AI Accounting for Networks: Revenue Reconciliation

In a network of clinics, restaurants, stores, or service locations, accounting reconciles several versions of the same revenue every day: the record or order in the industry system, the cash receipt, card acquiring, SBP, the bank statement, and the final result in 1C.

When there are only a few units, discrepancies are found manually.

When there are more than a hundred, the spreadsheet turns into a control room: the accountant checks not accounting, but the money routes between systems. In such a setup, the AI accountant is neither a conversational partner nor an autonomous replacement for a specialist.

This is an agent that pulls data from approved sources, applies deterministic rules, explains deviations, and passes only exceptions to the accountant. Posting, classification changes, and other material actions remain under human control.

In a single unit, revenue can already have five views: the rendered service, the receipt, the payment provider transaction, the bank deposit, and the accounting entry. In a chain, different legal entities, settlement delays, refunds, fees, certificate payments, advances, and payment-method selection errors are added.

Traditional automation matches rows well using exact keys.

But the real tail consists of ambiguous cases: the payment arrived the next day, the purpose was filled in as free text, the amount arrived net of fees, or the refund is tied to another shift.

Here, the agent is useful as an interpretation layer.

It does not invent an accounting solution; it gathers evidence around the transaction and offers the accountant a verifiable hypothesis.

Integrations and rules form the core of the setup: they pull data, normalize reference data, and calculate control sums. AI connects to unstructured context and ambiguous exceptions such as payment purpose, correspondence, documents, and the history of similar decisions. This separation is fundamental.

Arithmetic and allowed deviations are safer to handle deterministically.

The model must not guess VAT or execute a payment.

Its task is to find links, formulate an explanation, and show which sources it is based on.

The human-led, agent-operated approach matches the direction KPMG identifies in its 2026 finance functions study.

For a chain business, the first candidate is daily revenue reconciliation by unit and payment method.

It repeats every day, has a clear input and a verifiable outcome, and most exceptions can be built from history.

Such a pilot creates the basis for the next tasks: automatic processing of bank statements, finding those responsible for payments and documents, and controlling revenue and VAT classification.

Inventory write-offs for closed procedures are better handled in a separate stage.

It has a different accounting object, its own regulations, and its own cost of error.

The integration layer, reference data, activity log, and exception interface will remain shared, so the first reconciliation becomes not a one-off robot, but a check of the architecture of the future setup.