Accounting software already imports bank feeds, proposes matches and applies rules. The useful AI opportunity is often the unresolved remainder: missing evidence, unusual descriptions, conflicting records and review notes spread across systems.
This guide proposes a workflow for accounting and bookkeeping firms. It does not replace the firm’s accounting policies, professional standards or accountant review, and it does not claim Binarify client results. The firm must define permitted data, materiality, approval and posting rights for each engagement.
Target the exception queue
A broad promise to “automate bookkeeping” hides several distinct decisions. A safer first outcome is:
Prepare a sourced exception queue for one recurring close process, with deterministic calculations reconciled, suggested categories separated from approved postings, and every unresolved item assigned to a client or accountant.
Use the native product first. QuickBooks describes reconciliation as comparing accounting records with bank or card statements and reaching a zero difference. Its current workflow can also upload a statement for AI-powered reconciliation. A custom layer should address an evidenced gap around the platform, not duplicate matching that the subscribed product already performs.
A controlled exception workflow
1. Freeze the review population
Define the client, entity, ledger, account, period, source systems and cut-off time. Keep the imported transaction ID, original description, amount, currency, date and bank-feed state.
Take a snapshot or use supported period controls so the population cannot change invisibly while it is being reviewed. Late entries and reopened periods need their own logged route.
2. Run deterministic integrity checks
Use code for totals, duplicate identifiers, date windows, balance roll-forwards and reconciliation arithmetic. Compare opening and closing balances with the authorised statement or source.
Do not ask a language model to decide whether a ledger balances. It may explain a prepared exception, but displayed figures must reconcile to source records.
3. Retrieve the supporting evidence
Link invoices, receipts, statements, prior treatment, supplier or customer records and relevant client explanations. Keep evidence provenance and access rights.
An exception may be caused by missing evidence, duplicate import, timing difference, changed historic transaction, new counterparty, unclear business purpose, tax-treatment question or an incorrect opening balance. Keep these possibilities separate until reviewed.
4. Propose a category and reason
The model can propose an existing chart-of-accounts category, tax code or review reason from the approved set. Show the evidence and comparable approved examples that influenced the proposal where the firm permits their use.
Record confidence, model version and alternatives. Low confidence, material amounts, new suppliers, related parties, cross-border items, unusual tax treatment and policy conflicts should route to an accountant.
Do not create a new account, journal or tax code simply because the closest existing option looks imperfect.
5. Separate client questions from accountant questions
Prepare a plain-language client request when a business fact or document is missing. Route accounting-policy, materiality, recognition, tax and prior-period questions to the appropriate professional.
The workflow should avoid asking the client to choose an accounting category. Ask for the underlying fact: what was purchased, why, for whom, when it was delivered and where the supporting document can be found.
6. Put the reviewer in control
The reviewer workspace should show:
- original transaction and source state;
- proposed match or category;
- supporting evidence and client explanation;
- deterministic checks and reconciliation effect;
- previous approved treatment where relevant;
- materiality or escalation rule;
- unresolved question and owner; and
- a complete action history.
Allow approve, correct, request evidence, defer and escalate. Bulk approval should be restricted to a defined, tested pattern and still preserve item-level evidence.
7. Write only approved outcomes
Use the accounting platform’s supported API or import. Apply least-privilege permissions and idempotency controls. The integration may create a draft transaction, attach evidence or add a review note. Posting, journals, period locks and filing-related states should follow the firm’s approved authority model.
Reconcile failed writes and display them before the close is marked complete. Retain the model, rule and human action that produced each posted change.
8. Prepare the close review brief
Summarise the completed population, outstanding exceptions, material movements, reconciliations, late changes and client dependencies. Link each statement to the ledger or evidence.
The brief supports the accountant’s review. It should not state that the accounts are correct, complete or compliant before the responsible professional concludes that work.
Binarify close-exception ledger
This proposed artifact gives every unresolved item an evidence trail and owner.
| Exception | Deterministic evidence | AI may propose | Human decision | Close condition |
|---|---|---|---|---|
| Unmatched bank item | Feed ID, amount, date, candidates | Likely match and reason | Confirm, reject or investigate | Approved match or owned exception |
| Missing document | Transaction and request history | Exact evidence request | Adequacy and treatment | Evidence accepted or documented alternative |
| Changed reconciled item | Audit log and prior report | Change summary | Correction and period action | Reconciliation restored and reviewed |
| New or ambiguous spend | Description, supplier and evidence | Existing category candidates | Account and tax treatment | Approved posting with source |
| Balance variance | Source and ledger calculation | Narrative explanation of prepared figures | Cause, correction and sign-off | Difference resolved or formally carried |
Preserve professional judgement
The ACCA’s AI risk guidance states that professional judgement, objectivity and competence should not be undermined by undue reliance on technology. The ICAEW’s 2026 summary of AI and accountancy guidance similarly explains that AI can support work within existing professional frameworks while judgement and confidentiality remain with the accountant.
Translate those principles into operational controls: approved tools, client-data boundaries, independent source checking, review thresholds, staff training, incident escalation and a record of how AI affected the file.
Measure exception quality and close performance
Compare similar clients, periods and account types. Track:
- exceptions per thousand transactions;
- accountant minutes per exception and per completed close;
- age of unresolved items and client-dependent items;
- first-review acceptance and correction rate;
- incorrect matches, categories and tax-code suggestions;
- material items missed by the workflow;
- post-close corrections and reopened reconciliations;
- failed or duplicate system writes; and
- performance across supported document, language and transaction types.
High acceptance is not sufficient. Review a sample of accepted suggestions for automation bias, especially when the same treatment has been copied from historical data.
Pilot one ledger and one period
Select a recurring bookkeeping service, a small number of accounts and one reviewer team. Run the proposed queue beside the current process through at least one complete close.
Agree sources, field mappings, posting permissions, materiality, review actions, test cases, quality thresholds, outage procedures and rollback before assisted use. Include missing documents, duplicate imports, changed reconciled transactions, unusual but legitimate purchases and deliberate source conflicts.
Begin earlier in the workflow with AI client document collection for accounting firms. Use the accounting-firm AI ROI guide to calculate the value of changed capacity and rework. Explore our custom AI integration approach or book a 30-minute conversation about one exception queue.