Agentic AI: where to start
Our checklist before putting an AI agent in production, especially in a healthcare context.
AI agents (models that use tools, plan and act) went from demo to infrastructure in two years. Before putting one in production, especially in healthcare, we ask ourselves three questions.
1. What happens when it is wrong?
An agent will make mistakes. The question is not if but what the error costs and who catches it. In healthcare, this mandates human validation loops on anything touching patients or regulated data.
2. Can we replay what it did?
Full traceability: inputs, tools called, outputs. This is a classic quality requirement, and healthcare standards prepare us well for it.
3. Is an agent the right solution?
Many problems are solved by a deterministic script. An agent is justified when the space of situations is too large to enumerate. The rest of the time, a classic pipeline is simpler, cheaper and safer.
In upcoming posts we will detail our agentic stack and concrete lessons learned.