On July 15, 2026, the HTA Coordination Group (HTA CG) adopted the first General Principles on the Use of Artificial Intelligence in the Preparation of Dossiers for Joint Clinical Assessments (JCAs). Although the guidance is only a few pages long, it represents an important milestone for the industry. Rather than debating whether AI can be used in HTA, the document acknowledges that it already has a role to play — and instead focuses on how it should be used responsibly.

AI is welcome, but accountability remains human

The strongest message throughout the guidance is that AI does not change accountability. Health Technology Developers remain fully responsible for the evidence, methods, and conclusions presented in a JCA dossier, regardless of where AI has been used in the process. The document also reminds developers that AI use must comply with applicable legislation, including copyright, data protection, and the EU AI Act (1).

This is an important distinction. AI can support evidence synthesis, but it cannot replace scientific responsibility.

Human oversight is non-negotiable

The guidance also makes it clear that no stage of dossier preparation or analysis should be fully automated without a human ultimately responsible for the quality and accuracy of the outputs. Furthermore, any AI-assisted process must still allow reviewers to verify compliance with established HTA methodologies.

This reinforces what many HTA professionals have already experienced in practice: AI is most valuable when it augments experts, allowing them to spend less time on repetitive tasks and more time applying scientific judgment.

Transparency will become standard practice

Another key principle is transparency. Organizations are expected to document every meaningful use of AI during dossier preparation, whether for literature retrieval, study screening, data extraction, risk-of-bias assessment, analysis, or reporting. The guidance also recommends recording the AI system used, its version, developer, intended purpose, and any adaptations made to commercial tools. Even prompts should be retained and made available if requested during the JCA process.

These expectations move AI governance closer to the same level of documentation already expected for statistical methods and other regulated processes.

A direction we have already been following at Cytel

The principles published by the HTA CG closely align with the approach we have been developing at Cytel (2). Earlier this year at ISPOR 2026, we presented an AI-assisted framework for multilingual HTA landscaping to support EU JCA PICO simulations (3). Our solution combines retrieval-augmented generation (RAG), structured prompts, and expert review to extract information from historical HTA reports across multiple European jurisdictions.

One of the most important findings from our work was that human validation remains essential. While prompt refinement substantially improved extraction accuracy and reduced hallucinations, expert review was still required to verify completeness, interpret nuanced HTA decisions, and ensure the reliability of the final outputs. In many ways, this reflects the principles now formalized in the new guidance.

Looking ahead

The publication of these principles should be seen as an enabler rather than a constraint. They provide much-needed clarity for organizations investing in AI for evidence generation and JCA readiness. The future of AI in HTA is unlikely to be fully autonomous. Instead, it will be built on transparent, validated, and human-supervised systems that improve efficiency while preserving the scientific rigor expected by European HTA bodies.

1. European Union. Regulation (EU) 2024/1689 (Artificial Intelligence Act). Official Journal of the European Union, 2024.

2. Sarri, Grammati, and Ipek Ozer Stillman. “What “Human-in-the-Loop” Means in the Artificial Intelligence Era of Health Economics and Outcomes Research.” Value in Health.

3. Cossio, Manuel, and Lilia Leisle. “HTA27 ACCELERATING DYNAMIC HTA LANDSCAPING IN ONCOLOGY THROUGH AUTONOMOUS GENERATIVE AI-DRIVEN MULTILINGUAL DATA EXTRACTION.” Value in Health 29.6 (2026): S271.

Share Via