To choose automatic reconciliation software or accept any finance AI agent, run it on one defined duty using cases it has never been tuned on. Check what it clears as well as what it flags, test what it is allowed to do, and write down the scope it has earned.
A finance agent can produce a convincing supplier-balance packet. You still decide what your team can rely on, which work needs supervision and what the agent may do next. The method covers close preparation and other duties as well.
Start with one reconciliation duty
Write the job in terms your accounting and systems owners can both review. For example: reconcile one entity’s supplier statements to approved ledger records and prepare an exception packet with the difference, evidence, uncertainty and proposed next step.
Specify the entity, population, period, required sources and reviewer. Treat preparing a packet, recommending an action and carrying out that action as separate permissions. Reading a statement gives no right to change supplier details or send an external message.
Name the evaluation stage: a historical read-only review of records and proposed outputs, a sandbox or simulation that tests action boundaries, or a supervised run on live work under permissions agreed for that stage. Each move between stages needs its own recorded decision.
Reserve cases the agent has not been tuned on
Have an accountant set expected dispositions on their own. Use some cases to configure the duty and reserve different ones for evaluation. Before you trust any result, check the history behind the agent’s answer.
A reserved case used to tune a rule becomes a setup example, so the next assessment needs fresh ones, or the improvement reflects familiarity with the corrected case.
Include ordinary matches, material differences and cases that should stay unresolved. In the illustrative credit example used across this series, missing application history leaves the available balance unresolved, and the expected behaviour is to name the missing period and route the question to an owner.
Judge each case on the evidence available at the review date. A later settlement shows how the case ended, and the original reviewer never saw it.
Review cleared items as well as flagged exceptions
A strong packet tells you nothing about the exceptions the agent missed. Review a sample of the work it clears, and account for excluded, failed and waiting items.
Take an illustrative set of 20 constructed cases. The accountant identifies four exceptions. The agent flags five cases: three valid findings and two incorrect flags. It misses one of the four exceptions.
You now hold two separate observations: three of five flags are supported, and the agent found three of four expected exceptions. Neither describes the unresolved or excluded work unless you record it. The figures are illustrative arithmetic.
Report rare failure cases you added on purpose apart from the ordinary population, since their frequency in a constructed test says nothing about operations.
What to look for in automatic reconciliation software
Buyer guides compare matching, exception queues, approvals, audit trails and access. Map each feature to one of seven acceptance dimensions, record each dimension on its own and test it on your reserved cases.
| Dimension | What the reviewer needs to inspect | Feature it tests |
|---|---|---|
| Population | Processed, excluded, failed and waiting items, tied to the intended source scope | Coverage reporting |
| Correctness | Independent dispositions, reproducible calculations, valid flags and missed exceptions | Matching rules |
| Evidence and uncertainty | Supporting source versions, conflicting records and explicit missing information | Exception detail |
| Escalation | A named owner receives the unresolved case and usable partial work | Exception routing |
| Authority | The tested roles cannot take actions outside the agreed scope | Approvals and read-only access |
| Decision history | Earlier outputs, corrections and review records remain retrievable | Audit trail |
| Operating effort | Actual preparation, review, correction and follow-up effort | Time-saving claims |
Score each as Pass, Fail or Not tested and keep the case evidence. Define before testing the errors that stop progression. A material authority failure stays visible even when the financial explanation is excellent.
No universal match rate settles acceptance; relate errors and tolerances to the duty’s significance and your existing review. On Rever, Matcher holds this duty: it agrees open items across vendor statements, customer accounts and intercompany ledgers, and Investigator traces the differences that persist.
Test the action boundary in a controlled environment
In an agreed simulation or isolated sandbox, attempt self-approval, a source change by a read-only role and a send without the required approval. Keep what happened.
Then change evidence after a packet has been reviewed. Check that the affected decision returns for review and that the earlier version stays explainable. Approval of one packet for one action covers that packet and that action. Human approval in finance AI sets out how to define that boundary.
Rever’s security and trust page states that agents prepare and your team approves, that connections are read-only and that each connection is a direct integration rather than an agent holding credentials. Inspect those statements for the proposed duty and record your own result.
Decide the next permitted scope
Choose one outcome: continue within tested limits, restrict the population, remediate and retest, or stop. Record the decision, the accounting owner, unresolved limitations and the conditions that would suspend the duty.
If supervised observation follows, measure the work as it arrives, because a curated test can hide a review queue your team cannot sustain. Keep checking cleared items, and treat changes to rules, sources or authority as changes to the accepted scope. Rever applies the same rule to its own agents: every new agent starts as an observer, and Eve reviews its performance, promotes what performs and demotes what does not. The Explore Rever page on AI agents for finance lists the fifteen agents and the published scope of each.
One Duty, One Acceptance Decision holds worked cases with expected behaviour and a blank scorecard for what may proceed and why, with every result starting as Not tested.
To plan the review before you run it, the One Duty Acceptance Planner turns one proposed duty into a printable acceptance review agenda, with every check Not tested until your team records evidence.
