AI expense categorization adds a label to each transaction using the context in the row. Magic CSV uses Jev to choose from the category names you supply. This works well for a monthly card export, a reimbursement list, or a mixed spreadsheet that needs consistent reporting labels.
What the result looks like
Illustrative example with synthetic data and authored answers, not a recorded model run. Your results will depend on the source data and instructions.
Scroll tables sideways to see every column.
| # | Merchant | Description |
|---|---|---|
| 1 | Figma | Design software subscription |
| 2 | Amtrak | Train to client meeting |
| 3 | Amazon | Order 7841 |
| # | Merchant | Description | Category |
|---|---|---|---|
| 1 | Figma | Design software subscription | Software |
| 2 | Amtrak | Train to client meeting | Travel |
| 3 | Amazon | Order 7841 | Other |
Try it, step by step
Prepare a small transaction file
Include Merchant and Description. Keep Amount and Date if you need them in the exported result, but you do not have to send every field to the model. Start with a representative selection that includes ambiguous purchases, not just obvious subscriptions.
Set the category list
In the workspace, choose AI actions → Categorize. Name the new column Category. Enter these Allowed answers, one per line: Software, Travel, Meals, Office supplies, Other. Under Read these columns, select Merchant and Description.
Define the boundary cases
Open More guidance → Extra instructions. Enter: “Use the purchase description as well as the merchant. Software includes digital subscriptions. Travel includes trains, flights, and hotels. Meals covers food and drink. Office supplies covers stated office equipment or stationery. Choose Other when the purchase purpose is missing or does not fit.” Run the column.
Review and export
Inspect uncertain cells and sample answers from every category. Correct individual answers by clicking an AI cell. Export the file as CSV or Excel, then group totals by Category in your reporting tool. Magic CSV adds the labels; it does not create an accounting report automatically.
Why the description matters more than the merchant alone
The same retailer can sell office equipment, gifts, groceries, and personal items. “Amazon” is weak evidence for a category; “Replacement keyboard for office laptop” gives the model a specific purpose. If the description is only an order number, Other is a useful result rather than a failure.
Do not make categories compete unnecessarily. “Subscriptions” and “Software” overlap for a design app billed monthly. Either choose one reporting dimension or create separate columns for Purchase type and Recurring purchase?. With separate questions, a row can be Software and recurring without forcing a single combined label.
How to check the categorization before using it
Write your own expected categories for a small test file before running the AI. Include refunds, unclear descriptions, travel-related meals, and vendors that sell many kinds of goods. Compare the returned labels against those expectations. If definitions need work, remove the test AI column and create a replacement with clearer instructions.
The review flag helps you find uncertain answers, but it does not replace checking the data. Review a sample of confident answers as well. If mistakes cluster around one pair of categories, clarify that boundary in a replacement test column instead of repeatedly correcting similar rows.
Keep reporting categories separate from accounting treatment
These labels organize supplied transaction descriptions. A merchant name does not establish whether an expense is deductible, reimbursable, approved, or personal. Keep any accounting or approval decisions in your existing review process.
This workflow starts from a file you upload. It does not connect to your bank, retrieve receipts, reconcile a ledger, or post entries into bookkeeping software. Keep the original transaction identifiers in your export so you can trace each label back to its source.
Common questions
- Can I use my own expense categories?
- Yes. Replace the Allowed answers with your reporting categories. Use a distinct name for each category and explain boundaries in Extra instructions. Include Other if some transactions cannot be classified from the available context.
- Will it categorize transactions with missing descriptions?
- It can use whichever source columns you select, but missing context can make the result unreliable. Give it an Other option and check those rows rather than asking it to invent a purchase purpose.
- Can I upload an Excel expense file?
- Yes. Magic CSV imports CSV and Excel files and exports either format. AI sheets currently support up to 1,000 rows; test a small file before processing your full export.