Can AI do your bookkeeping? What the auto-posting apps get wrong

Every fall, our firm takes in a set of books that one of the "set it and forget it" bookkeeping apps kept on its own all year. The P&L usually has a line like $61,000 of Uncategorized Expense — or worse, no Uncategorized line at all, because the AI confidently filed everything *somewhere*, and now nobody knows which numbers to distrust.

So: can AI do your bookkeeping? Honest answer from a firm that builds AI bookkeeping software: AI can do most of the keystrokes. It cannot do the accountability. The products that pretend otherwise are the ones generating our cleanup work.

What actually goes wrong

The failure pattern is consistent enough that we can list it from memory:

• Vendor-name categorization. The AI sees "Amazon" and books Office Supplies — all of it, the $9 phone case and the $2,400 laptop that should have been a fixed asset. Vendor names identify who was paid, not what it was for.
• A generic chart of accounts. Most tools categorize into *their* idea of a chart of accounts, not yours. Every mapping that doesn't match how your preparer files your return is a reclass someone does later — or doesn't.
• Transfers booked as income and expense. The single most damaging one. Money moves from checking to savings; the tool books an expense in one account and income in the other. Do that weekly and your revenue is overstated by thousands, your expenses match, and the P&L *looks* plausible. Owner's draws and credit-card payments get the same treatment.
• Confidence hidden. The tool was 55% sure about that $847 charge, but it posted it anyway, with the same silent certainty as the Verizon bill. A wrong answer delivered confidently is worse than no answer — you can't find it later without re-reviewing everything.

Notice what's *not* on the list: "the AI is bad at categorizing." It isn't. Eighty-plus percent of a typical feed is the same thirty vendors — payroll, rent, utilities, software, fuel. A rules engine from 2010 handles those. The other 15–20% is the actual job, and the entire question is what the software does with a transaction it shouldn't be sure about.

Good automation knows what it doesn't know

The dividing line between AI bookkeeping that helps and AI bookkeeping that manufactures cleanup work is not accuracy. It's epistemics — whether the system can tell you which of its answers to trust. Four tests, usable on any product including ours:

1. Does the AI propose, or post? Every transaction that reaches the general ledger should have passed a human approval — one click on the confident ones, real attention on the rest. If the tool's pitch is that nobody has to look, the tool's output is that nobody did.
2. **Does it learn *your* chart of accounts from *your* corrections?** Recategorize something once and the next occurrence should come back proposed your way — without you writing a rule by hand. If your corrections don't teach it, you're the software's employee, not the reverse.
3. What does it do with transfer-looking transactions? The right answer is: hold them until the other side shows up, and never, ever post half a transfer as income or expense. Ask a vendor this question directly. The pause before they answer tells you everything.
4. Does it show its confidence and queue the uncertain ones? The dangerous transactions aren't the ones the AI flags — they're the ones it should have flagged and didn't. You want the uncertain 20% in a review queue, visibly, not distributed silently through your P&L.

"But reviewing everything defeats the point of automation"

It doesn't, and this is the part the marketing gets backwards. Reviewing a proposed categorization with the document attached takes seconds — the work of finding, reading, deciding, and typing is already done. What used to be an afternoon of data entry becomes minutes of approvals. That's the real productivity gain, and it survives contact with tax season, because a human saw every entry.

The review was never the busywork. The review is the bookkeeping — it's the moment someone accountable decides what a transaction *was*. Automation should erase everything around that moment, not the moment itself.

Where Woven stands on this

Disclosure: Woven is our product, and the position above is built into it rather than bolted on. The bank feed's AI proposes categorizations with its confidence shown; a person approves before anything posts. It learns each client's own chart of accounts from corrections — deterministic learned rules, not vibes. Anything that looks like a transfer — including transfers *between* related companies — is held until its counterpart appears, so half a transfer never lands as income. And every AI feature has a per-client off switch, because some clients don't want AI near their books at all, and that's a legitimate choice.

It was built by the owner of an accounting firm whose own practice runs on it — which means our bookkeepers hit the AI's mistakes before your books do.

The short version

• AI handles the repetitive 80% of a bank feed well. The other 20% is the job.
• Auto-posting tools fail loudly in cleanups: vendor-name categories, generic charts of accounts, transfers booked as revenue, confidence hidden.
• Evaluate any AI bookkeeping tool with four questions: propose or post? learns your chart of accounts? holds transfers? shows confidence?
• Human-verified beats human-free. The review is the bookkeeping.

If your books spent the year inside an auto-poster and you're not sure what to trust, that's a fixable problem — [book a demo](/demo) and we'll show you what proposed-and-approved looks like on a real feed.