The promise of AI in small-business accounting is real — repetitive admin handled at machine speed, exceptions flagged proactively, summaries drafted on schedule. The cost is the loss of human visibility if the automation is set up without proper controls. Books drift. Mistakes compound. By the time someone notices, recovery is expensive.
The fix isn't to avoid AI. It's to design the governance around it from the start. What follows is the pattern we use for Acuitas clients — the controls that keep AI agents useful and the owner in command.
The five-part governance pattern
1. Define what each agent does — narrowly
Each AI agent should have a specific, bounded job. Not “handles accounting” — “reads incoming receipts from email, extracts vendor / date / amount / GST, drafts the classification against historical patterns, flags low-confidence matches for human review.”
Boundaries make agents auditable. A vague agent is impossible to verify; a narrow agent is testable.
2. Name the reviewer for each agent
Every agent draft has a named human reviewer. For some agents (Categoriser), the reviewer is a process manager who handles routine review. For others (Close Captain), the reviewer is the qualified accountant whose sign-off carries accountability.
The pattern: agent prepares, named human reviews, nothing posts without sign-off.
3. Log everything the agent does
Every action logged: what the agent saw, what it drafted, what confidence score it assigned, who reviewed it, what changes were made on review.
The audit log is the governance backbone. Without it, when something goes wrong, you can't trace the failure. With it, you can.
4. Set tolerance for autonomous action
Some categories of work are low-risk enough to autonomously post (recurring supplier payments that match historical patterns exactly, for example). Others should always require human approval (anything unusual, anything material, anything new).
The tolerance line is set deliberately, documented, and reviewed quarterly. The default is “require human approval”; auto-post is the exception that has to be justified.
5. Build recovery into the design
When the agent gets it wrong (and over time, it will), the recovery path is part of the design:
- How is the error detected?
- What's reversed automatically vs manually?
- How is the agent retrained or adjusted?
- How is the operator notified?
A workflow without a recovery path is fragile by design. A workflow with one is robust.
What to automate vs what to keep human
Good candidates for AI agents
- Pattern-matching at scale. Receipt classification, recurring supplier coding, bank reconciliation rules.
- Following up. Vendor chasing for outstanding invoices, missing receipts, expired approvals.
- Exception detection. Unusual transactions, deviation from historical patterns, duplicate payments.
- Routine summarisation. Weekly management summary, exception feed digest.
- Checklist tracking. Monthly close, compliance calendar, audit prep.
Should stay human
- Tax positions and structural decisions. Whether a deduction is allowable, whether a structure works.
- Edge cases requiring business context. Why this supplier should be held in dispute, why this promotional cost should be coded differently this quarter.
- Final sign-off on anything that posts. The accountability for the books stays with a qualified human.
- Client communication and relationship work. Operators want to talk to a person, not be chased by an agent that sounds increasingly irritated.
- Strategic and advisory decisions. The numbers inform the decision; humans make the decision.
The recovery pattern
When an agent gets something wrong, the recovery discipline is:
- Stop the drift. Pause the agent on the workflow where the error happened until the cause is understood.
- Identify the scope.How many transactions are affected? What's the cumulative impact?
- Reverse what needs reversing. Journal corrections, communication with any affected counterparties, IRAS notification if material.
- Retrain or adjust. Why did the agent err? Bad training pattern, new transaction type, system change? Address the underlying cause.
- Resume with tightened controls. Lower the tolerance for autonomous action on the affected workflow until accuracy is back to expectation.
Each step is documented. The recovery becomes part of the audit log, demonstrating the governance is live, not just on paper.
What “losing control” looks like — and how to avoid it
Four patterns we've seen go wrong:
- Auto-posting without human review. Fast and looks efficient — until the agent classifies wrongly with confidence and the books drift. Don't auto-post.
- Single point of failure on the AI vendor. If your entire workflow depends on one vendor's AI service, vendor changes break everything. Diversify or own the orchestration layer.
- No human still understands the workflow. Six months in, the only person who knew how the agents were configured leaves. Nobody can audit or change anything. Document as you go.
- Agents asked to make business decisions. Categorising a receipt is an agent job. Deciding whether to absorb a supplier price increase is a human job. Don't cross the line.
Common questions about AI governance for SMEs
Do I need a formal AI policy as a small business?
Not in the corporate-governance sense. But you do need clarity on what your AI agents do, who reviews their work, and how errors are handled. A one-page document covering those three things is the practical equivalent for an SME.
What if my accounting firm uses AI but won't explain how?
That's a flag. A firm using AI on your data should be able to explain what specifically it does, who reviews its work, and what happens when it errs. Opacity here is the warning sign — the firm either doesn't understand its own systems well enough or is hiding something. Either way, push for transparency.
How do I know if AI is actually safe to use on financial data?
The same way you know any tool is safe: governance and audit. AI itself isn't the risk; AI without controls is. Confirm the controls — review gates, audit logs, recovery paths — and the safety follows.
What happens to my data when an AI agent processes it?
Depends on the architecture. For Acuitas clients, client data stays within the operating platforms (Xero, Aspire) and the agent layer operates on it within those platforms' security perimeter. We don't send client data to consumer AI services. See our security & data handling page for detail.
Can I roll back AI automation if I don't like how it's working?
Yes. The agents are layered on top of standard Xero / Aspire setup. They can be paused or disabled without changing the underlying ledger. The workflow reverts to manual; recovery is graceful. This optionality is part of the governance — the operator is never locked in.
AI in small-business accounting works when it's governed properly. The cost of governance is small relative to the gain in throughput; the cost of skipping governance is everything. The pattern above is what keeps the gain without the cost.
— The Acuitas team