How to do bookkeeping with AI

What AI actually does well in bookkeeping

Stripped of marketing, there are four jobs where machine reading and machine matching are clearly better than a person doing it by hand, and they cover most of the hours.

Notice the shape of the right-hand column. It is not a list of things AI does badly — it is a list of things that require knowledge of the business rather than knowledge of the document. That distinction is the most useful test you can apply when deciding whether to automate a step: if the correct answer is visible on the page, automate it; if the correct answer requires knowing something about this business that is not on the page, keep a human in it.

flowparse.io

What AI must not be allowed to do

Being specific here saves more money than any feature. These are the boundaries that keep automated books defensible.

It must not decide tax treatment. Whether a cost is capital or revenue, whether a vehicle expense is allowable, whether VAT is recoverable on a particular purchase — these are legal questions with jurisdiction-specific answers, and the document rarely contains enough information to settle them. FlowParse does not decide deductibility, does not compute tax, and does not file returns anywhere.

It must not post unreviewed to a live ledger. Automation that writes directly into the books with no review step converts a small extraction error into a systematic one, because the same wrong rule applies every month. Post through a review queue until a rule has proved itself over several periods.

It must not be the only check on itself.A model's confidence score is an opinion about what it read; it says nothing about what it never saw. Every automated pipeline needs at least one check that is arithmetic rather than probabilistic — which is the subject of step four.

The five-layer stack

Whatever tools you choose, an AI-assisted bookkeeping setup has the same five layers. Naming them makes it obvious what you are missing.

Most stacks that disappoint are missing the validation layer entirely. It is the least visible of the five and the only one that tells you when the other four have failed — which is exactly why it gets skipped, and exactly why the errors surface months later.

Step 1 — Connect the feeds and fix the foundations

Before any AI touches anything: connect every bank and card feed your ledger supports. A feed comes from the institution itself, arrives continuously, and cannot be mistyped or half-read. No extraction tool — ours included — beats that, and any vendor pretending otherwise is arguing against your interest.

Then do the unglamorous foundation work, because AI amplifies whatever structure it finds. A chart of accounts with sensible, distinguishable categories makes automated classification far more reliable; one with forty near-identical expense codes guarantees inconsistency whether a human or a machine is choosing. Set opening balances correctly. Decide which accounts are in scope. Agree who reviews what.

Then list the accounts a feed cannot cover — and there are always some. Closed accounts, foreign accounts, small institutions, older credit cards, the months before the feed existed. Those become the document workflow, and they are where statement conversion earns its place.

flowparse.io

Step 2 — Make document intake boring

The failure mode of every document workflow is not extraction quality. It is that documents arrive four different ways, three of them undocumented, and a quarter of them do not arrive at all until someone chases.

Pick one route per document type and enforce it. Receipts and purchase invoices: a capture app or a single email address, not WhatsApp, not a shared drive, not the bookkeeper's inbox. Bank and card statements for the accounts without feeds: a monthly download into one folder, named consistently by account and period. Anything else — a loan statement, a merchant settlement report, a broker statement — gets a named home too.

The reason to be strict is not tidiness. It is that a consistent intake is the difference between processing a month in twenty minutes and spending an afternoon establishing what is missing. If you are choosing a capture tool for this layer, Dext and AutoEntry are the two most common answers, and the receipt roundup covers the wider field.

Step 3 — Extract instead of retyping

This is where AI earns its keep first, because retyping is pure cost: it produces no judgement, it introduces errors at a stable rate no amount of care removes, and it is the reason bookkeeping backlogs exist.

For each document type the extraction should return typed data rather than text: a bank statement becomes dated, signed transactions with descriptions rejoined and the running balance kept separate from the amount; an invoice becomes header fields plus line items with tax; a receipt becomes merchant, date, total and tax. Scans and photographs go through OCR first and anything uncertain is flagged rather than silently guessed.

Two details separate a workflow that scales from one that annoys everybody. First, debit and credit columns must collapse into one signed amount — and the running balance must never be exported as the transaction amount, an error that puts a wrong number on every row while looking entirely plausible. Second, dates must be resolved rather than assumed, because 03/04/2026 is March in one country and April in another, and getting it wrong moves transactions between periods.

flowparse.io

Step 4 — Prove the data is complete, do not assume it

This is the step almost every AI bookkeeping guide skips, and it is the one that decides whether your automated books are trustworthy.

Accuracy and completeness are different properties. Accuracy asks whether the numbers that came out are right. Completeness asks whether all the numbers came out. A statement extracted with three transactions missing scores perfectly on accuracy: every field on every extracted row is correct, and nothing is flagged, because a row that was never emitted was never uncertain. The books look plausible and are wrong.

The statement, however, carries its own proof. Opening balance, plus every transaction, equals the closing balance the bank printed. Run that sum per account and completeness stops being a matter of trust — FlowParse runs it automatically on every statement, names the rows where it breaks and returns a score you can gate on. Invoices have a weaker equivalent: the sum of the lines against the printed total.

If your tools do not do this, do it yourself in a spreadsheet before posting. It takes a minute per statement and it is the only check available that can prove an extraction wrong with no reference data and no human in the loop.

flowparse.io

Step 5 — Categorise with AI, review like an adult

Automated categorisation is genuinely good on repeating payees and genuinely unreliable on anything requiring context. Both facts are true at once, and the workflow has to reflect it.

Start supervised. For the first month or two, review every proposed category and correct rather than accept. The corrections are the training signal — for the tool if it learns, and for the rules you write if it does not. After that, move to reviewing exceptions: anything new, anything above a threshold, anything the system was unsure about, plus a random sample of confident classifications to catch systematic drift.

Some categories should never be automated at all, and it is worth naming them explicitly in your process: transfers between own accounts, director or owner transactions, anything that might be capital rather than revenue, and any payment that could be a personal expense. Those are the four that quietly distort a set of accounts, and all four require knowing the business rather than reading the page. Our guide to categorising bank transactions goes deeper on the mechanics.

Step 6 — Reconcile, because it is the only real test

Reconciliation is not an administrative chore at the end of the process; it is the test that everything before it worked. If the ledger agrees with the bank, transaction by transaction and balance to balance, then capture, extraction, completeness and posting all held. If it does not, one of them failed and you now know to look.

AI helps here in a specific way: matching. Fuzzy payee names, split payments, batched deposits that cover several invoices, small FX differences — these are pattern-matching problems and machines are good at them. What machines cannot do is decide what an unexplained difference means, which is exactly where the bookkeeper's attention should be concentrated.

Reconcile monthly, not quarterly. A month of mismatches is an afternoon; a quarter is a project; a year is the reason catch-up bookkeeping exists as a service. Our reconciliation guide covers the method and the tooling in detail.

flowparse.io

Step 7 — The monthly review that catches what automation misses

Automation does not remove the need to look at the result; it changes what you look at. Instead of checking whether numbers were typed correctly, you are checking whether the picture is plausible.

Half an hour of this per client per month catches more real problems than any amount of transaction-level checking, because the errors that survive automation are systematic rather than random — and systematic errors are visible in ratios long before they are visible in rows.

The gaps AI cannot fill

Four things that no amount of automation addresses, and pretending otherwise is how projects lose credibility.

• Documents that do not exist. AI reads what it is given; a purchase with no invoice stays a purchase with no invoice.
• Cash that was never recorded. If the till discipline is weak, the books are weak, and automation changes nothing about that.
• Business context. Whether the transfer was a loan or income is knowledge, not extraction.
• Responsibility. Someone signs off the accounts, and it is not a model.

There is a fifth worth adding for anyone evaluating vendors: archiving. Storing an image is not the same as meeting a jurisdiction's record-keeping standard. FlowParse in particular deletes the original document immediately after extraction — good for privacy, explicitly not an archive — so retention of the source paperwork stays your responsibility.

Budgeting review time honestly

The most common planning error in AI bookkeeping is budgeting zero for review, then discovering the real number and concluding the automation failed. A realistic budget looks roughly like this for a small business month.

The pattern is consistent: total time falls sharply, and the proportion spent on judgement rises. That is the actual promise of AI in bookkeeping, and it is a better promise than the one usually advertised — a practice that spends its hours on the parts a client would pay more for.

Controls and the audit trail

An automated ledger needs to be explicable months later, to a partner, a client or an auditor. Three controls make that possible and none of them are expensive.

Traceability. Every posted transaction should be traceable to the document it came from. If you consolidate statements, keep a source-file reference on every row — that is what Smart Merge writes, and it is the difference between explaining a row in ten seconds and re-deriving a year.

Evidence of the check. Keep the validation result, not just the data. A stored reconciliation status per statement — reconciled, failed, not verifiable — is exactly the evidence an auditor asks for when they want to know how you knew the data was complete.

Separation of automation and approval. Whatever posts should not also approve. That is trivial in a small practice — one person automates, another reviews — and it is the control that keeps an automated error from becoming an unnoticed one.

How automated books actually go wrong

Not dramatically. The failures are dull, systematic and slow, which is precisely why they survive.

A statement is imported twice, so a month of spending appears doubled and the reconciliation is forced with a balancing entry nobody revisits. A recurring supplier is miscoded once and the rule repeats it for eleven months. A page break drops four transactions and the closing balance is never checked, so the difference sits in a suspense account until year-end. Debit and credit columns are read as one unsigned column, so refunds increase costs. The running balance is exported as the amount, so every row is wrong by a plausible number.

Every one of those is caught by the same two habits: reconcile monthly, and validate completeness at the point of extraction. Neither takes long. Both are skipped constantly, because the output looked fine — and looking fine is exactly what a systematic error does.

flowparse.io

Starting from a backlog rather than a clean month

Most AI bookkeeping projects do not begin with a tidy business. They begin with a client who has not touched their books for a year and a folder of PDFs.

Handle that as a project rather than as a heavy month. Gather statements for every account and every month first, so the gaps are known before any data entry starts. Convert in bulk rather than one file at a time — a year across three accounts is thirty-six documents and several hundred pages, and consolidation with duplicate detection is what stops overlapping periods being counted twice. Then work month by month, reconciling as you go, so an error is contained rather than repeated.

The full method — including what to do about missing receipts and how to stop the backlog re-forming — is in the catch-up guide.

flowparse.io

Running a practice on this

At practice scale two things change. The first is standardisation: the same intake, the same tools and the same monthly checklist for every client, because variation is what makes a portfolio unmanageable. A client who insists on emailing photographs to a partner is a process problem, not a technology problem.

The second is where staff time goes. With extraction automated, junior time shifts from typing to review — which is more valuable and also more demanding, so it needs training rather than assumption. The practices that get the most out of automation are the ones that explicitly redefine the junior role as reviewing exceptions and asking clients better questions.

Volume-specific tactics, including how firms batch statement work across clients, are in processing statements at scale and the accountants' workflow page.

What it actually saves

Be sceptical of percentage claims, including ours, and measure your own baseline before changing anything: how long does one client month take today, end to end, including chasing.

The consistent, defensible saving is in transcription. Statement entry that took hours becomes minutes; purchase document entry falls by a similar factor; categorisation falls once rules settle. Reconciliation improves less, because the hard part was never typing — it was investigating differences, and that remains human work.

The second saving is quieter and larger: errors caught at the point of extraction instead of at year-end. An hour spent finding a missing transaction in July is an hour; the same missing transaction found in March of the following year costs a re-reconciliation, a conversation with the client and often an amended filing.

Common mistakes

A sixth, less obvious: buying a platform when the problem is a workflow. Many firms evaluate expensive document-AI platforms for a job that is a converter plus a review habit — the platform comparison sets out when the heavier purchase is genuinely warranted.

The setup checklist

• Every available bank and card feed connected in the ledger.
• A written list of accounts no feed covers, with a monthly statement routine for each.
• One intake route per document type, agreed with the client and enforced.
• Extraction that returns signed amounts, resolved dates and rejoined descriptions.
• A completeness check on every statement, per account, before anything posts.
• Categorisation reviewed fully for the first two months, then by exception plus sampling.
• Monthly reconciliation of every account, with differences investigated rather than forced.
• A monthly review of ratios, suspense balances and missing documents.
• Source-file traceability from every posted row back to a document.
• Originals retained wherever your jurisdiction requires them — the extractor is not an archive.

Frequently asked questions

Keep reading