How to teach payroll administrators to bank corrections that improve an AI agent

Payroll administrators bank a correction by recording why they changed the agent's pay input, validation or payslip reply, and sending that reason to the agent owner before the next cut-off. The fix becomes a change to the pay rules the agent reads, its brief or its bounds, and is re-certified. Unbanked by cut-off, it repeats on every payslip.

What does the agent do in payroll?

A payroll agent validates timesheets and variable pay inputs against pay rules, calculates proration for starters and leavers, flags anomalies before the run, and drafts replies to employee payslip queries. It works at depth 1 and 2: it checks and drafts, and an administrator approves inputs and sends replies. Final pay runs and payments stay with people.

Administrators correct it in familiar ways. It applies the standard overtime rate to a grade covered by a collective agreement with a different rate. It prorates a starter by calendar days when the policy uses working days. It misses that an employee's tax code changed after the last run. It answers a query about a deduction by quoting the generic pension rule rather than the employee's scheme.

The broader role map is in AI agent skills by role.

Why does the pay calendar drive the process?

Payroll has a hard rhythm. Inputs close on a fixed day, the run happens, payslips go out. A correction made on the day inputs close is useful only if it reaches the agent before the next run. Miss the window and the same error appears on every affected payslip, often across hundreds of employees.

So banking in payroll is organised around the calendar. Corrections from one run are reviewed in the days after payday, changes are made and re-certified mid-cycle, and the agent runs on the improved rules for the next close. Teaching this rhythm matters more than any single technique.

What are the steps?

1. Agent, depth 1: validates inputs against pay rules, flagging anomalies with the rule it applied.
2. Agent, depth 2: drafts proration calculations and payslip replies, citing the rule and source.
3. Person (payroll administrator): corrects and records a reason: wrong pay rule, wrong agreement, stale tax data, proration method, unclear brief, or one-off.
4. Agent, depth 1: groups reasons after payday by rule, grade and site.
5. Person (payroll manager): decides what to bank in a review held in the first days after payday.
6. Person (agent owner): makes the change. A collective agreement rate goes into sources from the HR or reward owner; a working-days proration rule goes into the brief; anomalies above a value threshold become a hold in the customer's limits.
7. Attest replays the agent against past pay periods mid-cycle, so the improved agent is certified before inputs close.

What do payroll administrators need to learn?

Where pay rules live. Most payroll corrections are source problems: the agent read an out-of-date rate, a missing agreement or an old tax code. Administrators who learn to ask "what did the agent read?" before "what did it do?" find the fix faster. The answer is often that a rule changed in the HR system and never reached the agent's sources.

How to separate a pay rule error from an approved one-off. A manager may approve a one-off payment outside policy. If an administrator corrects the agent's flag on it and someone banks that, the agent stops flagging unusual payments. One-offs approved by a manager are recorded with the approver's name and never banked.

How to write reasons without pay details. "Applied standard overtime rate to grade covered by site agreement" is enough. Names and salaries do not belong in the record of corrections. The review habits beneath this are covered in how payroll administrators review an AI agent's decisions.

Which controls and evidence matter?

• Pay rules from the system of record. Evidence: provenance on each rate and rule cited.
• Post-payday review. Evidence: the review notes and banked changes, dated.
• Re-certification before input close. Evidence: a certificate dated before the cut-off.
• Manager approval for one-offs. Evidence: approver name on exception reasons.

What should payroll measure?

Repeat corrections per rule from one run to the next. Corrections made after inputs closed, which are the expensive ones. Time from a rule change in the HR or reward system to the agent's sources. Payslip queries that came back a second time after the agent's reply. And, critically, anomalies the agent missed that a person caught only after the run.

Where does it usually go wrong?

Banking misses the cut-off. The review happens too late, re-certification runs into input close, and the team decides to wait a cycle. Hold the review in the first days after payday, every time.

Rule changes do not reach the agent. Reward agrees a new rate, HR updates the system, payroll learns at the run. Make the source update part of the rule change.

Administrators loosen anomaly thresholds to reduce flags in a busy month. Fewer flags feel like progress until an overpayment gets through. IT teams face similar access-related pressure on payroll systems, covered in teaching IT administrators to bank corrections, and support teams face the same reply-quality issues at higher volume, as teaching customer support agents to bank corrections describes.