AUTOMATION
Automate the bank match. Keep the explanation.
A reconciliation should explain the difference between two records. A high match rate is useful only when the matches are correct and the remainder is understood.
By CA Tony Thomas · Chartered Accountant · 4 minute read
Prepared
PRACTICAL PERSPECTIVES / FINANCE · AUTOMATION · AI
Start with complete and comparable records
Bank reconciliation is an attractive automation candidate because many transactions follow recognisable patterns. The difficult part is deciding which patterns justify a match, confirming that both populations are complete and preserving an explanation for what remains.
Begin with a dated bank statement and the corresponding ledger extract. Confirm account identity, currency, period and opening balances. Record file versions, row counts and totals. A missing statement page or an incomplete ledger export cannot be repaired by a more sophisticated matching algorithm.
Keep statement dates, transaction dates and ledger posting dates as separate fields. Document which date is used for each comparison. Treat a date window as an approved tolerance requiring context, not permission to connect unrelated transactions of similar value.
A match needs more than an equal amount
In a fictional reconciliation, the ledger records a customer receipt of AED 100,000. The bank shows AED 99,950, with AED 50 separately identified as a bank charge. If the remittance reference, payer details and fee evidence support the relationship, a reviewer may be able to explain the net settlement.
The AED 50 difference should not disappear inside a broad matching tolerance. Establish whether the bank charge was already recorded and whether a journal is needed under the normal approval process. A matching rule should identify the supported components rather than silently write off the difference.
Now imagine two ledger receipts of AED 50,000 and one bank receipt of AED 100,000. A many-to-one match may be valid, but only if the references and customer settlement evidence support it. A subset of numbers that adds to the bank value is a candidate, not proof.
Keep competing candidates visible. The workflow should abstain when evidence cannot distinguish between them.
Use a transparent matching hierarchy
Process high-quality matches first: unique transaction reference, correct currency, exact amount and compatible dates. Next consider approved combinations such as a documented batch settlement. Leave ambiguous amount-only matches for review.
For each match, retain the rule used, source identifiers, relevant tolerances and whether the result was automatic or approved by a person. Do not let a later rule reuse a transaction already matched elsewhere.
Distinguish reversals, refunds, chargebacks, inter-account transfers and bank fees. They may look similar in an extract but have different explanations and evidence requirements. Transfers need both sides checked so a receipt is not mistaken for customer cash.
If an AI component reads remittance text, use it to suggest candidate references or summarise evidence. Validate extracted fields against the source and keep uncertain interpretations out of automatic posting paths. Restrict access to the bank and customer data involved.
Make reruns safe
A reconciliation process may run more than once because a statement was corrected or the ledger changed. Use stable source identifiers and a run record so the same imported transaction is not duplicated and a previously approved match is not silently overwritten.
If the bank supplies no reliable unique identifier, document an approved deduplication method and its limitations. A combination of date, value and narrative can collide for legitimate recurring payments. Preserve the original records needed to investigate those collisions.
Separate importing, proposing matches, approving exceptions and recording journals. A technical rerun must not cause a second journal or a second payment. Where the process triggers a downstream action, use an explicit action register and check whether the action has already completed.
When a correction affects a reconciled period, create a controlled reopening process. Keep the prior approved result and the reason for change.
Make unresolved items actionable
Classify exceptions by their next step: timing difference, missing ledger entry, unidentified receipt, disputed charge, suspected duplicate or missing evidence. Give every material item an owner, age, evidence link and due date.
A timing difference should have an expected clearing event. If an outstanding payment persists beyond that expectation, investigate rather than carrying it forward indefinitely. An unidentified receipt needs allocation work, not a balancing entry to a convenient account.
Reconcile the final unmatched population back to the overall difference. The exception register, approved adjustments and updated ledger should tell the same story. Do not close the reconciliation merely because the displayed difference is zero after an unexplained adjustment.
Measure the right result
Track confirmed incorrect matches, recurring exceptions, reviewer effort, aging and completeness of evidence alongside the automatic match rate. Sample automatically matched transactions and deliberately test valid transactions with identical amounts.
Run the first version in parallel with the established reconciliation. Include fees, reversals, split settlements and duplicate statement imports in the test population. Set stop rules for unexplained source-total differences and unauthorised changes.
A useful first deliverable is one account reconciled from controlled source files, with a reproducible match register and a clear exception queue. Expand only when another reviewer can follow the explanation without relying on the automation's confidence.
Personal educational perspective, independent of my employer. All business examples and figures are fictional and illustrative. Proposed workflows do not claim completed deployments or measured productivity gains. This is not entity-specific accounting, tax, legal or investment advice. Apply professional judgment, the applicable reporting framework and your organisation’s approval and data-protection policies.