AI Bookkeeping: An Honest Guide to What It Can and Can't Do

AI is genuinely good at four bookkeeping jobs: categorizing routine transactions at high volume, spotting anomalies a human would miss, pulling data off receipts and invoices, and keeping dashboards current daily instead of monthly. It is genuinely bad at novel transactions, judgment calls, tax positions, and accountability. Good bookkeeping now uses both: software for scale, a person for everything that requires judgment.

We run an AI-powered bookkeeping firm, so you'd expect us to sell you on the technology. This is the more useful version: an honest map of where it earns its keep, where it fails, and why we won't send you a report no human has reviewed.

What AI does genuinely well

Categorization at scale

Most businesses generate hundreds or thousands of transactions a month, and the majority are repetitive: the same vendors, the same subscriptions, the same payroll runs. Modern models categorize these accurately and instantly, applying the same logic to transaction 4,000 as to transaction 4. That consistency matters as much as the speed. Human bookkeepers doing pure data entry get tired, drift, and make different calls on identical transactions in week one and week four. Software doesn't.

Anomaly detection

AI is very good at noticing that something is off: a vendor payment 3x its usual size, a duplicate charge, a subscription that jumped in price, a deposit that doesn't match any open invoice, spending in a category that broke sharply from its trend. A human reviewing a 900-line ledger will skim; the software checks every line against the pattern, every time. The best systems don't act on the anomaly, they flag it for a person, which is exactly the right division of labor. Detection is cheap and safe to automate; the decision about what the anomaly means is not.

OCR and data capture

Reading a receipt, extracting the vendor, date, amount, and tax, and attaching the image to the right transaction used to be the most tedious work in bookkeeping. Document extraction now handles the overwhelming majority of clean documents without human keystrokes. That kills the shoebox problem: receipts get captured when they happen, not reconstructed at year end from faded thermal paper.

Real-time dashboards

When categorization runs continuously instead of in a monthly batch, your numbers can be current every morning: cash position, revenue against last month, spending by category, who owes you what. The traditional model made you wait for the close to see anything. Continuous processing means the close confirms what you've already been watching, rather than revealing it.

Where AI fails

Novel transactions

AI categorizes by pattern, and a new pattern has no precedent. Your first equipment lease, an insurance payout, a legal settlement, the deposit from selling a truck: a model will confidently file these somewhere, and confidence is the problem. A wrong guess doesn't look wrong. It sits in your books distorting margins until a human happens to notice. The transactions AI handles worst are usually the largest and least frequent, which means the smallest share of your ledger carries the biggest share of the risk.

Judgment calls

Is that payment to a contractor a cost of goods sold or overhead? The answer changes your gross margin, and it depends on what the contractor actually did, context that lives in your head, not in the transaction data. Is the owner's vehicle a business expense, and what share? Should that big invoice be recognized now or spread across the project? These aren't pattern-matching questions. They're judgment questions, and they're worth real money.

Tax positions

How a transaction is treated for tax purposes is a decision with consequences, sometimes years later. Software can suggest treatments; it cannot weigh your risk tolerance, know your full situation, or stand behind the position if it's questioned. Anyone who lets an unreviewed model make tax decisions has confused a suggestion engine with an advisor.

Accountability

When a number is wrong, someone has to notice, own it, fix it, and explain it. Software cannot be accountable; it can only be corrected. If your books are wrong and your bookkeeping is a subscription with no human attached, the accountability lands back on you, which is precisely what you were paying to avoid.

Why human review is non-negotiable

The failure mode of AI bookkeeping isn't chaos, it's plausible-looking error. A model that miscategorizes a loan deposit as revenue produces a clean, professional income statement that overstates your income. Nothing about the report looks broken. That's what makes unreviewed automation dangerous in accounting specifically: errors don't crash, they compound quietly, and you find them at tax time or during a loan application, when they're expensive.

So the question to ask any bookkeeping service, including us, isn't "do you use AI?" Everyone will, soon enough. It's "who reviews the output, how often, and whose name is on my close?" If the answer is nobody, you don't have a bookkeeper. You have software with better marketing.

How we split the work at FigureWise

Our model is simple: AI does the volume, a human owns the result. Software handles first-pass categorization, document capture, continuous reconciliation checks, and anomaly flagging. A real bookkeeper reviews the flags, makes the judgment calls, verifies every reconciliation, closes your books each month, and answers your questions, the same person, month after month.

Pure-software tools are honest value for very simple books and owners who enjoy the work. Traditional human-only firms do careful work at a slower pace and higher labor cost. We just think the pairing beats both: a strong bookkeeper with strong tools catches more, closes faster, and costs less than either alone. That's not a slogan about the future of AI. It's how the work gets done here every month.