There is no single AI wizard that can do everything for you. In pharmacovigilance, putting AI agents to work in the real world means choosing the right agent for each task, deciding how much autonomy to give it, and building guardrails that hold up when things do not go according to plan.

In this conversation, Dr. Jan Petracek, our Founder and Chief Executive Officer, sits down with Dr. Henry Seto, our Chief Medical Officer, to explore an analogy Henry uses to describe AI agents: working animals.

Some are slow and methodical. Others are fast and energetic. Some are powerful and need to be handled with care. Once you start thinking about AI agents in these terms, some of the practical questions around how to use them become much easier to understand.

How much freedom should an agent have? What tools should it be allowed to use? What should its workflow look like? And how do you train and retrain it over time?

The answers depend on the job you are asking the agent to do. If your priority is minimising hallucinations, you give the agent a narrow remit, tightly defined tools, and clear boundaries. If you want it to generate something more open-ended and creative, you give it considerably more freedom. Either way, the challenge is not simply building the agent. It is designing the environment in which that agent can operate safely and reliably.

The conversation also gets into what happens when those boundaries fail.

Jan and Henry discuss jailbreaks, why soft guardrails buried in configuration files are not always enough, and why robust governance is essential to getting AI agents into production. The goal is not to assume that an agent will never fail. It is to build systems that can detect, contain, and learn from those failures when they happen.

Henry also gives an example of how, with strict guardrails in place, an agent can surface things you have forgotten. It is a powerful illustration of what these systems can do, but also of an important limitation: you can only recognise whether an answer is right if you have the expertise to evaluate it.

That is ultimately what makes deploying AI in pharmacovigilance different from simply experimenting with AI. The question is not just “Can the agent do the task?” It is “Can we trust the way it does the task, and do we have the oversight to know when it gets something wrong?”

If you work in pharmacovigilance and are being asked to put AI agents to work on real-world tasks, this conversation offers a practical way to think about autonomy, oversight, governance, and where the real value of these systems comes from.

Have a pharmacovigilance question of your own? Put it to our experts.