Schedule backwards from the first day the agent acts, not the day it is switched on. Teach the shared vocabulary a few weeks ahead, run reviewer practice inside the shadow period on the agent's own drafts, certify reviewers before the agent reaches depth 3, and give a short update before every later step up in depth.
Why does timing matter so much?
Train too early and people forget before they need it. Train too late and they learn on live customers. Agent rollouts make this sharper than most system launches, because the agent does not arrive fully formed. It starts by reading, then drafts, then acts within bounds, and each stage asks something different of the people around it.
The shadow period is the key. While the agent runs at depth 1 or 2, every draft passes through a person and nothing irreversible happens. That is the safest possible classroom: real cases, real volume, no harm if the reviewer gets one wrong. A schedule that leaves the shadow period empty of training wastes it.
What does a good schedule look like?
| Phase | What the agent does | What people do |
|---|---|---|
| Six to eight weeks before | Being built and certified | Everyone learns the vocabulary: depth, verdicts, provenance, lapse. A readiness check shows the gaps |
| Four to six weeks before | Certification on the business's own cases | Reviewers take the reviewer modules; builders read the certification results with the process owner |
| Shadow period, two to four weeks | Depth 1 or 2: reads and drafts, a person decides everything | Reviewers decide its drafts as their practicum; weekly calibration on disagreements |
| Go-live | Depth 2 or 3, as certified | Only certified reviewers approve held actions |
| Each step up in depth | The agent moves up, on fresh evidence | A one-hour update for the people whose queue changes, before the change |
| First 30 days after | Live | Weekly calibration, then a retrospective |
The detailed planning steps are in how to schedule AI agent training around a go-live.
What belongs inside the shadow period?
Three things. First, the practicum itself: reviewers decide real held drafts and a lead checks their decisions. Second, calibration: once a week, everyone decides the same handful of cases and the team discusses where they split. Third, a drill: walk through what happens if the agent's certificate lapses or a run clears no depth and it drops to depth 1 that day. Teams that have rehearsed that moment do not panic when it comes.
Keep the formal exam out of the final days. Sit it in the first week of shadow so that a failed attempt can be retaken before go-live, not after.
What does go-live readiness mean for people?
It means you can show, on paper, that every person who approves a held action is certified and current, that the team lead can reconstruct a decision from the record, and that cover exists for absences. The fuller checklist, agent and people together, is in what go-live readiness is. If any reviewer is not ready, keep their queue at depth 2 until they are.
An example
Take a hypothetical insurer launching a claims intake agent. Eight weeks out, all 60 claims staff take the vocabulary course, mostly as audio on commutes. Five weeks out, the 18 handlers who will approve held claims start reviewer modules. For three weeks of shadow, the agent drafts first-notice-of-loss files and the 18 handlers decide every one, logged as their practicum. Go-live is at depth 2 for complex claims and depth 3 for small glass claims within the insurer's limits. Two months later, before the agent takes on simple motor claims at depth 3, the handlers spend an hour on what changes.
What happens after go-live?
The first month is when habits set, and it deserves as much planning as the weeks before. The argument is made in the first month after go-live is the real training. At day 30, run a retrospective using the training retrospective template after an AI agent go-live: what reviewers missed, what they over-held, and what the next update should cover.
Where the Academy fits
Pryme Intelligence Academy's Operator track is built for this timing. Its modules fit the weeks before go-live, and its two-week practicum of real review decisions can run inside the shadow period, signed off by a Pryme assessor. The admin view shows who is ready on the morning of go-live.
Questions people ask
How long should the shadow period be?
Long enough for each reviewer to see real volume, including some held exceptions. Two weeks is a minimum for most queues; slow queues need longer.
What if go-live is moved forward?
Keep the agent at depth 2 until reviewers are certified. Moving the date is a business decision; letting untrained people approve actions at depth 3 is a risk decision.
Who owns the schedule?
The process owner, with learning and development running the calendar. The guide to training teams for AI agents covers how those roles divide.