How to Handle HTACG General Principles on AI in JCA Dossiers

On 15 July 2026, the Member State Coordination Group on Health Technology Assessment (HTACG) adopted its general principles for the use of AI in preparing dossiers for Joint Clinical Assessment. It runs three pages. Read it here:

General principles on the use of Artificial Intelligence in the preparation of dossiers for Joint Clinical Assessments

If you read our previous piece on the NICE position statement, you may recognize the posture. The HTACG is not expressly forbidding the use of AI. It says plainly that AI could make parts of the JCA process more efficient, then adds that careless use risks the completeness and methodological rigour of the assessment itself. It also emphasizes accountable and transparent use, backed by rigorous reporting. Coincidentally, these principals are where NICE landed, as well as representing core tenants for us, ever since we started shipping what we call “Smart” features powered by large language models (LLMs).

Accountability

One facet of accountability is quite clear: the health technology developer owns the dossier, as well as the SLR that serves as the base of evidence for the submission. Now, according to this new guidance, the dossier must cover whether AI was used at all, how, what it did to the evidence synthesis, and whether the resulting output holds up. It also covers copyright and data protection, since the dossier gets published, plus consistency with the EU AI Act. Human oversight has to run through every AI-assisted step. This recent guidance is most blunt here: no part of dossier preparation or analysis should be fully automated without a person ultimately answerable for quality and accuracy. Existing HTACG methodological guidance does not bend to accommodate AI, and it must be possible to verify that the methods were actually followed. 

What you need to do:

• Human-in-the-loop. Remember, you are accountable for every aspect of the “evidence synthesis presented.” Ensure a human expert double-checks every decision along the way, as well as the final output. In fact, the guidance clearly states that “none of the steps in the dossier preparation or analysis process should be fully automated without a human being ultimately responsible for the quality and accuracy.”
• Document, document, document. Because dossiers are published, it’s equally important that you have covered your copyright, data protection, and if you are using AI, your AI Act (Regulation (EU) 2024/1689) obligations before drafting the dossier.
• Adherence to best practice. Health Technology Developer’s (HTD’s) dossier specifies methodology; if AI is used, the guidance states that “it should also be possible to verify the compliance with the methods.” In the context of the literature review, this means maintaining an audit log of every decision made at each stage in your review, as well as the ability to trace

Transparency

If AI was used in the preparation of the dossier, you should document it and name the steps. The guidance lists information retrieval and screening, data extraction, risk of bias assessment, analysis, and reporting. Anything AI-generated or AI-informed gets flagged as such. Spellcheck and grammar tools are exempt, however. 

What you need to do:

• Disclose every AI-assisted step by name. For example, the guidance lists information retrieval (incl. screening of studies), data extraction, risk of bias assessment, analysis, reporting as common examples.
• Flag each AI-generated or AI-informed output or judgment in the dossier and document what you did and how you e

Reporting

At minimum, report the tool name, its version and date, who built it, and what you used it for. Describing any adaptations you made to a commercial tool is listed as good practice rather than a requirement, though it is explicitly expected that all prompts used to instruct an LLM be documented and made available.

What you need to do:

• Record, for each tool: name, version, date, developer, and the purpose you used it for.
• Save your prompts and be ready to produce them if asked during the JCA.
• As good practice, note any customizations you made to an off-the-shelf tool.

Alignment For Nested Knowledge Users

This new guidance from HTACG on the usage of AI is best understood when situated in the context of the broader methodological requirements outlined by the dossier template guidance – Medicinal products (V1.0, adopted 28 Nov 2024). While the template and overarching methodology does not change with this most recent guidance, it does specify the disclosure requirements if you are using AI. Here is list of how Smart features list aligns with these steps in the following order:

• Information retrieval: Smart Search. §4.2.1 requires at least MEDLINE and Cochrane CENTRAL plus registries; Smart Search covers MEDLINE (via PubMed) and ClinicalTrials.gov natively. To fully meet this requirement, you need to query CENTRAL and Embase by RIS import.
• Screening: Smart Screener. §4.2.2/§5.1.2 require a documented selection process and PRISMA chart; PRISMA ; charts are created automatically in Nested Knowledge, and two-pass screening produces the counts and logs every decision which can be viewed in Study Inspector.
• Data extraction: Adaptive Smart Tags. §5.2 requires study-characteristics tables; you accept or reject each recommendation, and that step is your audit trail.
• Risk of bias: Smart Critical Appraisal. §5.3.1.2 requires you supply the signalling-question information and source passages. Smart critical appraisal supports both ROB 2 and Robins-I, and attaches passage-level annotations to each response. Dual modes are available.
• Analysis: Adaptive Smart Tags, Insights, and Synthesis Chat. §4.3 and the synthesis guideline govern meta-analysis and indirect comparisons; we recommend using Index Table Tags if you need to do an ITC feasibility analysis.
• Reporting: Smart Abstract and Synthesis Chat. §1.3 calls for a per-PICO executive summary. Your prompts can be exported as well, to align with the reporting requirements for AI: “As well as it is expected that all prompts used in AI tools are recorded and made available if requested during the joint clinical assessment process.”