The problem
Coaching skill is one of those things everyone agrees matters and almost nobody gets real practice at. Managers read about how to handle a defensive, deflecting, or checked-out direct report, but the first time they actually try it is with a real person and a real relationship on the line. There’s no equivalent of a flight simulator for a hard coaching conversation, no safe place to get it wrong first.
What I built
I built a coaching practice agent that plays the role of a direct report in a realistic, difficult workplace scenario while the manager practices coaching them. Before a session starts, the agent asks a short set of diagnostic questions: what skill the manager wants to practice (pushing past deflection, building accountability, listening without jumping to solve), what kind of conversation it should resemble (a performance struggle, a motivation issue, a behavior problem), and how resistant they want the fictional employee to be. It then proposes a specific scenario and fictional direct-report profile before starting.
During the practice conversation, the agent stays fully in character, deflecting and pushing back the way a real employee would, and never breaks to coach from inside the roleplay. At two or three key moments, chosen for how much they actually matter rather than after every exchange, it explicitly steps out of character to give direct feedback: naming the specific moment, connecting it to a coaching principle, offering a concrete alternative the manager could have tried, and inviting them to run the exchange again. Every session ends with a short summary: one thing that went well, one thing to keep working on, and a suggested scenario for next time.
It’s scoped deliberately as a practice tool, not a substitute for HR. It never simulates a real employee by name and won’t touch a real performance or employee-relations situation. If a manager describes something that sounds like an actual issue, it redirects them to their HRBP instead of trying to help.
What happened
This is the tool the HR Buildathon produced: built in that session and running on Microsoft Copilot Studio, the same platform behind the assessment agent. It’s shipped, live for managers to use whenever they want to rehearse a coaching conversation before having it for real.
What I’d do differently
I’d expand its scope beyond coaching specifically. The same practice-and-feedback structure, rehearsing a hard conversation against a realistic, resistant AI counterpart, applies just as well to the other difficult conversations managers face, not only coaching moments. Right now it’s built for one type of conversation when it could reasonably handle several.