How to onboard an AI agent like a new hire

Give the agent what you would give a new starter: a written role, a named manager, access to what the job needs and nothing more, induction material it can cite, supervised first tasks and a probation review at the end. Then mind where the analogy breaks: the agent does not learn from experience, so its manager changes its brief.

Why use the new-hire model at all?

Most organisations already know how to bring a person into a regulated role. They write a job description, assign a manager, provision access, train, supervise and review. Each step exists because someone once learned the hard way what happens without it. An agent doing the same work needs the same steps for the same reasons.

The model also helps people accept the agent. Staff understand a colleague on probation. They are less comfortable with "a system that has been switched on". And managers already know what their role is when someone new joins the team.

The hiring manager is the process owner: the person whose team the agent joins. They keep the decisions about the role, the access, the probation outcome and any change in responsibility.

What are the steps?

1. Process owner: writes the role. One page: the work, the systems, what good looks like, what the agent must never do, and when it must ask. This becomes the agent's brief.
2. Process owner: names the manager and the buddy. The manager is accountable for the agent's work. The buddy is the experienced team member who reviews its drafts in the first weeks.
3. IT and process owner: provision access for day one. Read access to the systems the job needs. No write access yet. The question of what is appropriate at the start is answered in what a new AI agent should be allowed to do on day one.
4. Process owner: assembles the induction pack. The procedures, policies, templates and worked examples the team uses, each version-controlled. These become the agent's sources.
5. Agent, depth 1 and 2: first tasks. It reads cases and drafts what it would do. The buddy compares with what the team did.
6. Buddy: gives feedback by recording what was wrong in each draft and why.
7. Process owner: revises the brief and sources in response. This is the agent's equivalent of learning.
8. Risk or quality: runs the probation assessment. The agent is replayed against a set of real past cases and scored against a floor for each gate.
9. Process owner: confirms or extends probation and records the decision.

Where does the analogy break?

The agent does not learn on the job. A new hire who is corrected tends to improve. An agent repeats its mistake on the next case unless its brief, sources or bounds change. Feedback must be turned into a change in those, by a person. We call this qualification, not training: nothing about the underlying model is updated for your organisation.

Probation is not passed once. A person who passes probation is employed until something changes. An agent's certification should lapse after a fixed period, 90 days in our model, and be renewed on fresh evidence. A failed run during that period drops its depth the same day.

Its work can be replayed exactly. You cannot ask a person to redo last month's cases to see whether they would decide the same way. You can ask an agent. Use this: the probation assessment should replay real cases, not invented ones.

Responsibility is per task, not per person. A new hire who proves reliable is usually trusted more broadly. An agent earns one depth, and each action needs its own. Doing well on address changes says nothing about refunds.

Decisions about people need meaningful review. If the agent's job touches decisions with legal or similarly significant effects on individuals, the review step must be real, not a signature. The standard is set out in GDPR Article 22 human review: what meaningful intervention looks like.

What controls apply, and what record does onboarding leave?

The agent's file should mirror a personnel file:

• The role, version-controlled, with the date and the approver.
• The manager and buddy, with any changes over time.
• Access granted, per system and per capability, with dates.
• The induction sources, with versions.
• Feedback and revisions: every correction the buddy recorded and the change it led to.
• The probation assessment: cases used, scores, floor, result.
• The confirmation decision, signed by the process owner.

This file is the basis for every later decision about the agent, and it is what an auditor will ask for. It fits inside the wider record kept through train, shadow, certify, widen.

Where does it usually go wrong?

No one writes the role. The agent is configured by the team that built it, from a conversation. Six months later no one can say what it was meant to do.

The buddy is too busy. Reviewing drafts is added to someone's full workload. Feedback becomes thin, and the agent's brief never improves. Give the buddy time.

Access arrives all at once. IT provisions the agent's service account with everything it might eventually need. Grant per capability, as you would for a junior.

Probation is skipped for a good demo. A convincing demonstration is not evidence on your cases. Run the assessment anyway.

The manager leaves. Agents outlast reorganisations. Make the manager a role, with a named holder, and update the file when it changes.