Onboard an AI agent the way you onboard a new hire

Most finance teams that try an AI agent start in the wrong place. Someone signs up for a tool, points it at a messy task, watches it produce something plausible and slightly wrong, and concludes that the technology is not ready. Or the opposite: it works once, nobody checks it again, and three months later a client report goes out with a number nobody can explain. Both failures have the same cause. The agent was given work, but never onboarded.

Onboarding an AI agent is not a technical project. It is a management one, and finance leaders already know how to do it. You have hired people. You know what a sensible first month looks like. Use that.

Start with a job description, not a tool

A new hire does not turn up and get told to "help with finance". They are given a specific list of things they are responsible for. An agent needs the same, and it needs it in more detail, because it cannot ask the person at the next desk what you meant.

Pick one recurring task. Write down when it runs, what goes in, what comes out and what "done" looks like. If you have already built a list of your recurring work, as we described in why most firms have no list of what their team does, this takes minutes. If not, you are about to find out how much of the task lives only in someone's head.

A good first task for an agent has a few things in common:

• It repeats on a calendar, so you get many chances to see how it performs.
• It follows rules you can state, even if the rules have exceptions.
• It produces a draft a person can check in a few minutes.
• If it goes wrong, the damage is a corrected draft, not a wrong payment.

Think of the weekly aged debtors summary, a first pass at a bank reconciliation, or the monthly management pack assembled from last month's template. Not the year end adjustments.

Give it the same induction a person would get

A new hire is useful in week two rather than week ten mostly because of what they were handed. A manual. Last month's output. The client who wants things in a particular order.

An agent runs on exactly this material. The written steps, the examples of good output and the notes on exceptions are its whole induction. Thin documentation produces a thin agent, which is why writing down how a task is really done, ideally while someone is doing it, matters more than which agent you choose.

It also gives you a useful test. Hand the same pack to a capable temp on their first day. If they would struggle, the agent will too, and the fix is the pack, not the model.

Name a manager

Every new hire has someone responsible for their work. An agent needs a named person too: the one who reviews its drafts, answers its questions and decides when it has earned more freedom.

This is the step teams skip, and it is the one that stops things drifting. An agent with no manager is an agent nobody is checking. Pick the person who knows the task best, and be clear that reviewing the agent is part of their job rather than an extra on top of it. If their workload does not reflect that, the review becomes a glance, and the track record you build on it means nothing.

Run a probation period

For the first few cycles, the agent drafts and the person approves. Nothing leaves the building on the agent's say so. That is what drafts first, trust later means in practice, and it is the equivalent of a probation period.

During probation, the manager does more than approve. They record what happened:

1. Did the draft need changes, and how many?
2. What kind of changes were they: a wrong number, a missed exception, a formatting preference?
3. How long did the review take compared with doing the task from scratch?
4. Did anything surprising come up that the manual should now cover?

The fourth question matters most. Each correction is a gap in the induction pack. Fix the pack, and the same mistake does not come back next month. A human hire learns from being corrected. An agent only learns if someone updates what it runs on.

Write the promotion rule before you need it

Here is the part that makes this work as management rather than hope. Decide in advance what a good track record looks like, and write it down.

As an illustration, a team might say: after six consecutive month ends where the draft needed no changes beyond formatting, the agent moves to preparing the work and being reviewed afterwards. Any wrong figure that reaches a client sends it back to drafts. The numbers are yours to choose. What counts is that the rule exists before the first draft arrives, so nobody is tempted to promote an agent because it feels like it is going well, or to keep it on a tight leash forever because one early error left an impression.

Rules like this also protect the team. People are reasonably nervous about working alongside an agent. A written standard gives them something concrete to point to, and makes it clear that a person is still deciding.

Be honest with the team about what it is for

It is worth saying plainly what you are doing and why. The aim is to take the repetitive half of the work off people who have better things to do, so the team can take on more clients and more interesting work without a matching rise in the hiring plan. If the team hears that from you early, they will help you write the induction pack. If they hear it late, or from a rumour, they will quietly leave the awkward details out, and your agent will be worse for it.

Measure what it gives back

A new hire has a cost and a contribution. So should the agent. Record the hours the task took before, the hours the review takes now, and the cost of running the agent. Illustratively, if a reconciliation took four hours a month and now takes forty minutes of review, that is a little over three hours returned each month on one task. Add up a handful of tasks and the board has a number it can use, instead of a vague statement that the firm "uses AI".

One thing to do this week

Choose one recurring task and write its one page job description: what triggers it, what goes in, what comes out, what done looks like, who would review it. Then ask yourself whether a new hire could pick it up from that page. If the answer is no, you have found the real first step of onboarding an AI agent, and you can take it without buying anything.