Human-Agent Operations Playbook

Start with the role map, before any control layer or training. Split each affected process's work three ways, agent-run, human-run, shared, and name the roles whose jobs change. Everything else is guesswork without it: you cannot staff escalation paths, size oversight, or build training paths until you know which work moved and who is affected. Then make staffed, named escalation coverage a launch condition for every agent.

Convert their process knowledge, which is the asset, into oversight capability: reading agent outputs, judging exceptions against policy, spotting degradation early, escalating with context. Define a written skill profile per oversight role, build a training path with practice cases and checkpoints, and hold a hard certification gate before anyone owns a queue alone. Recruit incumbents first; hiring oversight from outside while the people who know the process go untrained is the standard failure.

Watch the seam, because mixed operations degrade there first, long before quarterly numbers move. The two instruments that matter: how often handed-off exceptions bounce back or need rework, and how long decisions take from handoff to resolution, trended by process. Add exception volumes and patterns reviewed on a fixed cadence, in quiet weeks too. If those signals are not owned by someone who can say whether the seam got better or worse this month, degradation is invisible.

Retrain the incumbents first. The people who executed the process before agents did are usually the best judges of its exceptions, which is the core of oversight work. A defined path from execution to oversight, with certification and visible career routes beyond, converts that experience and answers the workforce's real question about agents. Hiring outside for oversight while incumbents go untrained loses the process knowledge and the people at the same time.