Can AI Do Our Bookkeeping? What It Handles and What It Gets Wrong

Can AI do our bookkeeping?

AI can do most of the keying, matching and coding in your bookkeeping. It cannot take responsibility for the numbers.

Split the function into four layers and the answer gets precise. Capture, meaning getting documents and transactions into the system with their data read correctly. Classification, meaning coding each transaction to the right account. Reconciliation, meaning agreeing the books to the bank and to the subledgers. And judgment, meaning accruals, cut-off decisions, unusual transactions, tax treatment and the sign-off that says these books are right.

The first two layers are where AI is strongest, the third works with review, and the fourth is not a software problem. The practical shape of a working system is that the software drafts and a person approves, with the approval recorded.

Adoption has already outrun measurement in this profession, which is worth knowing before you benchmark yourself against peers. Financial Cents surveyed 486 accounting and bookkeeping professionals across North America for its 2026 State of AI in Accounting and Bookkeeping report and found 95% already using or exploring AI, while only about one in five could point to a clear, measurable return and 90% had no written AI policy. Inside Public Accounting covered the same gap in September 2026.

So the honest answer to the question as asked: AI can take the part of bookkeeping a careful junior would do, at higher speed and with a different error profile, and the part that makes a bookkeeper worth paying stays exactly where it is.

Which bookkeeping tasks can AI actually handle today?

The bookkeeping tasks AI can handle today sort into a short list, and the difference between them is how much judgment the task carries and what it costs to be wrong.

Task	Can AI run it today	What a person still does
Document capture: bills, receipts, statements, remittances	Yes, proven	Checks flagged low-confidence reads, handles unreadable documents
Transaction coding and categorisation	Yes, with review	Reviews low-confidence items, sets the rules for recurring exceptions
Invoice matching to purchase orders and receipts	Yes, proven for standard three-way matches	Handles partial deliveries, disputes, price variances
Bank and card reconciliation	First pass, yes	Clears the unmatched items, which is where the real work was anyway
Chasing missing documents and approvals	Yes, proven	Nothing, until somebody does not respond
Expense policy checks	Yes, with review	Decides the edge cases and the conversations
Accruals, prepayments, period cut-off	Drafts, needs a reviewer	Owns the decision
Tax treatment and filings	No	Everything
Sign-off on the books	No	Everything

Two things on this table matter more than the rows. The first is that accuracy improves most where the work is repetitive and the volume is high, which is why capture and coding pay first. Ardent Partners' Accounts Payable Metrics That Matter in 2025 puts the average cost of processing one invoice at $9.40 against $2.78 for best in class, and average processing time at 9.2 days against 3.1 days; its separate State of ePayables 2025 research counts labor, overhead and technology together and lands at $10.89 an invoice fully loaded. The gap between average and best in class is the size of the opportunity in accounts payable alone, and the detail of that build is in how to automate invoice processing with AI.

The second is that "with review" is a design decision and not a disclaimer. A system that drafts 400 codings and surfaces the 30 it is unsure about is doing the job. A system that drafts 400 codings and asks you to check all 400 has moved the work rather than done it.

What does AI still get wrong in the books?

What AI gets wrong in the books falls into three failure modes, and one of them is about your people rather than the software.

Wrong confidently. Models produce a plausible coding for a transaction they have no basis to code, and plausible is the problem, because an obviously wrong answer gets caught and a reasonable wrong answer gets posted. Jung Ho Choi of Stanford and Chloe Xie of MIT Sloan found both sides of this in Human + AI in Accounting: Early Evidence from the Field, published in the Journal of Accounting Research in June 2026, which combines a survey of 277 accountants, platform records for 79 small and mid-sized clients, and a framed field experiment with 99 accountants. In the experiment, AI assistance raised transaction classification accuracy by about 18 percentage points. Experienced accountants intervened selectively where the system's confidence was low, which is the behaviour you want. The same experiment also found users following AI suggestions that deviated from what they would have chosen on their own, and that is the pathway to a wrong number getting posted with a signature on it.

Rules it was told to follow. Work with one correct answer defined by a rulebook is not automatically work a model gets right. TaxCalcBench, built from Column Tax's own test cases and published on 22 July 2025, gave frontier models complete 2024 federal tax returns to compute. The best models got fewer than a third right even on a simplified sample set, and the study reports them consistently misusing the tax tables, miscalculating and misjudging eligibility. Keep tax out of scope for automation and the rest of the system gets easier to trust.

Anything that depends on context nobody wrote down. A payment that is a deposit on a job rather than revenue, an invoice that belongs in last month because the goods shipped then, an owner's transaction that is not a business expense. Those are coded correctly by a person who knows the business and incorrectly by anything that only sees the transaction.

The design answer to all three is the same. Route by confidence, sample the output against what a person would have done, and track a rework rate from the start. Our general approach to that is in what happens when AI gets it wrong.

Does AI replace our bookkeeper, or change what they do?

AI changes what a bookkeeper does rather than replacing them, and the change is towards the work that was getting squeezed out.

A bookkeeper's week in most small and mid-sized businesses is dominated by handling: chasing documents, keying, matching, fixing what came in wrong, assembling the pack. The analysis, the cash forecast, the conversation about the margin on a particular client, the reason the payroll number moved: those happen if there is time left. Automating the handling changes the ratio.

For a practice, the same shift decides what you can take on. A bookkeeping firm whose staff spend most of their week on capture and coding can take on more clients per person once that work is drafted rather than done, and that is what decides how many engagements the same team can carry.

Two claims worth retiring while we are here. AI is not about to file your accounts unsupervised, because the profession's accountability model does not permit it and the error profile does not support it. And the gain is not evenly spread across staff: assistance helps the less experienced most, and the Choi and Xie study found that complementarity directly, with experienced accountants adding value by knowing when to override.

What the capacity gets spent on is a management decision rather than a technical one, and worth making in advance. The broader version of that argument, including what it means for roles, is in will AI automation replace my employees.

Who is responsible if the AI miscodes something?

You are responsible if the AI miscodes something, and the arrangement has to be written down before anything posts.

Four things belong in writing. A named reviewer for each class of output, because "the team checks it" means nobody checks it. A confidence threshold that decides what gets reviewed and what posts automatically, set deliberately rather than by whatever the vendor shipped. An audit trail that records what the system proposed, what the person approved and when, so a question six months later has an answer. And a stated scope: the transactions the system is allowed to touch, and the ones that always go to a person.

For a practice, add two more. What you tell clients about AI in the engagement, and which client data is permitted to leave which system, agreed with your professional indemnity insurer rather than assumed. For a business, the equivalent is your accountant knowing what the system does before they sign anything that relies on it. Worth noting that 90% of the firms in the Financial Cents survey had no written AI policy at all, which is where most of this exposure actually sits; does a small business need an AI policy covers how to write one.

This is also where the vendor question gets concrete. Ask any provider what happens to your transaction data, whether it is used to train anything, where it is stored, and who can see it. Our general position on that is in is AI automation safe with my business data.

Systems BLACKSIG builds run on our infrastructure and we operate them from there, so the monitoring, the vendor changes and the drift checks sit with us, and the approvals stay with your people. The accountability model is not something we take off your hands, because it cannot be: the sign-off is yours and the system's job is to make it defensible.

How do we put AI into the close without breaking anything?

Put AI into the close in order of least risk per unit of time saved, which is not the order a software demo suggests.

Start with capture. Documents arriving, data read, filed against the right supplier or customer, missing ones chased. Nothing posts, nothing is decided, and it removes the single largest block of handling time. Measure the share of documents read correctly without a touch.

Then coding suggestions with review. The system proposes, a person approves in bulk, and low-confidence items get surfaced. Measure the acceptance rate and, more importantly, sample the accepted ones, because the acceptance rate on its own tells you how trusting your team is rather than how accurate the system is.

Then matching and first-pass reconciliation. The system clears what agrees and presents what does not. Measure the share clearing without a human touch, and the cost per invoice against the Ardent figures above.

Then, only after two closes have run clean, consider automatic posting for the narrow categories where accuracy has held and the cost of an error is low. Keep accruals, cut-off and anything tax adjacent with a person indefinitely.

Two conditions make the difference between this working and stalling. The chart of accounts has to be in a state somebody can defend, since a system coding into a messy chart produces confident nonsense, and is our data good enough for AI covers how to tell. And the measurement has to start before the first build, not after, which is the subject of how to measure whether AI automation is saving money.

Is this different for an accounting firm than for a business doing its own books?

An accounting firm and a business doing its own books run the same technology under a different constraint, and the constraint changes what you build first.

A business doing its own books is solving for one chart of accounts, one set of suppliers and one set of habits. Build for the specific mess you have: the suppliers who send PDFs, the one who sends a photo of a docket, the approval that happens by text message. The payback comes from volume in a narrow set of processes, and it arrives quickly because nothing has to generalise.

A practice is solving for many clients at once, which is harder in one way and easier in another. Harder because every client has a different chart, different source documents and different standards of record keeping, so anything built for one client is worth little on the next. Easier because the volume is enormous and the same four or five workflows repeat across the whole book. The right build for a practice is workflow-shaped rather than client-shaped, with client-specific rules as configuration.

Both readers hit the same wall eventually, which is that the professional work is not the bottleneck; the administration around it is. That is the pattern across professional services generally, and the capability by capability picture for firms like these is on our legal and accounting board, with maturity marked on each one.

One more number worth holding onto for either reader. The Thomson Reuters Institute's 2026 AI in Professional Services Report, built on more than 1,500 professionals across legal, tax, accounting, risk and government work, found organisation-wide AI adoption at 40%, close to double the 22% of a year earlier, while only 18% said their organisation tracks return on investment on AI at all. Being in that 18% is a cheaper advantage than being early.

Frequently asked questions

How much does it cost to automate our bookkeeping?

What moves the number is transaction volume, how many systems the books touch, how clean the chart of accounts is, how variable the source documents are, and how much review you want before anything posts. A business with 300 supplier invoices a month arriving as PDFs is a different job from a practice running 80 clients on four different ledgers. BLACKSIG does not publish a rate card; scope gets set on the call after the strategy work has said which parts are worth doing. What drives cost either way is broken down in how much AI automation costs.

How does AI bookkeeping actually work in practice?

Documents and transactions arrive, a model reads them and proposes the coding, matching and reconciliation, and each proposal carries a confidence score. Anything above your threshold goes into a batch a person approves; anything below it gets surfaced individually with the reason. Approvals are recorded, so there is an audit trail showing what the system proposed and who accepted it. Accruals, cut-off decisions and tax treatment stay outside the automated path.

Is AI bookkeeping software better than hiring a bookkeeper or outsourcing?

Different things, and most businesses end up with two of the three. Software changes the cost per transaction and leaves the judgment with whoever you already have. A bookkeeper or an outsourced provider gives you the judgment and the accountability, and the good ones are already running software underneath. The question worth asking a provider is which parts of their process are automated and what their review step looks like, because you are paying for the judgment either way. The build versus buy version of this decision is in automation consultant versus in-house.

Can our bookkeeping be automated if we run QuickBooks or Xero plus spreadsheets?

Yes, and that combination is the common case. Both ledgers have documented interfaces, so the capture, coding and matching layers can read and write without replacing anything. The spreadsheets are the part to look at carefully: a spreadsheet that holds a calculation is fine, a spreadsheet that is the only record of something is a risk before AI enters the picture. Whether a stack like that is ready is covered in does AI automation work with my software.

Will AI get our month-end close down to a day?

Not on its own. The close takes as long as its slowest dependency, which is usually waiting on documents, approvals or an answer from somebody outside finance. Automating capture and chasing compresses that waiting more than automating the accounting does. Firms that have shortened a close meaningfully did it by removing the handoffs, with coding and reconciliation speed as the second order effect.

Can AI Do Our Bookkeeping? What It Handles and What It Still Gets Wrong

The short version