How do you train new joiners when AI agents already run part of the work?

Teach the work and the agent together, in order. New joiners learn how governed agents decide, then the workspace, then the work through the agent's completed cases, and finally shadow the live queue before they approve anything alone. The trap is letting them learn the job by watching the agent and assuming it is right.

Why is onboarding harder once agents are live?

In a team without agents, a new joiner learns by doing the routine cases first and the hard ones later. Once an agent handles the routine cases at depth 3, the new joiner never sees them. They start on the held queue, which is by definition the harder work, without the grounding the routine cases used to give.

There is a second risk. New joiners see the agent producing confident, well-written drafts and assume it knows the job better than they do. They approve because they lack the confidence to challenge. That is the opposite of what a reviewer is for.

What should onboarding cover, and in what order?

1. Foundations, in week one. What agents do, depth, the five bounds of the minimum chain, the four verdicts and what a lapsed certificate means. Without this, nothing in the workspace makes sense.
2. The work itself. The policies, the rules and the reasons behind them, taught using the agent's completed routine cases as worked examples. The new joiner reads cases the agent handled alone and explains why each outcome was right.
3. Reading the workspace. Sources, provenance, the bound that held an item, and how to find the record of past decisions.
4. Writing reasons. How to change or decline a draft so a colleague could follow the reasoning.
5. Shadowing. Two weeks or so deciding live held items alongside a certified reviewer, whose decision takes effect.
6. Supervised approval. Approving alone on a narrow set of held items, with a higher sampling rate, before full approval rights.

Job descriptions change too; the shifts are covered in how job descriptions change when AI agents do part of the work.

What does it look like in a collections team?

Take a lender's collections team where an agent sends reminders and proposes payment arrangements, with arrangements above a set value held. A new joiner starts. In week one, they complete foundational training and read fifty arrangements the agent agreed alone, explaining why each was within policy. In week two, they learn the workspace and practise writing reasons on past held cases. In weeks three and four, they shadow the held queue. On day two of shadowing, they would have approved an arrangement for a customer whose notes mentioned a recent bereavement; the certified reviewer held it for a specialist. The new joiner learns, from a real case, why vulnerability is a red line. A role-specific version is in how to onboard a new collections agent to an AI agent workspace.

Which controls protect the work during onboarding?

• No approval rights until the shadow period is complete and a defined standard is met.
• A higher sampling rate on new joiners' first weeks of approvals.
• A named buddy, ideally the team's champion, for questions.
• A record of each new joiner's shadow decisions and clearance.

Teams that run shifts need the same programme for every shift, which takes planning; see how to train reviewers who work different shifts. Where a second agent arrives while the first is live, existing staff face a similar challenge, covered in how to train a team for a new AI agent when another is already live.

Where the Academy fits

New joiners can complete Pryme Intelligence Academy's free Foundations track in their first week. Those who will review held items go on to the Operator track, whose two-week practicum on real review decisions matches the shadow period above. Seats are included with paid Pryme Intelligence plans, so new joiners can start without a separate purchase. The topic guide on training teams for AI agents covers the wider sequence.