What an AI bookkeeper actually does all day.
Strip the marketing and the job has five parts. It READS documents: an emailed supplier invoice, a photographed fuel receipt, a PDF statement — extracting the vendor, the line items, the tax, the total. It MATCHES: this bank feed line of $768 against that open invoice, this card settlement against yesterday’s till total. It DRAFTS: a complete double-entry journal entry — debit this account, credit that one — for each thing it read or matched. It CATEGORISES the stragglers: the bank line with no document, assigned to an account with a confidence score. And it FLAGS what smells wrong: a duplicate invoice, a supplier price jump, a payment with no counterpart.
Notice what is not on the list: deciding. In a well-designed system the AI prepares and a person disposes. The draft sits in a queue with its evidence attached — the source document, the extracted fields, the proposed entry — and the human approves, edits, or rejects. The unit of work changes from typing (minutes per document) to reviewing (seconds per document). That is the entire economic argument, and it is enough.
The line that matters: automation is not autonomy.
Every credible AI bookkeeping product in 2026 draws the same line: the AI may PREPARE anything, but nothing POSTS to the ledger without a human decision. This is not caution theatre. Your books are a legal record — the basis of your tax filing, your bank’s lending decision, your own view of reality. An error that posts silently compounds silently: a mis-read invoice becomes a wrong expense, a wrong margin, a wrong tax figure, discovered months later by someone billing by the hour.
So when you evaluate any product, find this line and check which side of it each feature sits on. “AI categorisation” that writes straight to the ledger with no review queue is autonomy wearing automation’s clothes. A draft queue with evidence attached, an immutable audit log of who approved what, and the ability to reject without a trace left on the books — that is the shape of the safe version. The speed difference between the two is minutes per month. The risk difference is the whole point.
A worked example: from receipt photo to posted entry.
Concrete beats abstract. A driver photographs a fuel receipt: $45.00, paid in cash. The AI reads the vendor, the amount, and the payment method, and drafts: debit Vehicle Expenses $45, credit Cash $45 — with the photo attached and a note of what it read. The owner sees the draft, the photo beside it, taps approve. Elapsed human time: about four seconds. The same evening, the bank feed shows a $1,240 card settlement; the AI matches it against yesterday’s till total, drafts the clearing entry, and attaches both sides of the match.
Multiply by a month: eighty supplier invoices forwarded by email, three hundred bank lines, a shoebox of receipts. The typing that consumed a bookkeeper’s week becomes a review queue an owner clears with coffee. And because every draft carries its evidence, the review is genuinely fast — you are checking a claim against its source, not reconstructing where a number came from.
How accurate is it, honestly?
On the mechanical work, the AI is simply better than people. It does not transpose digits, never types $4,500 as $5,400, does not skip line eleven of a forty-line statement, and applies the same categorisation rule at 11pm as at 9am. On clean, recurring documents — the same suppliers, the same shapes of transaction — draft accuracy in mature products runs high enough that review becomes a skim.
Where it fails is judgment and novelty. The first invoice from a new supplier in a new currency. A settlement that nets three things a human would book separately. Whether that equipment purchase is an expense or an asset — a question with a right answer that depends on your jurisdiction, your policy, and your accountant’s preference. Good products handle this honestly: low-confidence drafts are flagged as such, unusual transactions are routed for a human decision rather than guessed at, and the system learns your corrections. The failure mode to refuse is confident nonsense — a system that guesses novel cases with the same swagger as routine ones.
AI bookkeeper vs human bookkeeper vs bookkeeping service.
The three options are not rivals for the same job. AI bookkeeping SOFTWARE (roughly $29–115 a month, checked August 2026) removes the typing and the matching; you or your team still glance at the queue and someone still closes the month. A done-for-you SERVICE — Bench from about $190 a month, Zeni from about $549 — removes the queue too: humans plus their software do everything and hand you statements. A part-time human bookkeeper (a few hundred to a couple of thousand a month depending on market) brings judgment, local tax fluency, and someone to call — increasingly spending their hours on review and close rather than data entry, because their tools now include the same AI.
The honest decision rule: if you or a staff member can spare minutes a day, AI software is the cheapest path by an order of magnitude — and if your business sells physical products, software that also runs the till and the stock sees the whole picture and drafts better entries for it. If nobody in the business will ever open the books, pay for the service and count it as what it is: a full-service subscription. And keep a human accountant for what humans are for — tax strategy, year-end, judgment calls — whichever path you take.
Why one ledger changes what the AI can see.
An underrated detail: an AI bookkeeper is only as good as its context. A standalone bookkeeping tool sees your bank feed and your uploaded documents — and nothing else. It meets a $1,240 deposit cold, guesses, and asks you. An AI that lives inside the same ledger as your till and your inventory meets the same deposit already knowing yesterday’s card totals, the invoice it settles, and the customer it belongs to. The match stops being a guess.
This is why the architecture question — is the AI bolted onto the books, or built into the system that also runs operations? — matters more than any model benchmark. The same intelligence with richer context drafts entries that need less correcting. When you trial products, test exactly this: forward the same messy week of documents and bank lines to each, and count how many drafts you had to fix.
Six questions to ask any AI bookkeeping product.
Take this list into every demo and free trial. The answers separate marketing from machinery.
- Is there a REAL double-entry ledger underneath — journals, trial balance, audit trail — or a categorised transaction list dressed as one?
- Can anything post without human approval? (The only acceptable answer: no.)
- Does every draft carry its evidence — source document, extracted fields, match reasoning?
- What happens on low confidence — flagged and routed to you, or guessed silently?
- Does the AI see your whole operation — sales, till, stock — or only a bank feed?
- Can you export every record, any day, in open formats — including the day you leave?
Where Nonari fits. (Full disclosure: ours.)
Nonari is the product we build, so weigh this section accordingly. Its AI bookkeeper is the draft-and-approve design this guide describes — forwarded invoices, receipt photos, and bank lines become drafted entries with evidence attached, and nothing posts without a person — built into a real double-entry ledger that also runs the till, the stock, and multi-branch books. That one-ledger context is the design bet: the AI drafts from the whole picture, not a bank feed in isolation.
The commercial part is simple: every plan carries the full AI allowance — $29 a month for a single location, $70 for up to three branches, unlimited users, 15-day free trial with no card. If you are comparing the whole field first, the honest tool-by-tool comparison is linked below; it names who should NOT pick us, too.