AI Categorize Bank Transactions for QuickBooks: Suggestions, Rules and a Mapping You Own

AI can categorize bank transactions for QuickBooks in the sense that QuickBooks Online now pre-fills a category when it has high confidence, and bank rules you write apply a category automatically. Neither replaces the step that makes categories reliable: turning a raw bank description like "POS PURCHASE AMZN Mktp US*2K4L7 SEATTLE WA" into a stable payee ("Amazon") and mapping that payee to an account name that exists, spelled exactly the same, in your chart of accounts. This post shows that workflow with a real before/after file, and where the AI layer helps and where it guesses.

Formatting the bank CSV itself (dates, signs, the 3-column and 4-column layouts) is covered in how to clean bank statement CSVs for QuickBooks, Xero and Wave and QuickBooks 3-column vs 4-column bank CSV. This post is only about categories.

What QuickBooks' AI does, and what rules do

On the AI-powered banking page, Intuit says QuickBooks "suggests matches or categories when it has high confidence" and "only pre-fills categories with high-confidence suggestions". Each suggestion carries a Review Signal: green when there is "a clear pattern in your history", blue when there is some data but variation, orange when the data behind it is limited. Intuit tells you to confirm before posting even for the strong ones, and says QuickBooks "will learn from your changes".

Bank rules are the deterministic layer underneath. A rule has up to five conditions on Description, Bank text or Amount (contains, doesn't contain, is exactly) and sets the transaction type, category, payee and tags. Rules apply in order, one per transaction. There is an auto-add option, which Intuit's help describes with a warning worth quoting: if you leave it on "you won't get a chance to review transactions before QuickBooks adds them to your books". Transactions posted that way carry an AUTO badge.

Xero's equivalent, for comparison, has four methods: your rules, exact matches, your own history ("Memory") and "Prediction", which its reconciliation page describes as suggestions "based on how other Xero users reconciled similar transactions". That last one is worth knowing if you care where the pattern came from: it is other businesses' categorisation history, not yours.

Why the import file has to match your chart of accounts

QuickBooks Online's own CSV upload for bank transactions takes only Date, Description and Amount (or Credit and Debit); Intuit's help articles for the upload describe exactly those two layouts and no category column. Categories are then applied in the For Review tab, by you, by rules, or by the AI suggestion. So for a native upload, the category lives in QuickBooks, and the file's job is to make the Description clean enough for rules to hit.

If instead you import categorised transactions through a third-party importer, as journal entries, or into Xero or Wave with a category column, the account name in the file is matched against your chart of accounts by name. A file that says "Office Supplies" when the chart says "Office supplies & software" typically either fails the row or creates a new account, and the second outcome is worse because it looks like success. Either way, the fix is the same: keep the mapping from payee to account name in one table, copy the account names out of the chart of accounts rather than typing them, and re-use the table every month.

Before and after: raw description to payee to category

Here is a six-row slice of a (fictional) US bank export as it arrives:

Date,Description,Amount
03/04/2026,"POS PURCHASE  AMZN Mktp US*2K4L7  SEATTLE WA",-42.17
03/05/2026,"ACH CREDIT STRIPE TRANSFER ST-X1Y2Z3",803.00
03/07/2026,"ONLINE PMT  COMCAST   CABLE",-129.99
03/09/2026,"POS PURCHASE  AMZN MKTP US*9Q1F2  SEATTLE WA",-18.40
03/11/2026,"CHECK 1042",-1250.00
03/12/2026,"ADOBE  *CREATIVE CLOUD  SAN JOSE CA",-59.99

Two cleaning steps make this categorisable. First, normalise the description (trim, collapse spaces, uppercase) so the same merchant looks the same every month. Second, derive a Payee from a stable substring, because the suffix after the asterisk changes on every Amazon charge. Then a third, separate table maps Payee to the account name copied from the chart of accounts:

Applying the two steps and the mapping gives the file you actually want to review:

Date,Description,Amount,Description Clean,Payee,Category
03/04/2026,"POS PURCHASE  AMZN Mktp US*2K4L7  SEATTLE WA",-42.17,POS PURCHASE AMZN MKTP US*2K4L7 SEATTLE WA,Amazon,Office supplies & software
03/05/2026,"ACH CREDIT STRIPE TRANSFER ST-X1Y2Z3",803.00,ACH CREDIT STRIPE TRANSFER ST-X1Y2Z3,Stripe,Undeposited Funds
03/07/2026,"ONLINE PMT  COMCAST   CABLE",-129.99,ONLINE PMT COMCAST CABLE,Comcast,Utilities: Internet
03/09/2026,"POS PURCHASE  AMZN MKTP US*9Q1F2  SEATTLE WA",-18.40,POS PURCHASE AMZN MKTP US*9Q1F2 SEATTLE WA,Amazon,Office supplies & software
03/11/2026,"CHECK 1042",-1250.00,CHECK 1042,(check number),Ask My Accountant
03/12/2026,"ADOBE  *CREATIVE CLOUD  SAN JOSE CA",-59.99,ADOBE *CREATIVE CLOUD SAN JOSE CA,Adobe,Office supplies & software

Notice what the mapping did not do: it did not guess. The check went to Ask My Accountant with the reason written down, and the Stripe deposit went to Undeposited Funds rather than Sales, because booking a net payout as revenue is how Stripe deposits get double counted. Both are decisions for a person, and the file says so.

Doing this as a deterministic step

In PipeSheets the three steps above are a saved pipeline: trim whitespace (with internal-space collapsing), then a find-and-replace or map-values step that assigns Payee from a contains-match, then a second map-values step from Payee to Category using the table above. The mapping table is editable, lives in the pipeline, and runs the same way on next month's file. The preview shows every row whose Category is empty, which is your review queue, before anything is downloaded. Nothing here involves a model; it is a lookup you control, and the same input always produces the same output.

That is also our view of where AI suggestions belong in this workflow: proposing a new mapping row when a payee appears for the first time ("ADOBE looks like software; add to the table?"), with you approving it, rather than categorising individual transactions on the fly where no one sees the decision. QuickBooks' Review Signals are a step in that direction; a mapping table you can read is the rest of it.

Where AI categorisation goes wrong

A monthly QA checklist for categorised transactions

For the file-format side of the import, the QuickBooks CSV import cleanup page covers the 3-column and 4-column layouts and date handling, and the bank statement CSV cleaner page shows the description-cleaning steps that make rules hit consistently.