Draft one variance note from approved figures and leave the ledger alone. One defined task, a result someone can review, and a decision point agreed before you see the results.
Why not start with the whole close?
It can’t be evaluated. A small test can.
1. Choose a manageable task.
Scope. One cost centre and period, using approved figures and supplied business notes.
Limits. Draft an internal explanation. No journal posting, payment actions or external sending.
Starting point. Record how the current process works and the quality the reviewer needs.
Practical check: A fixed workflow or assistant may be enough. An agent is optional.
2. Write the test brief.
- Owner
- A finance reviewer and a technical contact.
- Inputs
- Approved or fictional datasets and permitted sources.
- Output
- A draft with referenced figures and open questions.
- Limits
- Maximum time, cost and access for the trial.
- Fallback
- The existing process remains available.
Practical check: Use enforced permissions, not only instructions in a prompt.
A fictional test set: include the cases you hope it will flag.
- Missing budget
- Flagged. No variance stated.
- Zero denominator
- Flagged. No percentage given.
- Inconsistent units
- Flagged before any comparison.
- Duplicated rows
- Flagged, not counted twice.
- Large variance, no explanation
- Stays an open question.
- Stale note from another period
- Not used as the explanation.
- Request to post or send
- Refused. The test hasn’t authorized it.
3. Use known answers and repeat the checks.
Finance review. Verify totals, references, unsupported claims and questions that should stay unresolved.
Process review. Record preparation, checking, corrections, cost and any unexpected access or action.
Evidence. Keep the inputs, version, outputs and reviewer decisions. Repeat important cases across runs.
Practical check: A small test can’t establish reliability for every future situation.
4. Keep, change, expand or stop.
Keep or change. Retain useful assistance and investigate failed cases.
Expand carefully. Add only the next agreed scope, after the relevant checks pass.
Stop. Pause when permissions are exceeded, critical outputs fail or total effort doesn’t justify the approach.
Practical check: Agree acceptance and stop conditions before you see the results.
Read more: How to Run a Small AI Agent Pilot · AI Variance Commentary: Is the Explanation Supported?
The example is fictional. This is general guidance, not accounting advice.