AI accruals: how to test the estimate before it reaches the ledger
Test AI accruals against contracts, actual invoices, and subsequent adjustments. A worked review method for finance teams.
Maebh Collins FCA7 min read
Published 1 October 2026
An accrual can match eleven months of history and still be wrong in the twelfth. The review needs evidence about the period being closed. A convincing explanation of last month’s pattern will not provide it.
On 28 September, Trintech announced an Accruals Intelligence Agent. Trintech says it recommends amounts from historical activity, tracks estimate accuracy, and attaches evidence and approvals. Those are vendor descriptions. I have not tested the product, and the announcement does not establish its performance on your ledger.
The useful finance question is what would make you accept an estimate. My existing article on AI in the month-end close covers how agent-assisted production changes the team’s work. Here I want to follow one estimate through preparation, challenge, and the later evidence that tells you whether it was any good. This belongs within a wider AI in finance strategy.
Start with the question the accrual answers
Before evaluating an output, write down the expense, entity, reporting period, and evidence cut-off. Identify the service or activity being estimated. State whether an invoice has already been posted and whether another process has recorded part of the cost.
This gives the reviewer something firmer than a generated amount. It also makes a duplicate estimate easier to spot. Two separately plausible schedules can describe the same obligation.
Define the comparison method before running the tool. A recurring support contract might use the current agreement and the period of service. A variable distribution charge might use shipments completed in the period, subject to the information available at close.
Historical averages can help identify an unusual result. They should not quietly replace the evidence that explains this month’s activity. A seasonal business needs a review method that can recognise a seasonal month.
A worked test with three different outcomes
The following figures are fictional, exclude VAT, and illustrate a review method. They are not product outputs or a prescribed accounting treatment. Assume that none of these costs has already been recorded and that the reviewer has checked the applicable accounting policy.
| Cost | Initial estimate | Evidence available at close | Reviewed amount | Later evidence |
|---|---|---|---|---|
| Monthly IT support | €4,000 | Current agreement covers September at €4,000 | €4,000 | Invoice for €4,000 |
| Seasonal distribution | €6,000 | September shipment records support €9,000 | €9,000 | Invoice for €9,300 |
| Terminated advisory contract | €2,500 | Agreement ended in August; no September service identified | €0 | No September charge, confirmed with the supplier |
The initial estimates total €12,500. The reviewed amounts total €13,000. The later evidence totals €13,300. The initial total appears closer to the reviewed total than the individual errors deserve.
That is the first lesson from the schedule. The distribution estimate is too low, while the advisory estimate includes a cost unsupported by the stated facts. Those errors partially offset in aggregate. Reviewing only the total would conceal both.
The second lesson is that the €9,000 distribution estimate still differs from the later €9,300 invoice. A reviewer can make a reasonable decision from the available evidence without predicting the final invoice exactly. Keep the distinction between an unsupported estimate and a supported estimate with later uncertainty.
The later invoice also needs validation. Check its service period, scope, and allocation before treating it as the final comparison. A supplier invoice can contain a correction, another month’s service, or an error of its own.
Keep calculation and interpretation visible
For each line, retain the inputs and calculation used to reach the amount. If shipment count is multiplied by a rate, show both. If a contract is allocated across periods, show the dates and allocation basis.
Then keep the interpretation alongside it. Explain why those inputs relate to the entity and month under review. A correct multiplication does not establish that the underlying service was delivered in that period.
I would use a spreadsheet or controlled calculation for the arithmetic and ask the AI to identify missing evidence, conflicting dates, or unexplained changes. The reviewer should be able to recompute the amount without reconstructing the conversation. The calculation and the commentary have different jobs.
Do not treat a model’s confidence statement as evidence about financial accuracy. If the product supplies a confidence score, ask how it is defined and whether testing supports that interpretation. A high score can still accompany an estimate based on incomplete information.
Test the cases that could change the answer
A useful evaluation set contains more than stable recurring invoices. Include an agreement that ended, a new supplier without history, a price change, and a cost already partly recorded elsewhere. Include an item whose amount is ordinary but whose service period is wrong.
Run the same cases through the proposed process and the current process. Record whether each detected the issue, what additional evidence was requested, and how long review took. Use the same information cut-off so one process does not receive facts that the other could not know.
Keep the expected answer separate from the AI’s output. For a fictional exercise, calculate it independently from the stated facts. For a real case, have an appropriately authorised reviewer establish the reference decision from the documentation.
Do not use a good explanation to mark an unsupported amount as a pass. Equally, do not mark every difference from the final invoice as a failure. Your acceptance rule needs to distinguish calculation error, missing evidence, and uncertainty that existed at the close date.
Measure error at line level and over time
In the example, the reviewed estimates are €300 below the later validated amounts. The difference comes entirely from distribution. That is a more useful finding than saying the schedule was approximately right overall.
Retain both the signed error and its absolute size. Signed error shows whether an estimate was above or below the later comparison. Absolute error prevents an overestimate on one account from cancelling an underestimate on another in the performance measure.
Also record the reviewer adjustment separately. The review changed distribution by €3,000 and removed the €2,500 advisory estimate. Those changes tell you about preparation quality. The subsequent €300 difference tells you about the remaining estimation uncertainty.
Look across periods for recurring direction. An estimate that consistently understates a cost deserves investigation even when each individual difference falls within a review threshold. Check whether the rate, source data, or method needs changing before asking the model to write a better explanation.
Keep reversals, replacements, and final postings connected to the estimate identifier. Otherwise, performance tracking becomes a second reconciliation project. A well-labelled schedule should let the next reviewer follow the sequence without asking who remembers what happened.
Make acceptance a recorded decision
I would require a short acceptance record containing the original proposal, the reviewed amount, the evidence used, the reviewer, and the reason for any change. Capture the relevant workflow version as well. A later evaluation needs to know which method produced the output.
Where evidence is missing, record the unresolved question and its owner. A reviewer should not have to choose between approving an unsupported amount and silently completing the tool’s unfinished research. The exception needs a visible route to resolution.
This supports audit preparation because the estimate’s history is available. It also helps the finance team improve its own process. Repeated corrections should lead to a better input, calculation, or review rule.
Start with one account family and a small set of representative cases. Keep posting approval within the agreed control process while you evaluate the preparation work. Expand only when the evidence shows that the review is reliable and the overall effort makes sense.
The cost of a bad month-end close puts this review discipline in its wider finance context. The next useful step is to take one existing accrual schedule and add the columns needed to follow an estimate through to its outcome. That will tell you more about readiness than another demonstration. If you need senior finance leadership to build that discipline into the close, get in touch.
Maebh Collins is a Fellow Chartered Accountant (FCA, ICAEW) with Big 4 training and twenty years of operational experience as a founder and senior finance leader.