AI agents for accounting and bookkeeping firms
Accounting firms use AI agents on the repetitive parts of the close: coding supplier invoices, matching bank transactions, chasing missing client documents, and drafting month-end variance commentary. The constraint that shapes every design is segregation of duties — an agent may prepare an entry, but the same system must not both prepare and approve it.
The work that scales badly
A bookkeeping practice's cost base is dominated by work that grows linearly with client count and never gets easier: coding supplier invoices to the right account, matching bank lines, and chasing clients for the receipt they never sent. Adding clients means adding people, which is why margins compress as these firms grow.
This is unusually good agent territory because the work is high volume, the correct answer is knowable, and — critically — you have years of historical coding decisions sitting in the ledger to check against. Most industries do not have that. You can measure an agent's accuracy against what your own team actually did, on your own clients, before it touches anything live.
Segregation of duties is the design constraint
The rule that shapes everything here: whatever prepares a journal entry must not also approve it. An agent that codes an invoice and posts it has collapsed a control that exists for a reason, and an auditor will say so.
Build the agent as a preparer. It codes, matches, drafts and explains its reasoning; a human reviews and posts. As confidence grows, raise the auto-post threshold for low-value, high-confidence, well-precedented transactions — but keep the threshold explicit, configurable, and logged, so the control is documented rather than implied.
Close workflows ranked by agent fit
| Workflow | Agent fit | Why |
|---|---|---|
| Supplier invoice coding | Excellent | High volume, verifiable against ledger history, cheap to correct |
| Bank reconciliation | Excellent | Deterministic matching for most lines; the agent's value is the unmatched tail |
| Client document chasing | Excellent | Pure follow-up work, no judgement, endless volume |
| Expense policy checking | Good | Policy is written down, so the agent has criteria to apply |
| Variance commentary drafting | Good | Agent drafts from the numbers; the accountant supplies the story |
| Accrual and provision judgements | Poor | Genuine professional judgement — not repetitive, and consequential |
| Final sign-off | None | A licensed human signs. Non-negotiable. |
An AP coding agent, traced
Ingests the document
Invoice arrives by email, portal or photo. The agent extracts supplier, date, totals, tax and line detail regardless of layout — no per-supplier template needed.
Identifies the supplier
Matches against the ledger including near-misses and trading-name variants, rather than creating a duplicate supplier record.
Proposes the coding
Looks at how this supplier's invoices have historically been coded for this client, and proposes the account, department and tax treatment with that precedent cited.
Checks against the PO
Where a purchase order exists, three-way matches invoice to PO to receipt and flags discrepancies rather than silently accepting them.
Flags the anomaly
Amount materially outside this supplier's normal range, duplicate invoice number, bank details differing from the supplier record — the last of these is a fraud signal and escalates immediately.
Queues for approval
A human approves the batch. Approvals and overrides both feed the precedent the agent reasons from next month.
Integration: the ledger APIs are genuinely good
This is the rare industry where integration is not the hard part. Xero, QuickBooks Online, Sage and NetSuite all expose capable REST APIs with real write support, and the document-capture layer — Dext, Hubdoc, Bill.com — is API-first by construction.
Which means the engineering effort lands where it should: on the coding logic, the anomaly detection, and the approval workflow, rather than on fighting a system of record for permission to write to it.