By Eugen Stoicov, Senior Product Designer, Vienna. Published 2025-11-12.
In life sciences, the AI that wins is the one a team can explain to an auditor. Not the one that demos best.
Regulated software does not forgive magic. When an AI agent proposes a change to a requirements trace or a risk file, someone will eventually ask: what did it do, why, and can we show that in an audit?
That question should shape the interface long before anyone argues about the model. The job is not to hide the machinery, and not to dump it on the desk either. Someone came here to finish a release, not to audit your architecture.
Teams under compliance pressure already live in dense systems. A wall of confidence scores and token traces does not read as transparency to them. It reads as more work. Start with the outcome they came for. Have the trail ready for the moment they need to verify, challenge or document, which will come.
I treat agent runs like a good diff: a summary you can act on, steps you can open, and a straight line back to the source. And when the AI has nothing to point at, the interface should say so plainly. A green checkmark over missing evidence is not a design shortcut, it is a false statement.
Chat is a familiar metaphor and a dangerous one. Compliance work has shape: requirements, tests, risks, releases. A blank text box has none. Forcing structured work through a conversational void asks QA and engineering leads to hold the structure in their heads while the software pretends it does not exist.
Better to keep the agent inside the objects people already manage. Let it suggest, annotate, propose. Then leave the human an explicit accept, revise or reject, and write down which one they chose. That record is the product.
In this domain, polish laid over confusion is a liability with rounded corners. The interaction worth having is the one that lets someone make a hard call and defend it six months later. That is the bar, and it runs through everything else on this site.
Topics: AI, Compliance, Product design.