"Can AI do bookkeeping for a small business?" has a short honest answer: parts of it, yes, and those parts are worth having. The other parts, no, and the line between them is not about how clever the software is. It is about who answers when the number is wrong. This article is for the owner or CEO of a company with somewhere between twenty and three hundred people who typed that question, or "should I use AI for my company's books" or "AI bookkeeping risks", because a controller is asking for a tool, a vendor is offering one, or you suspect your own team is already pasting bank exports into a chat window.
I was COO and then President of a public company that closed more than forty acquisitions, and I ran the integration on most of them, which meant I spent a great many hours inside other people's books. I am a lawyer by training. Now I write fixed-fee AI plans for CEOs of private-equity-backed and owner-led mid-market companies through 013 Labs. Nothing here is accounting, tax or legal advice. Where judgement belongs, I say "your accountant", and I mean it.
The test: who answers when the number is wrong
Books are not a document. They are a record that other people rely on. Your lender sets covenants against it, your tax authority can ask for the support behind any line, and a buyer will pay a multiple of it. So the useful way to sort bookkeeping work is not "easy versus hard". It is this: if this entry is wrong, who has to explain it to someone outside the company?
If the answer is "nobody, someone will catch it in review", AI can probably help. If the answer is "a named person has to stand behind it", AI can prepare the work, but the person has to do the standing. That one test sorts nearly everything below.
What AI can safely do in the books today
These are the jobs where software that reads and drafts earns its place. In every one of them, the output is a suggestion that a person with the right access reviews before it becomes part of the record.
Categorising transactions. Reading a bank or card feed and proposing which account each line belongs to is repetitive pattern work. A suggestion that is right most of the time and flagged when unsure saves hours. The control is that suggestions wait for approval, or are sampled, and a person looks at the unusual ones.
Matching receipts to transactions. Pairing a photographed receipt or an emailed invoice with the card line it belongs to is tedious and well suited to the task. An unmatched item should stay visibly unmatched, not be forced to fit.
Drafting reconciliations. A first pass at lining up the bank statement against the ledger, with the differences listed, is a fair job for AI. The person who owns the reconciliation still signs it, and "the software said it ties" is not a sign-off.
Flagging anomalies. A duplicate invoice, a vendor paid twice, a round-number expense just under an approval limit, a cost that doubled without explanation. AI is good at saying "look here". Treat a flag as a question, and the absence of a flag as meaning nothing.
Chasing missing documents. Drafting the polite note to a manager who has not supplied a receipt, or to a vendor whose statement is missing, is low-risk. A person decides what goes out to outsiders under the company's name.
Notice what these have in common. Each one prepares work for a person. None of them moves money or closes a period, and none of them changes who the company's records say owes what. Reading the finished statements afterwards is a separate job, covered in can AI analyze my financial statements.
What must stay with a named person
This is the part owners are most often surprised by, because it is the part that sounds like "bookkeeping".
Closing the month. Closing is a decision, not a task. Someone attests that the period is complete, the accruals are reasonable and the cut-off is right. That person has a name, and the name is a human being whom the lender, the auditor and you can call. AI can assemble the checklist and draft the entries. It cannot be the one who closed.
Anything a lender or tax authority relies on. Compliance certificates, borrowing-base reports, covenant calculations, sales tax or payroll filings, the figures behind a tax return. If an outsider will rely on it, a named person reviews every number that feeds it, and your accountant decides the treatment where there is judgement. Do not let a draft from a tool travel to a lender or a tax filing without that review.
Bank-detail changes. An email arrives from a vendor: our bank account has changed, please pay the new one. This is a common route to a large loss, and the control is a human process, not software. The change must be confirmed by calling a number you already held, not one in the email, and by someone other than the person who pays. AI may help you spot that the details differ from the last invoice. It must never be the thing that accepts the change.
Payments. Approving and releasing money stays with named people with limits, and with two sets of hands above a threshold you set. Software that drafts a payment run is fine. Authority to release one should be a deliberate written decision, not something a switched-on feature gave away. Chasing customers who owe you is a different job, covered in can AI help collect late payments. Paying the bills you owe is covered in can AI do accounts payable.
How to try it without breaking the audit trail
An audit trail is the answer to "who did this, when, on what support, and who approved it". It is what your accountant, your lender and a buyer's diligence team look for first. AI does not break it by being in the room. It breaks it when work happens somewhere the record cannot see. Five habits keep it intact.
One: start with a copy, then a side-by-side. For the first month, run the tool beside the existing process instead of replacing it. You learn where it is strong and where it is confidently wrong before anything depends on it.
Two: keep a human approval step on anything that posts. Suggestions are suggestions until a named person approves them, and the approval is recorded with a date. If the tool can post directly to the ledger, find the setting that stops that, and turn it on.
Three: use company accounts, not personal ones. The cheapest way to lose the trail is a staff member pasting bank data into a personal free account. You then have no log and no way to show anyone what happened. The article on whether ChatGPT is safe for business use sets out the plan, terms and data questions to answer first, and the piece on how to approve an AI tool for your company is the one-page record.
Four: keep the support next to the entry. If the tool cannot show the document or feed line behind a figure, treat the figure as unsupported.
Five: write a dated note. This tool was switched on for these accounts, with these permissions, approved by this person. Months later, that note is the difference between a clean answer and a long afternoon.
If a vendor tells you what its product can or cannot do, the sales page is a claim and the written terms and the product's own documentation are the evidence. The article on how to judge an AI vendor's claims helps with the sales language. Whatever you rely on, write the date next to it, because products change often.
What your accountant will ask
Your outside accountant or CPA is the first person to feel a change in how the books are produced, and the best thing you can do is tell them before they find out. Expect these questions, and have answers.
What is the tool used for, and which accounts does it touch? Who reviews its output, and is the review recorded? Can it post to the ledger without approval? Where does the company's data go, and under what written terms? How do you know a suggestion was right, and how often has someone corrected it? What changed in the month it was switched on, and is there a note saying so? And who is the named person responsible for the close?
Ask your accountant what they would want to see. That costs an hour and removes a surprise.
The questions run both ways; for the ones to bring to your accountant, here is what to ask your accountant about AI.
What your lender will ask
A lender cares whether the numbers in the certificate are reliable and whether anything changed in how they are produced. Expect these. Did the process for preparing the financials change? Who reviews the figures before they come to us? Has anything been outsourced, and to whom? Has your accountant seen the change? Does anyone other than the preparer review the covenant calculation?
Check your credit agreement, with your own counsel, for any duty to give notice before you change how reporting is produced. If you cannot answer in a few sentences, the change is not yet documented well enough to defend, and the fix is to document it.
More on that conversation: what to tell your bank about AI.
Picture the week it goes wrong
Picture a services company of one hundred and ten people. The controller, short-staffed before a quarter end, connects a tool to the card feed and lets it categorise and post. It is fast and mostly right. Nobody turns on an approval step because it seemed like friction. A vendor email arrives with new bank details, a junior clerk updates the vendor record, and the next payment run sends a large invoice to the new account. Two weeks later the real vendor chases, just as the lender's quarterly request lands and the accountant asks why a month of entries has no approval history. The fix: the owner pauses direct posting, restores a named approver for every entry, writes a one-sentence rule that bank-detail changes are confirmed by a call to a number already on file by someone who does not release payments, asks the bank about recovery, tells the accountant and the lender what happened and what changed, and records it all in a dated note. It takes the better part of a month. This is a hypothesis, a composite and not a client.
What a sensible owner does this week
None of this needs a project team. Six steps, in order.
Find out what is already happening, by asking, and make clear nobody is in trouble for answering.
Name the person who owns the close and the question of AI in the books, with a half-page written scope of what they may approve on their own.
Write the sorting test on one page: what the tool may prepare, what a person must approve, and the short list that is never delegated, which is closing, lender and tax figures, bank-detail changes and payments.
Choose one narrow job, such as categorising a single card account, and run it side by side for a month with approval switched on.
Tell your accountant before you start, and ask what they would want to see. If a lender relies on your reporting, check what your agreement says about changes to how it is produced.
Set a review date. The article on an AI policy for employees that fits on one page shows where the rule goes.
Mistakes that repeat
Letting the tool post without approval. Treating no flag as clean. Personal accounts. Skipping the accountant. Delegating the payment step. Believing a reconciliation that "ties". No dated note of what changed.
Frequently asked questions
Can AI do bookkeeping for a small business?
Parts of it. It can categorise transactions, match receipts, draft reconciliations, flag oddities and chase missing documents, as long as a named person reviews what it prepares. It should not close the month, produce figures a lender or tax authority relies on without review, accept bank-detail changes or release payments.
Should I use AI for my company's books?
For preparation, with a person approving, yes. Start narrow, run it beside your existing process for a month and tell your accountant first.
What are the main AI bookkeeping risks?
Confident errors nobody reviews, entries that post without an approval record, company data in accounts the company does not control, payment fraud through a changed bank detail, and a change in reporting that your lender or accountant was never told about.
What will my accountant or lender ask me?
Which accounts the tool touches, who reviews its output, whether it can post without approval, where the data goes and who is responsible for the close.
Not sure where you stand? Take the 3-minute AI readiness scorecard.