AI Bookkeeping: An Honest Assessment of What Works
The short answer
AI is genuinely good at four bookkeeping tasks: reading documents and extracting structured data, suggesting expense coding based on history, detecting anomalies such as duplicates and unusual amounts, and reconciling transactions against bank feeds. It is unreliable at judgement calls that depend on context it cannot see, such as whether a cost is capital or expense, whether a transaction is material, and anything requiring accountability to a regulator. The realistic model is AI doing the reading and a human doing the deciding.
The phrase AI bookkeeping covers everything from a tool that reads a PDF to a claim that a business no longer needs an accountant. The first is real and useful. The second is a marketing position, and following it is how businesses end up with tidy books that are wrong.
Here is a task-by-task assessment based on what these systems actually do rather than what they are sold as.
What AI does well
Reading documents
This is the strongest use case by a wide margin. Modern models extract vendor, dates, amounts, tax and line items from an invoice they have never seen, in a layout they have never encountered, without configuration. Ten years ago this required a template per vendor. Now it does not.
The reason it works is that invoices are a highly structured genre with enormous training data behind them. The model is not reasoning about your business, it is recognising a document type it has seen a very large number of variations of.
Suggesting expense coding
Given history, a model reliably predicts that this vendor goes to this account. Accuracy climbs quickly with a few examples per vendor and then plateaus high.
This works because most coding is repetition, not judgement. The electricity bill goes to utilities every month. Where it stops working is exactly where a human would also hesitate: a payment to a vendor you use for two different purposes, or a mixed invoice that should split.
Anomaly detection
Duplicates, amounts far outside a vendor's normal range, invoices dated in the future, tax that does not compute against the subtotal, bank details that changed since last time. These are pattern deviations, and pattern deviation is what these systems are built for.
This is arguably more valuable than the time saving, because it catches things a human doing routine entry at speed reliably misses.
Transaction matching
Matching bank feed lines to bills and invoices, including partial payments, batched payments and amounts that differ by a payment processor fee. Fuzzy matching at volume is a machine task.
What AI does badly
Capital versus expense
Whether a $3,000 laptop purchase is an expense or a capitalised asset depends on your capitalisation policy, your tax position and sometimes a deliberate planning choice. The invoice does not contain the answer, and a model that guesses is guessing about tax treatment.
Materiality
Knowing that a $40 discrepancy does not matter but a $40 discrepancy in a specific control account does is a judgement about consequences. Models have no reliable sense of what matters, so they either flag everything or nothing.
Anything requiring accountability
Signing a tax return, taking a filing position, deciding revenue recognition timing, attesting to accounts. These require a person who is professionally responsible. This is not a technical limitation that will be engineered away; it is the point of the requirement.
Context that lives outside the documents
That an invoice relates to a job you already wrote off, that a vendor is disputed, that a payment is a deposit against a contract that has not been performed. None of that is on the PDF, and a system with access only to documents cannot know it.
Being confidently wrong
This is the failure mode that matters most. A template-based system breaks loudly. A model produces a plausible answer with no signal that it is unusual. Any AI bookkeeping tool worth using surfaces confidence and flags low-confidence extractions rather than presenting everything with equal certainty.
A realistic division of labour
| Task | AI | Human |
|---|---|---|
| Read the invoice | Does it | Spot checks |
| Suggest the expense account | Proposes | Confirms, corrects the first time |
| Flag duplicates and anomalies | Detects | Decides |
| Match bank transactions | Proposes matches | Approves exceptions |
| Capital vs expense | Should not decide | Decides, per policy |
| Approve payment | Never | Always |
| Verify changed bank details | Flags the change | Verifies by phone |
| Month-end judgements and accruals | Assists with data | Decides |
| File and sign returns | No role | Accountant |
What this means for bookkeepers
The part of bookkeeping that AI genuinely replaces is data entry, which is also the part clients least want to pay for and practices least want to staff. What it does not replace is the reason clients call: knowing what the numbers mean, catching what looks wrong, and being accountable.
The practices doing well with this are the ones that let automation absorb volume growth rather than cutting staff, and shift the freed hours toward advisory work that bills at a higher rate. The ones struggling are those still charging for hours spent on work a machine now does in seconds, which is a pricing problem rather than a technology one.
How to evaluate an AI bookkeeping claim
- 1Ask what happens when it is unsure. If there is no confidence signal and no review queue, it is presenting guesses as facts.
- 2Ask what it does not do. A vendor who cannot name a limitation has not thought about failure modes or is not telling you about them.
- 3Ask where the audit trail lives. You need to be able to show an auditor the source document and who approved what.
- 4Run your own worst invoices through it. Not the clean ones.
- 5Check whether it can be wrong quietly. Silent failure is the property that turns a time saver into a liability.
Where this is going
Document reading is close to solved for clean digital documents and will keep improving on messy ones. Coding suggestions will keep getting better because they benefit from accumulated history. Anomaly detection will get more sophisticated.
What will not change soon is the accountability boundary. Someone has to be responsible for the numbers, and responsibility does not delegate to software. A tool that reads every invoice perfectly still needs a person who understands the business to decide what the resulting picture means.