← Back
Pharma·AI

Where AI fits in a clinical trial, and where it doesn’t

Most conversations about AI in pharma start with the same worries: data leaking, decisions nobody can explain, an auditor’s question with no good answer. These are real concerns. Each one is met by work pharma already does, such as validation, risk assessment, traceability and keeping evidence on file.

Compliance means weighing a risk, justifying the decision and writing it down. So rather than ask whether AI is allowed, ask which task, carrying which risk. The answer changes a great deal from one task to the next.

Buy the regulated core

For a small biotech heading for its first approval, I’d recommend buying an eTMF, EDC, CTMS or IVRS rather than building one. I have implemented these systems and assessed them as a buyer. A good vendor gives you confidence that this part is handled while you get a drug to market, and that is worth paying for. A 50-person biotech has better uses for its people than running its own clinical data systems.

The rest of this piece is about the work around those systems.

Bounded, not black-box

Inside a tool like Claude Code the work is pigeonholed: scoped to one task, reviewable and logged as it goes. There is a defined scope, a record of what changed and a person who approves the result. That is a controlled change, the same process pharma uses for any system change.

Agents and plug-ins that record the evidence

Point the AI at the documentation as well as the code. Agents and plug-ins can write the logs, records and evidence a regulated system needs as the work happens. That covers what changed, why and against which requirement. The audit trail then comes out of the job itself, instead of being assembled afterwards from memory and half-finished tickets. epiClavis ↗ is built to do this. It works, but it has not yet been used on a live audited system.

The work around the regulated core

Most trials run on a layer of work that takes up people’s weeks without touching regulated data. Someone tracks when monitoring visits happened and when follow-ups are due. Someone checks whether documents get their quality control (QC) on time. Someone spots the outlier in a pile of eCOA output so it can be investigated properly. Then there are reports and issue logs.

I have built tools for this before, in Excel and VBA, because that was what I had. One spreadsheet turned raw eCOA data into graphs, so we could see which sites needed a closer look. It didn’t analyse anything or replace the investigation; it showed us where to start one. Another tracked monitoring visits. A third turned document QC timelines into metrics, so the team could see whether QC was on time.

That is the work I would use AI for now. Its risk is far lower than the regulated core’s. EDC data is a serious risk and should be treated as one; a monitoring visit tracker is not. You could paste much of this into Claude today and get something useful. A purpose-built tool makes it repeatable, and gives you something to hand to the next person.

Work in progress

Parts of epiClavis will be free to use. The regulatory parts are where the value is, and they are the hardest to build. If someone on your team maintains a critical spreadsheet that nobody else understands, I’d like to hear about it.

I would start by asking what it touches. That answer decides everything else.