The same merchant gets different categories because a language model produces a likely answer each time rather than looking one up. Small differences in wording, the amount, the other rows in the batch and the category names on offer can all tip the answer. The cure is to take decisions out of the model's hands: give it a fixed list of categories, keep a mapping table of known merchants, and use rules for anything that repeats.
Why does an AI give two answers to the same question?
Assistants generate text word by word with some built-in variation. Ask it to categorise Sharma Stores twice and you may see Groceries once and Shopping the next. Context pushes it further. A payment of ₹45 might look like a snack, while ₹4,500 to the same merchant looks like a bulk purchase. If the narration includes a reference number or a city, that can shift the guess too.
What causes the drift in practice?
- Open-ended category names: if you let the assistant invent labels, you get Food, Food and dining and Eating out for the same thing.
- Different batches: categorising January in one chat and February in another means no shared memory of earlier choices.
- Ambiguous merchants: a supermarket that sells groceries, clothes and appliances really can be several things.
- Messy narrations: the same shop shows up as a clean name on one row and a string of codes on another.
- Changed instructions: a slightly different prompt each month yields a slightly different outcome.
How do I make the answers consistent?
Start with the category list. Decide on 12 to 20 categories you actually use, and paste the exact list into every prompt with the instruction to choose only from it. If nothing fits, the assistant should answer Uncategorised rather than invent a name.
Next, add a mapping table. This is a two-column list you keep: merchant name on the left, category on the right. Paste it in with the statement and tell the assistant to use the table first and guess only for merchants missing from it. Each month, add the new merchants it guessed, after you check them.
| Merchant as cleaned | Category you fixed | Why |
|---|---|---|
| Milk delivery service | Groceries | Always daily milk |
| Neighbourhood gym | Fitness | Monthly membership |
| Metro card top-up | Transport | Commute only |
| Large e-commerce store | Review each order | Mixed purchases |
Should I use rules instead of prompts?
Where you can, yes. A rule in a tracker is deterministic: the merchant matches, so the category is applied, with no variation. Prompts and tables help an assistant behave better but never make it fully repeatable. Use prompts for the unusual rows and rules for everything that repeats.
SpendixAI follows the same logic. It checks your own rules first, then your history and similar merchants, and only asks an AI model about what is left, if AI is turned on. It sends a normalised merchant name and an amount band, not your name, account numbers or balances. You can also turn AI off for a feature, set a monthly AI budget, or turn it off entirely, in which case unmatched items wait for your review.
What should I do with merchants that are genuinely mixed?
Accept that a single label will sometimes be wrong. For a store that sells everything, either create a Shopping - mixed bucket and live with it, or review those orders one by one. What matters is that the same decision is applied the same way each month, so your trend lines mean something.
A short routine to keep categories steady
- Fix your category list and keep it in one note.
- Keep a mapping table and update it monthly.
- Convert repeat merchants into rules.
- Paste the same instructions every time, unchanged.
- Spot-check ten rows before you rely on the result.
Useful ideas start with a clearer picture of your own spending.
Explore your reports


