How MadeAi Supports HTACG AI Principles in JCA

HTACG AI Principles in JCA

The HTACG AI Principles in JCA are three rules- accountability, transparency and reporting, that the EU HTA Coordination Group adopted on 15 July 2026 for any use of AI in preparing a Joint Clinical Assessment dossier. MadeAi supports them with a human approval gate at every workflow stage, a timestamped audit trail of every AI action and human edit, and exportable AI disclosure documentation built for the dossier.

The principles do not ban AI. They make the health technology developer answerable for every AI-assisted step, require developers to identify those steps in the dossier, and set a minimum reporting standard for the tools used. For evidence teams, that changes the question from “can we use AI for the JCA” to “can we prove how we used it.” This article explains what the principles require, where they bite in a real dossier workflow, and how each requirement maps to a specific control in MadeAi.

Why the HTA Coordination Group Published AI Principles for JCA Dossiers

These principles address the need for consistent AI disclosure in JCA dossiers, which are legally defined, publicly available documents increasingly prepared with AI assistance. Under Regulation (EU) 2021/2282, the Joint Clinical Assessment has applied to new oncology medicines and advanced therapy medicinal products since 12 January 2025, with orphan medicines following in 2028 and all centrally authorized medicines by 2030. The developer’s dossier is the basis of the assessment; it must be complete with respect to available studies, and the final report is public.

That combination creates pressure in two directions. On one side, the dossier must answer every PI(E)COS question in the scope within roughly 100 days of receiving it, and each PI(E)COS can require its own systematic literature review. The earlier MadeAi article on the JCA anticipation window shows how early assessments produced 7 to 13 PI(E)COS per product. On the other side, assessors need to trust that a human reviewer checked the screening decision or extracted hazard ratio, and assessors can reproduce the result.

The HTACG’s own document states the tension plainly: AI can make parts of the JCA process more efficient, but careless use risks the completeness, methodological quality and rigor of the assessment. The HTACG adopted the General Principles on the Use of Artificial Intelligence in the Preparation of Dossiers for Joint Clinical Assessments on 15 July 2026 under Article 3(7) of the Regulation to address that tension through rules rather than restrictions.

Who Is Responsible for AI Compliance in JCA?

The principles apply to health technology developers (HTDs), meaning the pharmaceutical or medical device company submitting the dossier, and to anyone preparing evidence on their behalf. The scope is deliberately wide. The principles define AI as any model developed through machine learning, including generative models, regardless of architecture. A screening classifier trained in 2019 and a large language model drafting a results narrative are both in scope. Only AI used for spelling and grammar correction is exempt from disclosure.

The principles also sit inside a wider regulatory frame. They cite the EMA reflection paper on AI in the medicinal product lifecycle for the definition of AI, and they require consistency with Regulation (EU) 2024/1689, the EU AI Act. A platform that claims to support the HTACG AI Principles in JCA therefore needs to handle data protection, copyright and model governance, not just produce a disclosure paragraph.

What Are the Three HTACG AI Principles for JCA?

The HTACG document is three pages long and organizes its expectations under accountability, transparency, and reporting. Each principle translates into concrete obligations that an evidence team can audit against.

Three HTACG AI Principles for JCA

Accountability. The developer owns the content, methods and findings of the evidence synthesis, including the decision to use AI, the effect of that use on the synthesis, and the validity of every AI output. Developers must ensure human oversight at every AI-assisted step and keep a person responsible for quality and accuracy, even when using automation in dossier preparation or analysis. AI use does not change the applicability of HTACG methodological guidance, and it must remain possible to verify compliance with those methods. Legal compliance must address copyright, data protection, and the EU AI Act, particularly because the dossier is intended for publication.

Transparency. If developers use AI, they must disclose its use in the dossier and identify every AI-assisted step. The document lists information retrieval, including study screening, data extraction, risk of bias assessment, analysis and reporting as examples. All AI-generated or AI-informed outputs and judgments must be clearly identified and disclosed in the dossier to ensure transparency and accountability.

Reporting. At minimum, the dossier reports the name, version and date of each AI tool, its developer, and the specific purpose it served. Describing any adaptation of a commercial tool is named as good practice. Teams should record all prompts used in AI tools and make them available on request during the assessment.

How MadeAi Builds the HTACG AI Principles in JCA Into the Workflow

MadeAi supports the principles structurally: every AI step in the platform ends at a human decision, every decision is logged, and the log exports in a form the dossier can carry. The platform’s JCA module, MadeAi-JCA, connects anticipatory scoping with the full systematic literature review services workflow, so the same governance applies from comparator identification through to the synthesis narrative.

Workflow Architecture from Scoping to Dossier

MadeAi-JCA: AI-Assisted Evidence Generation Workflow

The JCA workflow in MadeAi runs in four stages, and a human approval gate sits between each of them:

1. Define product. The team enters the indication, population details and intervention. No AI runs yet; this stage fixes the P and I of every PI(E)COS that follows.
2. Generate comparators and outcomes. AI scans tiered sources, including published EU HTA reports, clinical guidelines, product labels, trial registries and the literature, to propose comparators and outcomes for each of the 27 Member States. Each proposal carries a source tier, a rationale and a link to the exact source document. Reviewers accept, edit or add entries before anything moves forward.
3. Generate PI(E)COS sets. The platform combines approved comparators and outcomes with the fixed population and intervention to produce the PI(E)COS sets assessors are likely to issue, with Member State applicability shown for each. The team approves the set it will prepare evidence for.
4. Run the literature review per PI(E)COS. Each approved PI(E)COS becomes a literature review project with protocol and search strategy generation, multi-database search, AI screening, risk of bias assessment, data extraction and synthesis. The team tracks overlapping articles across PI(E)COS sets to prevent duplicate screening and ensure consistent results for the same study.

The design choice that matters for the principles is that AI output never becomes dossier content by default. It becomes a suggestion that a named reviewer accepts, changes or rejects, and the platform records which of those three happened.

Mapping Each Principle to a MadeAi Control

The audit trail is the mechanism underneath most rows. Every AI action, human edit and approval is timestamped and attributed to a user. That record is what makes the disclosure export more than a template: the export is generated from what happened, not written from memory after the fact.

Traceability as the Common Thread

Transparency at the output level depends on traceability at the data level. MadeAi extracts data from text, tables and figures, including forest plots and Kaplan-Meier curves, and links every extracted value to its location in the source document. The synthesis narrative is built the same way, with each sentence traceable to the study it draws on. When an assessor asks why a particular hazard ratio appears in an evidence table, the answer is a click, not a search through email.

Key Features and Benefits for JCA Evidence Teams

The features that satisfy the HTACG principles are the same features that reduce dossier risk, which is why governance and productivity are not in tension here. The table below pairs each feature with the benefit a JCA team sees and the principle it serves.

For market access solutions teams, the practical result is that a single prepared evidence base serves the JCA dossier and the national submissions that follow it, since Member States must give due consideration to the JCA report and annex the developer’s dossier to national HTA documentation. Evidence that was traceable once stays traceable everywhere it is reused.

Where the Principles Apply Across the JCA Timeline

The principles apply wherever AI touches evidence that ends up in the dossier, which in practice means three distinct periods with different governance demands.

Anticipation window, before the scope arrives. Teams use AI to map comparators and standard of care across Member States and to generate likely PI(E)COS sets. The accountability requirement already applies: a reviewer must approve each comparator and outcome, and the source behind it must be retained, because this scoping work feeds the literature reviews that will be submitted. MadeAi logs these approvals from the first stage, so there is no gap between pre-scope and in-scope governance.

The 100-day submission period. This is where most AI-assisted screening and extraction happens, and where documentation is most likely to be skipped under time pressure. A platform that records AI actions and reviewer responses automatically removes that failure mode. The disclosure export is a by-product of the work rather than a task added at the end.

Assessment and national reuse. During the assessment, the HTACG can request prompts and clarification on AI-assisted steps. Afterward, developers include the same dossier as an annex to national HTA submissions. Evidence that carries its provenance travels well; evidence that depends on a departed contractor’s spreadsheet does not. The same governance logic applies to US submissions, which the MadeAi article on evidence generation for AMCP dossiers covers in detail.

A typical use case is an oncology asset entering JCA with 10 or more PI(E)COS, several requiring indirect comparisons. Each PI(E)COS runs as a separate literature review project in MadeAi with shared article tracking, and each produces its own audit trail and PRISMA diagram. The dossier’s AI use statement then draws on one consolidated export rather than ten separately maintained logs.

Best Practices for Using AI in a JCA Dossier

The principles reward teams that treat AI governance as a workflow property rather than a disclosure exercise. Six practices follow directly from the HTACG text and from how MadeAi is used in practice.

1. Decide AI use at protocol stage. Write into the review protocol which steps will use AI, who approves each, and what counts as an AI-informed judgment. The HTACG makes the developer accountable for the decision to use AI, so the decision should be a documented one.
2. Keep humans at every gate, not just the last one. A single final review does not satisfy “no step fully automated.” Screening, extraction, risk of bias and synthesis each need a recorded human action.
3. Record the tool stack before work begins. Capture tool name, version, date and developer at project setup. Versions change mid-project; the log should show which version produced which output.
4. Retain prompts as project artifacts. The HTACG expects prompts to be available on request. Prompts stored in individual chat histories are not retrievable; prompts stored with the project are.
5. Validate AI performance in the therapeutic area. Published screening and extraction accuracy figures vary by domain. Run a calibration set against human reference decisions before relying on AI prioritization for a new indication.
6. Generate disclosure from the audit trail, not from memory. If the platform can export the record of what happened, use that export as the basis of the AI use statement and have the methods lead review it.

Pharma evidence generation solutions that embed these practices in the tool remove most of the manual compliance work, but none of them remove the need for the methods lead to own the result. The HTACG is explicit that accountability stays with the developer regardless of the vendor.

Risks and Limitations 

The principles will evolve. The HTACG may introduce further guidance on AI-informed judgments and disclosure requirements. Teams should maintain detailed logs to prepare for future updates.

AI accuracy varies by domain. Therapeutic areas, study designs, and complex figures affect AI performance. Human reviewers must verify extracted data to ensure accuracy.

Governance does not replace expert judgment. Audit trails confirm approvals, not accuracy. Experts must still validate methods, search strategies, and evidence interpretation.

Transparency requires balance. Teams must disclose AI use while protecting copyrighted and commercially sensitive information. They can provide detailed prompt logs upon request rather than publish them in the dossier.

Future Trends in AI Governance for HTA Submissions

The HTACG principles mark the first EU-level statement on AI use in HTA dossiers. As assessment volumes grow, the HTACG will likely introduce more specific guidance and aligned reporting standards.

The HTACG estimates it will initiate around 50 JCAs in 2026, according to its 2026 work programme as summarized by Cencora, twice the 2025 target, with medical devices entering the process for the first time. Volume at that level makes manual governance impractical and makes platform-level logging the default.

Reporting standards are converging. ISPOR’s ELEVATE-GenAI guideline addresses reporting of generative AI in HEOR, and a standards-based governance framework presented at ISPOR Europe 2026 maps human-in-the-lead control, audit trail, reproducibility and provenance onto PRISMA 2020, PRISMA-S and the HTACG dossier guidance. Expect dossier templates to acquire a structured AI use section that pulls from exactly these fields.

National agencies are moving in the same direction. NICE has published a position statement on AI in evidence generation, and Canada’s CDA-AMC has issued its own expectations, so a disclosure approach that satisfies the HTACG will increasingly serve multiple jurisdictions. For AI in Life Science platforms, the design target is a single governance model that produces jurisdiction-specific exports, rather than separate compliance work per agency.

Conclusion

The HTACG AI Principles in JCA turn responsible AI use from an aspiration into a dossier requirement with three testable parts: a person accountable for every step, every step named, and every tool reported. MadeAi meets those parts by design rather than by documentation effort. The approval gate makes accountability a workflow event, the audit trail makes transparency a record, and the disclosure export makes reporting a by-product of the review itself.

For teams preparing a JCA, the useful next step is to run the three questions from the summary table against the current process: who approved each decision, which outputs were AI-informed, and whether the prompt behind a given number can be produced on request. Where the answer is unclear, the gap is a systems gap. An AI platform for Life Sciences that logs decisions as they are made closes it without adding work to the 100-day window.

To see how the MadeAi-JCA workflow handles scoping, screening and disclosure for a specific indication, request a walkthrough with the MadeAi team, or read the HTACG principles directly and compare them against your own protocol.

Author’s Note: This article was supported by AI-based research and writing, with Claude 5.1 assisting in the creation of text and images.