How to onboard a new accounts payable clerk to an AI agent workspace

Onboard a new accounts payable clerk in two weeks by teaching the agent's work before the clerk's own. Week one: how the agent matches and holds invoices, why an emailed bank change can never pay, and shadowing a certified reviewer. Week two: deciding held invoices with every decision checked, then sign-off. No solo approvals until the red lines are proven.

What does a new clerk walk into?

In a team where agents already work, the accounts payable agent typically reads incoming invoices, matches them to purchase orders and goods receipts, flags duplicates and drafts payment proposals. Matching and flagging happen at depth 1 and 2. Some teams let the agent approve exact three-way matches below a value within the customer's limits at depth 3. Releasing payments sits higher and stays with people in most teams.

So the new clerk's job is different from the one they may have done elsewhere. Less keying, more judgement on what the agent held: price variances, partial deliveries, invoices with no purchase order, and supplier bank detail changes. The day-to-day shift is described in what changes in an accounts payable clerk's day when AI agents arrive. The AP team lead owns the onboarding; a certified senior reviewer does the daily coaching.

What happens in the first two weeks?

1. Day 1. Person (team lead): explains the agent's job and limits. What it reads, what it drafts, what it may do alone, and what it must hold. Show the bounds on screen.
2. Day 1 to 2. Person (new clerk): learns provenance. A purchase order from the ERP is sealed. A supplier's invoice is signed or external depending on how it arrived. An emailed letter changing bank details is external, and an external claim cannot authorise a payment.
3. Day 2 to 3. Agent, depth 2: drafts matches and holds while the new clerk watches a senior reviewer decide them, explaining each verdict aloud.
4. Day 4 to 5. Person (new clerk): practises in a sandbox on past held invoices, with the outcomes hidden, then compares their decisions with what actually happened. A sandbox workspace keeps practice away from live payments.
5. Week 2. Person (new clerk): decides live held invoices, with the senior reviewer checking every decision before it takes effect.
6. Week 2. Person (new clerk): records reasons on every correction, so the agent owner can bank patterns.
7. End of week 2. Person (team lead): signs off limited approval rights if accuracy is high and no red line was missed; otherwise extends supervision.

Which red lines must be proven before solo work?

Three, and they should be tested deliberately in week two.

Bank detail changes. The clerk must never approve a change based on the email or letter that requested it. Verification goes through a contact already on file, and the agent holds every change for a person. Put a test case in the sandbox: a convincing change request from a known supplier. A clerk who approves it is not ready.

Invoices without a purchase order. The agent holds them; the clerk must know who may approve them and must not approve outside their own authority.

Duplicates with small differences. A resubmitted invoice with a new number and the same amount. The clerk must check before clearing a duplicate flag.

What should the clerk learn to correct, and how?

Corrections are how the agent improves, so a new clerk should learn the habit from day one. Every time they change the agent's match or overturn a hold, they pick a reason: wrong source, stale price list, tolerance unclear, or one-off. A plain reason lets the agent owner change the agent's brief, sources or bounds rather than leaving the same error to recur. The AI agent training plan template for accounts payable clerks sets this out as a module, and the reviewer onboarding checklist for the accounts payable team turns the two weeks into tick boxes.

Which controls and evidence matter?

• Supervised approvals in week two. Evidence: every decision shows the clerk and the checking reviewer.
• Sandbox test results. Evidence: the cases attempted and the outcome, including the bank change test.
• Approval rights granted, not assumed. Evidence: the access change dated after sign-off.
• Reasons on every correction. Evidence: reason codes against the clerk's name.

What should the team lead measure?

Accuracy of the clerk's decisions in week two, judged by the checking reviewer. Red-line misses, which must be zero. Quality of correction reasons. Time to decide a held invoice, which will be slow at first and should fall. And, after sign-off, a monthly sample of the clerk's solo decisions for the first quarter.

Where does it usually go wrong?

Speed pressure in month-end. A new clerk starts in the last week of the month, is handed a full queue, and approves what the agent drafted because there is no time to check. Start new clerks away from month-end where possible.

Skipping provenance. Clerks experienced in other teams know invoices but not provenance classes, and assume an emailed PDF is as good as a portal submission.

Granting approval rights on day one "to keep the queue moving". The supervised week exists to catch exactly the errors that a confident newcomer makes.

Teams onboarding into other agent-heavy roles face the same trade-offs, as the plans for a new property manager and a new collections agent show. The wider approach is in training teams for AI agents.