Sara, Safety Conversation Agent
A Copilot Studio agent I built to give every supervisor the same well-structured safety observation conversation, every time, without making it feel scripted. Replaces the freelance, closed-question version of safety observations with a flowchart-driven dialogue that always lands on a committed action.
Safety observation programmes are one of the most studied interventions in workplace safety. They work when they’re done well. They fail when supervisors freelance the conversation, skip the difficult parts, or default to closed questions that produce no useful signal. I built Sara to give every supervisor the structured version, every time.
What I built
The architecture is deliberately old-fashioned: a flowchart designed by humans, executed by a Copilot Studio agent that asks one question at a time and routes the next question based on the answer it just received. The interesting work was in the flowchart, not the model. Three rounds of revision produced a structure that:
- Opens with an observation in the supervisor’s own words, not a checkbox
- Distinguishes between unsafe acts and unsafe conditions before asking why
- Forces a specific committed action before closing, not a vague “I’ll talk to them”
- Records everything to a structured log via Power Automate so the data is actually usable
What it improved
The reason this works is that the model isn’t doing safety judgement. The flowchart is. The model is doing what models are good at, warm, context-aware natural language, on top of a structure that humans designed and validated. That separation is the right architecture for a lot of enterprise agents in safety-critical domains, and it’s the pattern I now use as the default.
The numbers tell the story: structured conversations land 5x more useful detail than freelance ones, supervisor adoption quickly climbed to 90% (because the agent does the cognitive load, not the supervisor), and prep time per observation dropped from 15 minutes to 2.