On 9 November 2026 NICE published its AI methods for evidence framework (ECD20), a 90-page document that replaces its 2024 position statement and its earlier statement of intent. It sets no threshold for good-enough evidence. It governs how any use of AI in producing that evidence must be declared, validated and overseen, and for several tasks it says where AI should not be used on its own.
From position statement to working standard
NICE's August 2024 position statement was the first step. The new framework supersedes it and the earlier statement of intent, and sets specific expectations. An ISPOR Europe 2026 session describes it as a best practice framework to support the use of AI throughout the HTA pipeline. It applies to generative and non-generative methods, and covers primary evidence generation, evidence synthesis, health economic evaluation and technical documents such as dossiers.
The framework has two limits. It excludes health technologies that contain AI, so it is not guidance for assessing AI devices. And it leaves agentic AI for a future update. NICE calls it a living framework.
Uptake so far has been thin. An ISPOR Europe 2025 poster screened 288 NICE technology appraisals and found none that reported AI for data screening or systematic data collection. The framework reads as an attempt to make that safe enough to start.
Proportionate reporting, not a quality floor
NICE states that the framework sets no minimum standard for evidence quality. Reporting is proportionate to the risk and influence of each AI use, and risk rises with automation. The framework separates three levels: support, semi-automation and full automation.
Every use of AI is declared in part 1 of NICE's AI declaration checklist. Part 2 is recommended wherever the use has any material risk or influence on the evidence. NICE also warns that when the motive is efficiency through less human review, the risk profile changes and automation bias and undetected errors become more likely.
Where NICE draws the lines
The use-case positions are where the framework bites:
- Screening can be automated, with conditions. Support use is low risk if humans make every decision. Semi-automation needs an independent adjudicator. Full automation carries substantially higher methodological risk, and false negative rates must be shown to be no higher than under standard human screening, with random sampling of AI-labelled exclusions as ongoing quality assurance.
- Data extraction is not for full automation. NICE does not currently recommend it. Support and semi-automation are acceptable if every AI-extracted item is checked against the primary source.
- Health economic models keep a human close. AI is for support only, for example code that a human verifies and runs, or parameter suggestions that are not substituted into a model automatically.
- Dossiers need named expertise. AI-drafted text must be reviewed by a human with relevant expertise, tied to verbatim quotes or source references, and the submission must state who reviewed it and in what role.
Validation is stricter than many teams will expect. Every validation exercise must use independent datasets, with a pre-specified protocol and, where possible, 95% confidence intervals. Vendor studies can support a case but do not replace an organisation's own evaluation, and re-validation is expected when the model, prompts, datasets, vendor or workflow change. NICE also says humans should not be treated as a perfect reference standard, and that a fixed target such as an 80% true positive rate may not be appropriate.
What it means for market access teams
Whatever tools a team uses, three practical points follow from the text.
- Keep a record from the first day. The declaration checklist, model versions, settings and prompts are easier to assemble as the work happens than after a dossier is written. Stochastic models will not reproduce exactly, so NICE asks for findings that stay consistent across repeated runs, and that needs documented runs.
- Treat vendor claims as a starting point. A screening or extraction tool sold on published accuracy still needs in-house validation on the kind of evidence your submission contains. A tool validated on randomised trials may not hold up on other study designs.
- Decide the human role before the work starts. NICE expects submitters to justify the level of oversight and to name who is accountable. For economic models and dossier text, that means a qualified reviewer at each step, not only at the end.
Two further points. The rule that AI should summarise only documents the user's own organisation created, not third-party documents, affects how teams review competitor or assessment group material. And because the framework sets no quality threshold, committees will still judge the evidence itself. What changes is the audit trail behind it.
To discuss how the framework fits your evidence plan and submission workflow, get in touch.