Part 1 of 2: What AI-enabled MLR means for the medical signatory
By Shazad Jafri, ABPI Medical Signatory and Founder, SignAffairs, and Vivanti Europe
This summer, two developments arrived within a month of each other. They deserve to be read together.
In June, Veeva acquired Copli and launched what it calls agentic MLR, with the potential, it says, to eliminate 70% or more of manual MLR labour within five years.¹ Moderna has described its ambition as an almost touch-free review process.²
Then, on 24 July, the Italian medicines agency AIFA published guidance on AI in promotional material, including a declaration concerning algorithm validation and oversight of AI-generated outputs.³
One direction takes the human touch out of review. The other asks a named professional to stand behind the machine.
As a Medical Signatory, I find that tension more interesting than whether AI can check references faster. With discussions on the next ABPI Code expected during 2026 and 2027, it deserves attention now.⁴
Two words: final form
Clause 8.1 of the ABPI Code requires promotional material to be certified in its final form, with no subsequent amendments, by a qualified person acting on behalf of the company.⁵ The PMCPA confirms that certification requirements cannot be fulfilled solely through AI.⁶
Dynamic content is not new. The Code recognises that every possible combination of approved components need not be certified separately.⁵
Generative AI raises a different question.
Assembling approved content under controlled conditions is one thing. Allowing a model to rewrite an approved sentence or generate a response to a healthcare professional in real time is another.
If the words an HCP reads did not exist when the material was certified, what exactly was the final form?
Existing PMCPA decisions illustrate why this matters. In one case, a promotional website went live with a visual element that differed from its certified version.⁷ In another, material treated as non-promotional was subsequently determined by the Panel to be promotional.⁸
Neither case involved AI. Both demonstrate that what reaches the audience matters, and that a well-executed process cannot compensate for an incorrect assessment of the material.
What stays human?
I would be comfortable with AI supporting a great deal of bounded checking: prescribing information, black triangles, references, approved claims and version consistency.
AI can also assess more complex issues when given appropriate context. The question is how far we can rely on those assessments.
Is the material promotional in the circumstances? What overall impression does it create? Is a comparison fair? Could individually approved claims, presented differently, create a message nobody approved?
AI may identify these issues. Having the relevant context does not, however, prove it will do so reliably.
Clause 8.5 still requires the signatory to certify, among other things, that material complies with the Code and presents the facts fairly and truthfully.⁵
Capability, validated reliability and professional accountability are different things.
When claims libraries become the foundation for AI
Earlier in my career, I helped establish standards for local signatories across 56 AstraZeneca affiliates and implement core claims documents to improve consistency.
Back then, claims libraries were largely about efficiency. Today, they are the foundation on which many AI-enabled content workflows operate.
If the library is incomplete or out of date, AI does not fix that. It scales it.
Medical Affairs and signatories therefore need a voice in defining the content controls around AI, not simply reviewing what emerges at the end.
When the signatory is asked to vouch for the algorithm
AIFA’s guidance is more prescriptive than current PMCPA AI guidance. It addresses approved sources, traceability, verification of outputs, validation, stress testing and human oversight.³
Much of this is sensible. Source control is not, however, the same as algorithm validation.
A system can use approved sources and still omit a qualification, misinterpret evidence or create a misleading conclusion.
AIFA also calls for a declaration concerning algorithm validation and supervision of outputs, supported by technical documentation.³
This makes me uneasy.
I am a pharmacist, not a machine-learning engineer. If I am asked to sign such a declaration, what evidence should I expect to examine?
Technical validation belongs with appropriately qualified specialists. A responsible professional should nevertheless understand what has been validated, for which purposes and with what limitations.
Without that evidence, the danger is a comfort signature.
Italy’s filing system differs from UK certification, and its Responsible Person for the Scientific Service is not an ABPI signatory. The underlying question remains relevant: what can a named professional reasonably attest to?
Three layers of assurance
Working with Shana Owen and the team at Vivanti Europe, I have been exploring a practical framework for AI-enabled MLR. It separates three responsibilities.
System assurance: Can we trust the system?
Has the workflow been validated for its intended use? Are its limitations understood? Technical, Quality and Security specialists need to establish this evidence.
Content assurance: Can we trust the content?
Are the sources and claims approved, current and appropriately controlled? Can we distinguish approved wording from newly generated content? Medical, Regulatory, Legal and Compliance expertise is essential here.
Item accountability: Can I approve this material?
What has been produced, what checks were completed and what remains unresolved? Where Clause 8 applies, certification remains a named human decision.
This is our proposed framework, not PMCPA policy.
System assurance supports certification. It does not replace it.
Five questions follow naturally:
1.What sources can the system use, and who controls them?
2.What may it generate, and what must remain approved wording?
3.What happens when it is uncertain?
4.Which changes trigger another review?
5.If a complaint arrived next year, could we reconstruct how the item was produced and checked?
The next ABPI Code
Medical Signatories have never simply been final checkers. The difficult decisions have always required scientific knowledge, Code understanding and professional judgement.
AI may move more of the supporting work upstream, into source governance, validation and workflow design. That makes signatory involvement in these controls increasingly important.
What should final form mean when content can be generated at the point of use? What evidence justifies reliance on an AI-enabled workflow?
I do not know exactly what the next ABPI Code should say. I would rather signatories helped shape that discussion than read about its outcome afterwards.
The opportunity is not simply to automate more review. It is to make the systems supporting professional judgement worthy of our confidence.
This is Part 1 of a two-part series from SignAffairs and Vivanti Europe. In Part 2, Vivanti examines how system assurance, content controls, validation, traceability and human handover can work in practice.
Read Part 2: coming soon
References
Veeva Systems. Veeva Acquires Copli, Launches Veeva Falcon MLR. June 2026.
Accelerating the content lifecycle: A 7-step guide for AI in MLR. 2026.
Linee di indirizzo: utilizzo di sistemi di Intelligenza Artificiale nei materiali promozionali. 24 July 2026.
Public consultation on proposed amendments to its Constitution and Procedure. 2026.
2024 ABPI Code of Practice, Clause 8.
Guidance on AI and promotional review. Revised June 2025.
CASE/1029/07/26. Completed August 2026.
CASE/0599/05/25. Completed February 2026.
