Validated AI for Pharma Summit: Six Ideas the Sessions Kept Returning To

Reading time:

time

min

November 4, 2026

On 24 September we hosted the Validated AI for Pharma Summit: 17 speakers, 11 sessions and six hours on one question. What does it take to run AI in validated pharma environments? The day before, three hands-on workshops gave practitioners time to work through the same problems in smaller groups.

Pharma has plenty of AI pilots and very few systems running in validated production. Each session approached that gap from a different direction, and the same six ideas surfaced again and again. This post summarises them. Every session is available on demand on RingCentral, and we are publishing two recordings a week on YouTube.

1. Validate what the AI produces

Chris Decker of CDISC opened the day by describing AI as the engine and data as the fuel. Standards are what turn a model's output into evidence. Traceability, he argued, has to be designed into the data from the start, because documentation written afterwards cannot supply it. Without it, an AI result is “a persuasive guess wearing a lab coat.” The audience agreed: in our live poll, data quality came out as the top blocker to validated AI.

Peter Krusche of Novartis gave the structural argument. GxP rests on two things, a qualified and accountable human and a deterministic, validated system. AI sits between them and is neither, so the validation effort belongs on what it produces. He also reminded the room that people are non-deterministic as well, which is the reason double programming exists.

Ross Sharp of Domino Data Lab turned this into a working practice he calls the freeze point. Teams record which steps in a workflow are deterministic and which are not, add a human sign-off directly after each non-deterministic step, and allow no AI to change an output once a human has signed it.

2. Make human oversight reviewable

Krusche observed that a merged pull request with thousands of changed lines says little about whether anyone was really in the loop. Line-by-line review is the wrong altitude. The stronger check compares an implementation against its specification, and pharma already has machine-readable specifications to use: USDM for study design, data contracts and controlled terminology.

Eric Nantz and Will Landau of Eli Lilly reached the same conclusion through design. In their Trusted Mini-Agents pattern, every input to a trusted tool gets human oversight, and that oversight is made easy on purpose. A reviewer approves an assumptions table, or checks a map where they would otherwise read a list of coordinates. Ben O'Brien of ERA Sciences described oversight as a spectrum of activities and warned that placing a human in the loop does not, on its own, amount to oversight.

3. Enforce least privilege in the infrastructure

Krusche described agents that run in their own sandbox with fewer permissions than the people around them. Sharp described platform-enforced sandboxing in which sensitive data is never mounted at all. Lilly's rule goes a step further: trusted tools produce every result directly, and the model never relays them.

Several speakers pointed to a single LLM gateway as the place to manage routing, PII guardrails, cost and access control. Pascal Bouquet raised it in his session with Magalie Wasem Tréguer on where AI should run in Europe. Sharp raised it too, and so did Marcin Dubel, Appsilon's Head of Technology, whose talk showed how to get AI into production in validated environments with Mediforce, our open-source platform for human-supervised workflows.

4. Give agents standards and context

André Veríssimo of Appsilon applied Decker's point about standards to codebases. An agent working in a large modular framework only ever sees part of it, so it tends to fix symptoms locally and leave problems for later. A new developer behaves much the same way in their first few months. His answer is to give agents a map of the project: skills that encode its conventions, specialised agents for different roles, and documentation that matches the exact package versions the agent works with.

5. Act while regulation is still forming

Gabriel Eichler and David Foster explained how eClinical Solutions built its EU AI Act compliance framework in four to six weeks. None of the company's use cases fell under the high-risk categories in Annex III, and every workflow kept a human in it by design. Their main lesson was that a short, clear policy serves a team better than a binder of SOPs.

Ben O'Brien focused on what organisations can do before the rules settle. His answer was to demonstrate maturity through shared AI literacy, a clear policy on where non-GxP use crosses into GxP, a standard lifecycle, and an inventory of every AI system in use. In his experience, that inventory is often the first thing an inspector asks for. A lack of formal guidance, he said, is no excuse for inaction. It is also worth remembering that the AI literacy duty in Article 4 of the EU AI Act has applied to all providers and deployers since February 2025, whatever the risk level of their systems.

6. Plan for the people side

Michael Rimler of PHUSE, speaking on AI in pharma roles, polled the audience live. Most attendees said they had no formal way to measure the return on their AI work and confidence in existing guardrails sat below the midpoint of the scale.

In the change management roundtable, Sarah Banks of THRIVE argued that deploying a tool and transforming how people work are different things, and that the last mile is behaviour change. She urged teams to measure the new value AI creates instead of counting adoption. Srinivas Veeragoni described Bayer reviewing every AI project every ninety days. His closing advice was to stop chasing the hype and apply AI, together with your own expertise, to the problems that matter most in your area.

Questions still open

Four questions came up that nobody could answer yet:

Does mandatory human review simply move the bottleneck, turning double programming into triple programming?
What does independence mean when both programmers use the same model against the same specification?
Who may see agent traces in a blinded study, when those traces can reveal unblinded information?
When an AI-assisted decision turns out to be wrong, who owns it, by name? Decker raised this one, and it stayed with the room for the rest of the day.

We expect these to shape much of the practical work in validated AI over the coming year.

Watch the sessions

All 11 sessions and the three workshops are available on demand on RingCentral. If your team is working out how to move AI from pilot to validated production, Mediforce is our open-source platform for human-supervised AI workflows, and our team is happy to talk through your use case.