An accountant should use AI during reconciliation where it makes the evidence easier to inspect: preparing comparable records, proposing uncertain matches and organizing questions about exceptions. Keep accepted matches, unresolved items and proposed accounting entries visibly separate. The useful output is a reconciliation another qualified person can retrace, with authority to approve changes still explicit.
This article follows a bank reconciliation, comparing the bank's transactions and balance with the company's records and explaining differences. Other reconciliations need their own sources and acceptance checks. The finance and FP&A guide places this task within a wider finance rollout, including variance explanations, forecasts and review responsibilities.
Preserve the records before asking for matches
Start with a defined bank account and period. Keep the original statement or feed export and the corresponding ledger extract, the company's record of transactions. Confirm the currency, date basis and opening and closing balances being compared. An assistant cannot repair an incomplete population by writing a convincing explanation of it.
Xero's July 2026 bank-reconciliation glossary describes comparing the records and explaining differences, including timing items and charges not yet recorded. Its small-business guidance is a useful orientation, rather than a control standard for every company. Your controller needs to define the actual account, process and review requirements.
Before a trial, preserve these details:
- Source identity. Record which account, statement, period and export version each file represents.
- Completeness. Check imported row counts and totals against the originals, with missing or duplicated lines investigated.
- Interpretation. Confirm whether an amount is money received or paid, which date is being used and how references are carried through.
- Traceability. Give each source row a stable reference that survives sorting, filtering and grouping.
Keep an unchanged copy of the inputs. If a description is standardized for matching, retain the original beside the transformed value. If the statement is revised, identify the new version rather than overwriting the evidence behind an earlier decision.
Use only an approved environment for the actual records. Bank exports may contain personal information, customer details and account identifiers. Follow the task-specific source-preparation checks, including whether the selected route preserves tables and permits the intended access. Removing an account number alone does not establish permission to upload the remaining data.
Keep working matching rules when they do the job
Run the established matching process against the preserved inputs before deciding what AI should change. A repeatable rule may already handle a known reference, amount and date relationship. Document its conditions and exceptions. A match based on a defined rule and a model's suggested association should not arrive with an indistinguishable status.
Microsoft's Business Central reconciliation documentation, updated in October 2026, illustrates a specific division of work. Ordinary automatic matching runs first, then Copilot proposes additional matches for the remaining transactions. The documented review stage allows proposals to be inspected and discarded before acceptance. That is one product's behavior, not evidence that every reconciliation assistant works this way or produces reliable matches.
The case for keeping a baseline is practical. In a June 2026 Reddit discussion, a self-described senior accountant reported having a working Excel reconciliation and being asked to recreate it with AI. Their later edit clarified that the finance leader wanted a broader month-end-close experiment, using the existing spreadsheet to validate results. Replies questioned the benefit of replacing something already automated; another reader saw a useful paid learning exercise. Neither view establishes a successful rollout. The distinction matters: learning a new method and improving the recurring reconciliation are different objectives.
An assistant may help describe an exception, find possible relationships in inconsistent descriptions or draft a question for the person holding a missing record. Test those contributions against the existing process. If it helps write a script or formula, review and test that artifact before using it; generating the code does not establish that the matching logic is correct.
Follow one ambiguous payment through review
Suppose Jamie Lee is an accountant at a catering company, reviewing a permitted bank extract against the ledger. A receipt of US$480 may correspond to two recorded receipts of US$200 and US$280. A separate bank payment of US$75 has two possible US$75 ledger entries. There is also a US$25 bank charge with no identified ledger counterpart.
The amounts help Jamie investigate, but do not settle all three cases.
| Bank item | Proposed relationship | Evidence or decision still needed |
|---|---|---|
B17, US$480 received | Receipts R41 and R42 total US$480 | The settlement record must identify both receipts in this bank deposit |
B18, US$75 paid | Payment P43 or payment P44 has the same amount | A bank reference or remittance must distinguish the candidates |
B19, US$25 paid | The description suggests a bank charge | Verify the charge and existing records before proposing a new entry for authorized review |
For B17, US$200 plus US$280 equals US$480. Jamie still needs the settlement evidence connecting those particular receipts to that deposit. Adding unrelated amounts until they equal a bank line is not a supported match. Preserve both receipt references when the relationship is accepted, and prevent either receipt from being reused in another accepted allocation.
B18 exposes a different problem. Both ledger entries fit the amount. Selecting the first one in a sorted list would hide the ambiguity rather than resolve it.
One bank payment, two possible ledger records
B18 records US$75 paid. The available amount does not identify the invoice.
Invoice 118, US$75
A payment to Harbor Foods is recorded against invoice 118. Jamie still needs the remittance or bank reference connecting it to B18.
Invoice 124, US$75
Another US$75 payment to Harbor Foods is recorded against invoice 124. It fits the amount too, so row order cannot decide the match.
Jamie should retain both candidates and request the missing reference from the colleague who arranged the payment. If the evidence identifies P44, record why P44 was accepted and what remains unresolved about P43. A later reviewer should see the decision, not just a green cell.
For B19, matching and creating an accounting entry are separate operations. A plausible expense classification is still a proposal. Jamie must check the source and whether the charge is already recorded elsewhere, then follow the organization's posting and approval arrangements. Do not add an entry merely to make an unexplained difference disappear.
These rows are an excerpt for examining the review work. Jamie still has to complete the account reconciliation, including the applicable balances and remaining timing items. Three plausible explanations do not establish that the whole account is reconciled.
Give the reviewer a record they can retrace
A useful review record connects each conclusion to the source and matching basis. It should let the receiving accountant move from a proposed relationship to the relevant bank line and ledger entries, then back again without searching through a narrative.
Microsoft's Financial Reconciliation agent documentation describes mapping and monetary keys, potentially matched groups and a detailed report with transaction identifiers. Its report summary and troubleshooting explanation are separate from those underlying results. The documentation also notes that transaction hyperlinks are optional and disabled by default. Inspect the actual configured report; do not assume that a generated summary includes a usable route to every source row.
For Jamie's review, retain:
- Inputs and rules. Identify the source versions, comparison period, matching conditions and any tolerances used.
- Proposals and decisions. Show the candidate records, supporting evidence, accepted or rejected relationship and responsible reviewer.
- Open work. Keep unresolved amounts, missing evidence, the next owner and the required follow-up visible.
- Changes and completion. Separate proposed entries from authorized changes, confirm what was actually saved and retain the final review result.
An amount tolerance is a rule to approve and test, not permission for a model to invent an acceptable difference. Recheck its behavior with small discrepancies and duplicated values. Preserve the reason for an exception instead of treating every item within a threshold as automatically understood.
Check the actions attached to acceptance too. The Business Central guidance describes an option that posts a fully applied reconciliation when proposals are kept, and a separate flow that posts differences to suggested accounts. A button that appears to retain a match may therefore have a wider configured effect. Establish the effective action boundary before a trial and keep posting outside an assistance-only scope.
Test accepted work rather than a reassuring match rate
Build the trial around cases whose expected result an experienced accountant can establish from the originals. Include an ordinary match, a supported grouped receipt, an ambiguous same-amount payment, a missing ledger item and a duplicate import. Use the preserved baseline and include cases the assistant gets wrong or leaves unfinished.
Examine both kinds of failure: a false match that clears the wrong records and a missed item that remains hidden or unexplained. A high matched percentage can conceal the former. A zero difference can also conceal offsetting mistakes or the reuse of a record.
Record preparation, investigation, review, corrections and support effort across the whole handoff. Check whether another qualified colleague can reproduce the relationship and maintain the process after a source format changes. The workflow-redesign guide helps compare that complete result with the work being replaced.
Questions about AI during reconciliation
Can an accountant give AI the whole bank reconciliation?
Define the intended assistance and effective actions first. A general instruction to reconcile an account does not settle source completeness, matching rules, exception handling or posting authority. Begin with permitted records and a checkable output, compare it with the established process and keep unresolved items visible. Use the input checks before widening the scope.
Is an equal amount enough to accept a match?
An equal amount identifies a possible relationship, not necessarily the correct one. Check references, dates, the relevant source documents and whether either record has already been allocated. Grouped payments need evidence connecting the individual entries to that bank transaction. If two candidates remain plausible, retain the ambiguity and ask for the missing evidence rather than choosing by row order.
Does a zero reconciliation difference prove the output is correct?
A zero difference is insufficient by itself. Offsetting errors, duplicate allocations or incomplete imports can leave a total looking satisfactory. Trace accepted relationships back to their sources, check the complete input population and inspect unresolved or adjusted items. The human-review guide explains why a convincing answer still needs a task-specific evidence check.
Should AI post an unmatched bank charge automatically?
Treat the suggested classification and entry as a proposal until the source, existing records and applicable approval requirements have been checked. Confirm what accepting a suggestion actually does in the configured tool. For an assistance-only trial, retain the proposed charge for an authorized person to review and post through the approved process; an unexplained difference is not a reason to bypass that boundary.
How can the team tell whether assistance is useful?
Compare complete reviewed work with the existing process on like cases. Include false matches, missed exceptions, corrections and the receiving accountant's effort, rather than reporting only the number of matched lines. Test whether another colleague can retrace and maintain the result. Agree what improvement would justify expansion before treating a successful demonstration as an operational change.
