Which product functions use AI, and does each meet the device definition in section 201(h)?
  2. Which marketing authorization covers each function, and is the product on FDA's list of AI-enabled medical devices?
  3. Does the authorized indication match the marketed indication?
  4. Is there an authorized change control plan, and does every planned update fall inside it?
  5. Did the validation data come from sites different from the development sites?
  6. Has performance been characterized across subgroups to test for AI bias?
  7. For drug development uses, what are the context of use and the assessed model risk, and for a discovery platform, what share of the model's predictions did experiments confirm?
  8. What does the postmarket performance monitoring plan measure, and what has it detected?
  9. Does the quality management system conform to 21 CFR Part 820 as amended?
  10. For decision support in certified health information technology, are the source attributes under 45 CFR 170.315(b)(11), which HTI-5 would remove, supported?
  11. Was each public performance claim substantiated before it was disseminated?
  12. Does the company hold documented rights to every training and validation dataset, and was any protected health information de-identified under 45 CFR 164.514(a) and (b), released in a limited data set under a data use agreement (164.514(e)), or used under a patient authorization (164.508)?

What It Means for a Limited Partner

A limited partner reads an AI claim in a portfolio company as it reads an unaudited projection, by asking what independent record exists. For a regulated product, FDA's authorization, the labeling, and any change control plan are that record. Scope: the authorized and marketed indications should match. Durability: a model that changes after authorization is either covered by an authorized plan or not, which decides whether each update is a regulatory event. Evidence quality: FDA's draft guidance describes a complete evidence file, so a company that cannot produce one has a gap whatever its stated accuracy.

In Life Sciences and Healthcare

In life sciences an AI claim usually attaches to a regulated act, such as a diagnosis, a dosing decision, a trial enrollment, or a manufacturing release, which makes the claim testable. FDA's Center for Drug Evaluation and Research reports a significant increase in drug application submissions with AI components across the nonclinical, clinical, postmarketing, and manufacturing phases. Diligence treats AI as a capability that runs through every healthcare sector rather than as a sector of its own.

Governing Authority and Sources

Frequently Asked Questions

Does FDA regulate AI used in drug discovery?

FDA's January 2025 draft guidance does not cover AI used in drug discovery. It addresses AI that produces information or data supporting regulatory decisions on the safety, effectiveness, or quality of drugs, and it also excludes AI used only for operational efficiencies.

How is AI used in biotech, and which rules apply?

AI appears across the drug product life cycle, from nonclinical studies to manufacturing. For drugs and biologics, FDA's January 2025 draft guidance sets a seven-step credibility framework for submissions that rely on AI, and it scales the required evidence to the model's influence and the consequence of the decision.

What is an AI healthcare company?

For diligence, it is a company whose product depends on an AI model operating under a defined regulatory channel: an AI-enabled device software function, a model supporting a drug regulatory decision, or a decision support intervention in certified health information technology. Each channel carries its own rule set and evidence standard.

Pillar: AI as Accelerant. Related entries: Software as a Medical Device, Laboratory-Developed Tests and CLIA, and Growth Equity.