One that names every role, starts with a low-risk agent, and moves people forward only as fast as the agent widens. It runs from the first lesson to a month after go-live: shared vocabulary for everyone, reviewer practice on real drafts in shadow, certification before the agent acts, and an update before each later step in depth.
What makes a first deployment different?
Nobody in the organisation has done it before. There are no experienced reviewers to learn from, no calibration cases from past incidents, and no settled view of what a held action should look like. The plan has to build all of that, as well as the skills.
That is why the choice of first agent is part of the training plan. Pick a process with real volume, low harm and clear right answers: drafting replies to routine enquiries, matching documents to records, preparing standard letters. The team learns the vocabulary, the review habit and the record on work where a mistake is cheap.
Who needs to be named in the plan?
Five roles, even if one person holds two of them:
- The owner, who decides each capability's depth and answers for the process.
- The builder, who writes the agent's brief, chooses sources and bounds, and reads certification results.
- The reviewers, at least two, who approve the agent's held actions.
- The IT administrator, who manages access, connections and roles, and must understand what each permission lets the agent reach. The training plan template for IT administrators covers that role.
- The team lead, who runs calibration and covers absences.
What does the plan look like, phase by phase?
| Phase | The agent | The people |
|---|---|---|
| Weeks 1 to 2 | Being chosen and briefed | Everyone learns the vocabulary; the team takes a readiness check; roles are named |
| Weeks 3 to 4 | Being built and certified on your own cases | Reviewers take the reviewer modules; the builder and owner read certification results together |
| Weeks 5 to 7 | Shadow at depth 1 or 2: drafts, a person decides everything | Reviewers decide its drafts as practicum; weekly calibration; one incident drill |
| Week 8 | Go-live at depth 2, or depth 3 for certified routine capabilities | Only certified reviewers approve held actions |
| Weeks 9 to 12 | Live; first step up in depth if evidence supports it | One-hour update before the step; retrospective at day 30 |
The agent side of the same timeline sits in the AI agent deployment plan template. Keep the two plans in one document so neither runs ahead.
What should reviewers practise most?
Two things. First, provenance: every held draft should be checked for where its facts came from, and reviewers should decline anything resting only on an external source. Second, giving reasons. A first deployment has no history, so the reasons reviewers write when they decline a draft become the material for calibration, for the builder's next revision, and for the next cohort's training. "Wrong" is useless; "the delivery date came from the customer's message, not the tracking record" teaches everyone. The method is in how to train people to give good reasons when they decline an agent's draft.
How should the formal learning be delivered?
In short pieces that fit around work. Many people in a first deployment are doing their normal jobs at the same time. Lessons people can listen to on a commute or between tasks get finished; day-long workshops get postponed. The case for that approach is set out in what audio-first learning is. Keep live time for calibration and drills, where people need to argue about real cases.
What counts as success?
Set criteria before you start. Reviewers certified before go-live with no missed red line in their practicum. The owner able to trace one decision through the record. Calibration disagreement falling over the shadow period. And at day 30, a retrospective that names what people still find hard. If those are met, the plan has done its job, whatever the agent's depth.
An example
Take a hypothetical regional housing association choosing its first agent. It picks an agent that drafts replies to repair status enquiries from the repairs system. All 40 office staff learn the vocabulary in week one. Four customer service staff and their team lead become certified reviewers during a three-week shadow. The IT administrator learns what each connection exposes. The agent goes live at depth 2. In week ten, with evidence in hand, it starts sending status updates drawn only from the repairs system at depth 3. Small businesses can run a leaner version, described in what good AI agent training looks like for a small business.
Questions people ask
How long should a first deployment's training take?
Around two months from first lesson to go-live is common for one agent, with the shadow period the longest phase. Rushing the shadow is the most common mistake.
Should the first agent be the one with the biggest return?
Usually not. Choose the one where your team will learn fastest with least harm. The bigger returns come from the second and third agents, run by people who know what they are doing.
What should we keep for the second deployment?
Calibration cases, reasons for declines and the retrospective. Pryme Intelligence Academy's admin view keeps certificates and expiry dates in one place, and the guide to training teams for AI agents covers the wider programme.