How accurate is it?

No categorizing tool is right on every list, and we do not publish an accuracy score we graded ourselves: a score from our test lists says little about your categories and your rows. This page covers what the add-in guarantees by design, where it gets it wrong, and how to check it yourself.

What it guarantees by design #

• Answers come only from your list. Every answer is one of your categories, or Other. It cannot invent a category that is not in your list.
• A confidence on every row. Each AI answer comes with a confidence and a status: Categorized, Review or Other.
• Review below your threshold. Every answer with a confidence below your threshold is marked Review with a light fill, and Show review rows filters the sheet down to those rows.

None of this makes every answer right. It makes every answer checkable, with the doubtful ones marked. See Review results and set the threshold.

Where it gets it wrong #

• Rows that could mean two things. AMZN MKTP could be office supplies or anything else Amazon sells; UBER *EATS is a meal, UBER alone could be a taxi. Rows like these often come back as Review, and sometimes as a confident answer that is right for most businesses but not for yours.
• Categories that overlap. When two of your categories could both fit, the AI picks one, and not always the one you would. A short description on each says where the line runs (see Write categories that work).
• Too little text. A bank line that is only a reference number, or a comment that says “ok”, gives the AI nothing to go on.
• Rules it cannot see. If the right category depends on the amount, the date, or something only you know (this client books software to a project), the text alone cannot decide it. Use a formula or a rule for those.

Check it on the samples #

Both samples in the add-in (menu, Insert sample data) come with an answer key, so you can compare every row, including the ones we expect to land in Review. Download them to check the result in your own Excel:

• Bank lines to P&L accounts: 25 bank lines and a 16-account chart of accounts, with the answer key.
• Customer comments to topics: 20 comments and 7 topics, with the answer key.

Bank lines to P&L accounts

Customer comments to topics

Other languages #

It was tested on English and on some German text; the comments sample has one German comment against English topics (row 24 above), so you can see how that row comes out. For any other language, or for your own wording, the free credits are the way to check: sort a slice of your own rows and look at the Review and Other rows.

When another tool is better #

• Your export already carries a category or merchant code you trust: use a lookup.
• A handful of fixed keywords covers every row and nothing new ever appears: a keyword table with XLOOKUP is free and exact.
• The category depends on amounts or dates: use a formula.
• Your policy allows no outside service to see the text: keep it in Excel with formulas.

AI Categorizer is for the rest: many rows, new wording every month, and a category list you already keep. The fairest test is your own data, and the first 100 credits are free for that.