AI bookkeeping: what is real and what is marketing

Every accounting tool now advertises AI bookkeeping. Some of it is real, most of it is marketing on top of a rules engine that has existed for a decade. This piece is the honest map of what AI genuinely does well in a bookkeeping loop, where it fails, and how to keep the audit trail clean when there is a probabilistic model in the middle of your ledger.

What AI is genuinely good at

Three things: reading messy documents, matching things that look similar but are not identical, and suggesting starting-point categories for merchants you have never seen. All three are extraction or suggestion tasks, and all three fit naturally on the input side of a bookkeeping loop.

What AI is not good at

Two things: making the final categorization decision at scale, and reconciling numbers. The first because a category assignment is a policy choice and needs to be consistent month over month; a probabilistic answer that varies with the model version is not consistent. The second because reconciliation is deterministic arithmetic, not judgment.

Why deterministic still wins on categorization

A rule-based categorizer does the same thing every time. If the rule is wrong you can fix it once. If the rule is right it will keep being right until the merchant string changes. An AI categorizer might be right today and wrong next week, and neither you nor your accountant will understand why. In a system of record you cannot audit, that is a real problem.

Karchu uses AI as a suggestion layer on unknown merchants and as an extraction layer on scanned documents, but the categorization step itself is deterministic. Every category assignment traces back to a rule you can read and change. This keeps the books auditable even as we lean on AI for the parts it does best.

Where AI earns its place

On scanned or photographed receipts, on non-standard PDF statements, and on unfamiliar merchant strings. In all three cases the alternative to AI is either a human transcribing text or a rule you have to write for every new merchant. AI compresses both of those into an initial extraction that a human confirms in one click.

The audit trail question

Any AI in a bookkeeping loop needs to write down what it did. What model, what version, what prompt, what output, what confidence. If a categorization gets flagged three months later, you need to reproduce the decision. Vendors that hide this behind a black box are creating audit risk. Vendors that surface it (with a diff and a rollback) are treating AI as an assistant, not a replacement.

How to evaluate an AI bookkeeping tool

Ask three questions. Where in the loop does the AI actually run? Can I see the reasoning and change the outcome? What happens six months from now when I need to explain a categorization to a tax auditor? If the answers are vague, you are looking at marketing. If the answers are specific and demonstrable, you are looking at a serious product.

The next two years

AI bookkeeping will not become fully autonomous, at least not for anyone who cares about their tax filing being defensible. It will become a very fast assistant that reads documents, suggests categories, matches receipts, and flags exceptions. The human stays in the loop as reviewer and policy setter, and that is a good place for them to be.