AI bookkeeping: what AI can actually do — and can't

AI bookkeeping is the use of language models and machine learning to automate the interpretive and administrative parts of day-to-day bookkeeping: reading receipts and invoices, suggesting account codes, and reconciling bank transactions against vouchers. What doesn't get automated is the bookkeeping obligation itself — responsibility for every entry being correct and approved stays with the business, regardless of which tool did the groundwork.

The difference from older automation is how the source documents get read. A rules engine from 2015 matched exact patterns — same supplier, same amount, same layout — and broke the moment something varied. A language model reads a receipt or invoice roughly the way a person does: it recognizes amount, VAT, supplier, and date even when the layout changes case by case, and it can reason about which account is probably right based on how similar entries were booked before.

In practice that covers three kinds of source documents: receipts and photos of receipts, supplier invoices as PDFs or e-invoices, and bank statements. For each receipt or invoice the AI extracts amount, VAT rate, supplier, date, and — where one exists — an OCR reference number, then compares it against how similar documents were coded before in the same company. The result is an account suggestion with a confidence level, not a verdict: uncertain suggestions get flagged rather than guessed.

Yes, partly: AI can interpret source documents, suggest account codes, and match bank transactions, but responsibility for the bookkeeping — and approving each entry — stays with you under Swedish bookkeeping law.

In practice, that means AI does the groundwork rather than the whole job. A supplier invoice landing in the inbox can be read, get an account suggestion, and sit ready for approval without anyone touching it by hand. But "ready for approval" is the operative phrase: a human signs off before anything is posted. That's not just a technical limit — it's a legal requirement, and honestly a reasonable safeguard too, since AI still misreads unusual or damaged documents.

Another common misunderstanding is that a spreadsheet counts as bookkeeping as long as the numbers are right. It doesn't. Swedish bookkeeping law requires that ongoing bookkeeping happen systematically, in a way that can be verified and reconciled afterward — a plain Excel sheet normally doesn't meet that bar even if every row is filled in correctly, because it lacks an audit trail and protection against rows being edited after the fact. It's the same point Sweden's business-registration guidance (verksamt.se) makes: the tool has to be a bookkeeping system, not just a spreadsheet that happens to hold numbers.

Where the line actually falls shows up clearest on unusual documents: a handwritten receipt, an invoice in a language the model has seen less of, or a faded thermal-paper receipt reads worse than a crisp PDF invoice from a large supplier. Where the AI is uncertain, it should flag that rather than guess — an account suggestion below a confidence threshold should always surface as a question, not a silent assumption. That's also why review doesn't disappear as interpretation improves: you still read every suggestion before approving it; only the drafting got faster.

There's no single answer, because "best" depends on the job you want done: audit tools for accounting firms, auto-categorization inside an existing bookkeeping program, and conversational AI that works directly inside Fortnox solve three different problems.

Audit tools like Lerry.ai are built for accounting firms with many clients — they review completed bookkeeping and flag anomalies before an audit, rather than handling day-to-day entry. Auto-categorization is the category most people mean when they ask about "AI bookkeeping": Zimply, Björn Lundén, and Spiris have each built AI-based receipt and invoice interpretation directly into their own bookkeeping programs, which fits well if you're already using that program and want to skip manual coding.

Nordsynk sits in a third category: it isn't a bookkeeping program at all, but an AI client — Claude or ChatGPT — that works directly inside the Fortnox you already use, over MCP. The difference is that you don't switch systems to get AI; you add a conversational layer on top of the Fortnox you already run, with the same approval requirement before anything gets written. A full comparison of all seven tools — with pricing, Fortnox support, and concrete pros and cons — is in our Swedish-language roundup of the best AI bookkeeping tools for 2026.

Which one is right depends on where you already stand. If an accounting firm already handles your day-to-day bookkeeping, an audit tool like Lerry.ai is the firm's problem to solve, not yours — you barely notice it as a client. If you code entries yourself inside a standalone bookkeeping program, auto- categorization in that program is usually the simplest path, since the AI already sits where you're working anyway. It's only once you'd rather steer bookkeeping from a conversation — with Claude or ChatGPT, without logging into a separate interface — that a tool like Nordsynk becomes the relevant choice, and then the question is whether you'd rather type what you want done than click through a form.

Partly: AI can prepare supporting records, run reconciliations, and export SIE files ahead of the close, but the final year-end close requires human judgment on accruals, valuation, and other calls that can't be automated away.

What AI genuinely takes off your plate is the groundwork: reconciling accounts against vouchers, flagging balances that don't add up, and packaging the fiscal year into an SIE file (in Swedish) the accountant or firm can load into their own tool. What AI can't take over are the judgment calls — how to value an uncertain receivable, which costs should be accrued into next year, or whether a depreciation schedule needs adjusting. Those decisions need someone who understands the business, and for a limited company (aktiebolag), the close ultimately has to be approved and signed by the board.

Concretely, that looks like this ahead of a close: AI can reconcile VAT against the filing basis, compare account balances to the prior year and flag entries that deviate without an obvious reason, and list which vouchers are missing an attached source document before the accountant has to ask. That saves time specifically in reconciliation, usually the most repetitive part of closing the books. But whether an entry should be accrued into the next fiscal year, or how to value an uncertain receivable, are judgment calls that rest on knowing the business — not patterns in past bookkeeping — and those belong with a person.

Swedish bookkeeping law requires that every business transaction be recorded systematically and be verifiable, that accounting records be archived for several years after the fiscal year ends, and that responsibility for correct bookkeeping stays with the business owner or board — regardless of whether the work was done by an employee, an accounting firm, or an AI tool.

The archiving requirement is concrete: accounting records must be kept for seven years after the calendar year in which the fiscal year ended, whether they live in Fortnox, with an accounting firm, or in another system. That covers receipts, invoices, and vouchers in their original form — not just a summary or an AI-generated reading of them — because the original document is what an auditor or the Swedish Tax Agency can ask to see.

Two things follow from that. First, every voucher has to trace back to its source document in order — that's why an AI account suggestion is always tied to the receipt or invoice it came from, not just a floating amount and account number. Second, approval isn't a formality; it's the act that actually makes you liable. As long as an entry is waiting for sign-off, nobody has booked anything yet, which is why writes in Fortnox always pause there until you say yes.

That shows up in how corrections are handled too: a posted voucher normally isn't edited after the fact — a mistaken entry gets fixed by posting a new voucher that credits or supplements the original one. Both stay visible in the history, which is the whole point — a system where old entries could be overwritten without a trace would make the bookkeeping impossible to trust, whether a person or an AI made the change.

This is a general description of how the bookkeeping obligation works, not legal advice for your specific company — for edge cases, the Swedish Tax Agency's and verksamt.se's guidance, or your accounting firm, is the right place to ask.