How to automate QuickBooks Online transaction categorization
Claude reads roughly three months of your already-categorized QuickBooks history, derives the vendor rules you have been applying by hand, and codes a fresh month of transactions in about 60 seconds. Anything it has not seen before, or that splits across categories in your history, it refuses to code and sends to a review sheet. Nothing posts to your books until you approve it.
The problem
A month of uncategorized transactions sits in the banking screen. Most owners spend a weekend a quarter working through it, or pay a bookkeeper to do the same. The work is not hard — it is repetitive, and it is the kind of repetition that quietly eats capacity you could be billing.
Why generic AI categorization fails, and this doesn't
The instinct is to have AI guess categories from a generic chart of accounts. That is the wrong shape of the problem, and it is why most tools in this space feel unreliable.
Your books already contain the answer key. Every Shell station you ever coded to Vehicle:Fuel. Every Sysco run to COGS. The pattern is sitting in your history — what is missing is something to apply it consistently. So the engine learns your coding, not a textbook's.
What you need
- A QuickBooks export of roughly three months of categorized history
- The fresh bank download for the month you want coded
That is the whole input. No credentials, no live connection to your accounting system.
What a run looks like
In the demonstration above, a real month produces:
| Stage | Result |
|---|---|
| Historical lines read | 45 |
| Vendor rules derived | 14 |
| New transactions processed | 17 |
| Coded automatically | 12 |
| Flagged for review | 2 |
| Unknown (new vendor) | 3 |
The interesting number is not the twelve. It is the five it declined to code.
The review sheet is the actual product
New vendors go to the top
A first-time Costco charge has no history behind it, so the engine will not pretend to know. You code it once. Next month it is automatic.
Ambiguous vendors get downgraded, not guessed
Amazon runs 80% Supplies and 20% bar inventory across the history. A naive tool picks the 80% and moves on. This one refuses to auto-code it.
In accounting, a confident wrong answer is the worst possible answer. A blank you notice costs a minute. A plausible miscode you don't notice can survive to the return. The engine is built to decline rather than guess.
Every automatic decision shows its evidence
Each auto-coded line carries its receipts — "6 prior lines, 100% to COGS:Beer." It is not a black box you have to trust. It is a decision you can audit in two seconds.
The approval gate
You flip the approve flags and save. The engine stages a
qbo_import.csv; you import it. Nothing touches your books until you
say go. Autopilot with a review gate — that is the part that makes it usable on
real client work rather than a demo.
It compounds
Every correction you make becomes history the engine learns from. Month one you review five lines. Month three, two. The review shrinks while the autopilot grows. That trajectory — not any single run — is the reason to set it up.
Common questions
Can AI categorize QuickBooks transactions accurately?
Yes, when it learns from your own history rather than a generic chart of accounts. Your books already contain the answer key: every Shell station you coded to Vehicle:Fuel, every Sysco run to COGS. The engine derives vendor rules from three months of your categorized history and applies them to new transactions, so it reproduces your coding decisions rather than guessing at them.
What happens when the AI is not sure how to categorize a transaction?
It refuses to code it and flags it for review. If a vendor splits across categories in your history — Amazon at 80% Supplies and 20% inventory, for example — the engine downgrades that line to REVIEW instead of picking the more common answer. In accounting a confident wrong answer is the worst outcome, so the engine is built to decline rather than guess.
Will AI post entries to my books without my approval?
No. Nothing reaches your books until you approve it. The run produces a review sheet and a staged import file; you flip the approve flags and import. Every automated line shows its evidence — how many prior transactions it matched and what percentage went to that account — so you can audit any decision before it posts.
Do I have to give an AI tool access to my QuickBooks account?
Not for this workflow. It runs on two exported files: a QuickBooks export of roughly three months of categorized history, and the fresh bank download for the month you want coded. The output is a CSV you import yourself. No credentials are shared and no live connection to your accounting system is required.
Want this running on your books?
Categorization learned from your own history, with the review gate in place. Scoped, fixed-fee, and you own what gets built.
Book a call