Almost every accounting firm now claims AI in some form. Most of it is marketing veneer — bank-feed rules from 2017 re-branded, OCR receipt capture re-described as “intelligent classification”, and a chat interface bolted on top of whatever was there before.
That doesn't mean AI is useless in small-business accounting. The opposite — there's real, specific, measurable lift available. It just isn't the marketing version.
This piece walks through what AI actually does well in a small-business back office, where it still struggles, and the human-review architecture that has to sit around it to keep the books trustworthy.
Where AI earns its place
AI agents are good at a specific shape of work: repetitive, pattern-matching, high-volume, with well-defined inputs and outputs. The accounting workflow has lots of work that fits that shape.
Receipt and document routing
Receipts arrive across channels — WhatsApp, email, camera, PDF. The job: read the document, identify the vendor, the date, the amount, the GST treatment, route it to the right account in the ledger, attach it to the matching transaction. Drafting that classification is a near-perfect use case for AI.
What the human still does: review low-confidence matches, handle the edge cases (ambiguous vendor names, missing GST identification, refunds, credit notes).
Vendor chasing
Outstanding supplier invoices, pending approvals, missing delivery notes. The agent watches for things that should have arrived and haven't, then sends the chase message to the right person. Not the bookkeeper forwarding emails — direct outreach from the system, with context, on a schedule.
Monthly close checklists
Close is a sequence of dozens of reconciliations and checks. An AI agent owns the checklist — knows what's done, what's blocked, what's waiting on a human, surfaces the bottleneck. The work doesn't change. What changes is the visibility of where the work is, and the discipline of the cadence.
Exception detection
Anomalies in transactions, drifts in cost ratios, duplicate supplier payments, miscoded entries that don't match historical patterns. The agent flags them for human review before they reach the close. This is the work that's closest to genuinely AI-native — the agent can spot things a human reviewer would have to comb through hundreds of transactions to catch.
Categorisation drafting
Based on historical patterns, the agent drafts the right account, the right outlet (via tracking categories), and the right tax code for new transactions. Drafts only. The human reviewer approves or corrects, and the corrections feed back into the model.
Management summary writing
End of week and end of month, an agent produces a plain-language narrative summary of what changed — sales by outlet, cash position, cost trends, exceptions worth knowing. Drafts the prose. The owner gets a 90-second read instead of a 30-minute report assembly.
Where AI still struggles
Equally important: the work AI is not good at, and where a small business relying on AI alone gets in trouble.
Edge cases and ambiguity
The 80% of transactions that look like every other transaction is where AI shines. The 20% that don't — credit notes, partial refunds, ambiguous coding, multi-leg journal entries — is where it overreaches. Without human review, the books drift, and the drift compounds.
Judgment about your specific business
AI doesn't know that your Outlet C is currently running a promotion that should be coded against the marketing budget instead of normal revenue. It doesn't know that this supplier just went into dispute and their invoices should be held pending resolution. Business context lives with humans.
Tax positions and structural decisions
AI agents can prepare draft tax computations. They cannot decide whether a deduction is reasonable for your specific structure, or whether a particular accounting treatment would survive an IRAS query. That's a qualified accountant's call.
Trust at the seam
Maybe the most consequential limit. An AI agent that posts entries directly to a client's ledger, without human review, eventually makes mistakes that cost the client real money. The only honest way to run AI on accounting data is to design it as a drafting layer, not a posting layer.
The human-review architecture
The pattern we've landed on at Acuitas: every entry that touches a client's ledger passes through two humans after the agent drafts it.
- The AI agent drafts. Classifying, chasing, reconciling, summarising. Actions are logged. Nothing is committed.
- A process manager reviews.Catches edge cases, the agent's low-confidence calls, the things that need business context. Either approves or sends back.
- A qualified accountant signs off. Final review on anything that posts. Carries the accountability for the work.
Nothing posts to your books without sign-off. This is what makes the AI claim defensible — and it's the single piece most “AI accounting” pitches quietly skip.
What this looks like, in practice, on an Acuitas engagement
We don't sell “AI-powered accounting”. We configure named agents on specific jobs:
- The Receipt Catcher — routes documents from WhatsApp, email, photo into Xero.
- The Vendor Chaser — follows up on missing invoices and approvals.
- The Close Captain — runs the monthly close checklist.
- The Outlier — flags anomalies before they reach the close.
- The Categoriser — drafts expense classification from historical patterns.
- The Compass — drafts the weekly management summary.
Each one has a named reviewer and a visible audit log. The full breakdown lives at /what-we-do/ai-automation-workflows.
What we'd expect to change in the next 24 months
The honest forecast: more of the drafting layer becomes AI-native. Receipts will be classified more accurately. Exception detection will catch more subtle anomalies. Summary writing will improve. The human-review layer will become more about exception triage than line-by-line review.
What won't change: the accountability for the books stays with a qualified human. The structural decisions about how a business is structured stay with humans. And the trust at the seam — “nothing posts without sign-off” — stays in place. Anyone who tells you otherwise is selling you something.
Common questions about AI in accounting workflows
Will AI replace my bookkeeper?
No, but it changes what the bookkeeper does. The rote data-entry layer becomes drafted automatically; the human role shifts to review, exception handling, and judgement calls. For most small businesses, the outcome is the same headcount producing more reliable books on a shorter cycle, not fewer staff.
Where is the data going? Is it secure?
For Acuitas clients, client data lives in Xero, Aspire, and a small number of connected platforms, each with its own security posture. AI agents we run operate on the data within those systems, with role-based access. Nothing is transmitted to consumer AI services in a way that exposes client data outside their own workflow. See our security & data handling page for the architecture in detail.
What happens when the AI gets it wrong?
The human reviewer catches it before posting. The agent's draft is exactly that — a draft. The qualified accountant signs off on what posts. If the agent's error rate on a particular task becomes problematic, we either retrain on better examples, tighten the human-review trigger, or stop using the agent for that task. Nothing about the architecture assumes the agent is infallible.
Can my existing bookkeeping setup use AI agents, or do I have to switch firms?
Many of the patterns above can be added to an existing Xero setup without a firm switch. Receipt capture tools like Hubdoc or Dext are off-the-shelf additions. Aspire integrates with most Xero setups. Where firm switching matters is on the deeper workflow design — named agents on specific jobs, structured human review, exception detection — which usually require a firm that's designed itself around running AI agents on accounting data, not just bolting on tools.
How do I evaluate whether a firm's AI claims are real?
Three questions: what specifically does the AI do (vague answers like “intelligent automation” are marketing); who reviews what the AI does and when (real firms have a named human at each review stage); can the AI post directly to your books (if yes, understand the risk; if no, that's the right default). The pattern of named agents with named reviewers, like the structure on our AI automation page, is the operational answer.
AI helps. It doesn't replace. The firms that get this right will be the ones whose claims survive a sales call.
— The Acuitas team