Most operators we talk to about AI in accounting are working from one of two starting positions: either they believe nothing (the “it's all hype” stance), or they believe everything (the “AI will replace my bookkeeper” stance). Both are wrong, in different ways.
The honest read is that AI agents are excellent at some specific tasks and predictably bad at others. The skill is knowing which is which — and designing the workflow so the agents do what they're good at and humans do what they aren't.
This piece walks through the capability map for each kind of work that shows up in a small-business back office.
Where AI agents are genuinely useful
Pattern matching at scale
Reading a receipt and identifying the vendor, the date, the amount, and the likely category. Doing it accurately across thousands of receipts a month, in different formats, in different languages, with different layouts. This is the single thing AI is now genuinely good at.
For a single receipt, a human is faster. For 3,000 receipts a month, the human-only approach is what quietly stretches month-end into a multi-week chase. AI handles the high-volume drafting layer.
Drafting categorisation from historical patterns
Given a year of your books, an agent can learn that payments to “ABC Restaurant Supplies” usually go to the food cost account for Outlet B with standard GST. It then drafts that same coding for the next transaction from the same supplier.
Drafting is the key word. The agent doesn't post — it suggests, and a human approves. Drafts that are right 90%+ of the time turn classification from work into review.
Exception detection
Surfacing things that don't look like the historical pattern. A supplier payment 30% larger than usual. A delivery platform payout that's short of expected. A transaction in a category that hasn't had activity in six months. A duplicate payment to the same vendor within a week.
A human reviewer would catch these if they had time to comb through every transaction. They don't. The agent surfaces the candidates; the human decides what to do with each one.
Following up — chasing
Vendor invoices outstanding past their expected receipt date. Approvals pending too long. Receipts missing for transactions above a threshold. Compliance deadlines approaching. The agent watches, the agent sends the appropriate nudge to the appropriate person, on the appropriate cadence.
The work itself isn't intellectually hard — it's just easy to forget when a human owns it. The agent doesn't forget.
Draft narrative writing
End-of-week or end-of-month management summaries. Given the underlying data — revenue, cost trends, exceptions — AI is good at drafting a plain-language summary an owner can read in 90 seconds. The factual base is the data; the agent does the assembly.
The human reviewer checks the narrative for accuracy and tone before it goes to the owner. Drafting saves hours; review takes minutes.
Where AI agents predictably fail
Edge cases that look like normal cases
Refunds posted as positive transactions. Credit notes that look like invoices. Reversals. Multi-leg journal entries. Transactions where the supplier name on the bank statement doesn't match the supplier name in your books because of a name change. Foreign currency adjustments.
These are the 5-15% of transactions that look standard until you read them properly. Agents handle them incorrectly with confidence — which is more dangerous than handling them incorrectly with uncertainty. Human review catches them.
Judgment about your specific business context
The agent doesn't know that Outlet C is running a promotion that should be coded against the marketing budget instead of revenue. It doesn't know that this supplier just went into dispute and their invoices should be held. It doesn't know that this catering job was done as a favour and the revenue should be coded differently.
Business context lives with humans. The agent is fast at executing patterns; humans are the source of the patterns in the first place.
Tax and structural judgement
Whether a particular expense is deductible. Whether a transaction is GST-applicable or zero-rated or exempt. Whether a payment to a related party is at arm's length. Whether a particular reorganisation has tax implications. These are not pattern-matching problems — they're judgement problems against rules that have context and exceptions.
AI can draft a position. A qualified accountant has to decide whether to take it. The accountability sits with the human, not the model.
Anything that's never been done before
Agents work from patterns. If your business is doing something genuinely new — a transaction type that doesn't exist in your historical data, a customer relationship that doesn't fit any template — the agent is guessing. Humans are better at first-instance reasoning.
Owning the final answer
This is the most consequential limit. Agents can draft; they cannot bear responsibility. If a tax filing is wrong, the qualified accountant who signed off is responsible — not the agent. If a financial statement is materially misleading, the firm is on the hook, not the model.
A firm that lets agents post directly to client books without human sign-off has either decided to absorb that risk silently, or hasn't thought about it carefully. Either is a red flag.
How to think about the trade-off
The right mental model is not “AI vs. human” but “AI as the drafting layer, humans as the review and decision layer.” The agent handles throughput; the human handles judgement and accountability.
The structural design that makes this work:
- Agent drafts. Classification, chasing, reconciling, summarising. Actions logged, nothing committed.
- Process manager reviews. Catches edge cases, low-confidence calls, business-context issues. Either approves or sends back.
- Qualified accountant signs off. Final review on anything that posts. Carries the accountability.
Nothing posts without sign-off. That is the rule. Without it, the AI claim is dangerous; with it, the AI claim is honest.
What this means for evaluating an AI-enabled firm
Three questions to ask any firm pitching AI in their accounting workflow:
- What specifically does your AI do?Vague answers (“intelligent automation”, “AI-powered bookkeeping”) are marketing. Specific answers (“our Receipt Catcher routes documents from WhatsApp into Xero with categorisation drafted”) are real.
- Who reviews what the AI does, and when? Real firms have a named human at each review stage. Vague firms don't.
- Can the AI post directly to my books? If yes, that's a structural choice they've made — make sure you understand the implications. If no, that's the right default.
Common questions about AI agents in accounting
How accurate are AI-drafted classifications in practice?
For well-defined recurring transactions (familiar vendors, common amount ranges, standard categories), accuracy is high enough that human review is largely confirmation. For new vendors or edge-case transactions, accuracy drops materially, which is why human review is the structural protection. We track agent accuracy per category and retrain or tighten the human-review trigger where accuracy slips.
What types of business benefit most from AI agents in their accounting?
Anything high-volume: F&B operators with daily POS settlement, retail with daily card and platform settlements, services SMEs with high expense volumes across many staff. The throughput problem is what AI agents solve. A low-volume desk business with 30 transactions a month gets less obvious benefit; the agents earn their keep when the manual alternative is genuinely painful.
Will AI agents catch fraud or theft?
Exception detection can flag unusual patterns — duplicate supplier payments, unusual transaction amounts, payments to vendors that don't match historical patterns. That's useful as an anti-fraud signal but it isn't a substitute for proper internal controls. The agent surfaces the anomaly; the human investigates and decides whether it's benign, an error, or something more serious.
Can the AI agent talk to my customers or suppliers directly?
For some workflows, yes — the Vendor Chaser sends follow-up messages to suppliers about outstanding invoices, for example. The agent operates within defined templates and escalation paths. It doesn't negotiate, make commitments, or handle disputes — those remain human work. The agent chasing pattern is documented to the supplier when we onboard them, so there are no surprises about who's sending the message.
What happens if I want to stop using AI agents partway through an engagement?
The agents are layered on top of standard Xero + Aspire setup. They can be disabled without changing the underlying ledger or system structure. The work that was happening through agents reverts to traditional manual workflow. The cost saving from agents is the recurring labour they save; turning them off recovers a more manual workflow at higher ongoing cost.
AI helps. It doesn't replace. The firms that will still be standing in five years are the ones that internalised both halves of that sentence.
— The Acuitas team