The cost of AI in an expense app has two sides: the money, and the data that leaves your workspace to earn it. To see both, find out which of your transactions actually went to an AI model, what was sent for each, and whether you have set a budget or switched the feature off.
A privacy-minded person usually cares about the second side more. A small bill is easy to live with; a model quietly reading your full statement is not. So start with what is sent, then count how often.
What gets sent when AI categorises a transaction?
In SpendixAI, categorisation runs in stages. Your own rules go first, then your history and similar merchants. Only if AI is on does a model see what is left, and it receives the normalised merchant name and an amount band, not your name, account numbers, full statement text or balances. Optional features such as reading an unfamiliar statement layout are different: they do process the uploaded document, so treat those as a separate decision.
How do I estimate how many rows reach the AI?
Rules and history absorb most repeat spends. Only the unfamiliar remainder goes to a model. Here is a hypothetical import to show the shape, not a typical result.
| Stage | Rows handled | Rows left |
|---|---|---|
| Imported rows | - | 120 |
| Your own rules (rent, EMI name, school fee) | 45 | 75 |
| History and similar merchants | 52 | 23 |
| AI model, if switched on | 23 | 0 |
| With AI off, held for your review | 0 | 23 waiting |
So in this made-up month 23 of 120 rows, a little under a fifth, would reach a model. Next month, after you have corrected those, history handles most of them and the number shrinks. Every correction you make becomes history, which is why AI use tends to fall as you tidy up.
What controls should a privacy-minded person look for?
- A switch for AI per feature, so you can allow categorisation but refuse document reading.
- A monthly AI budget that stops spending when reached.
- A way to turn AI off entirely and have unmatched rows wait for you instead of going out.
- Plain wording about what is sent: merchant name and amount band, or whole statements?
- Aliases for merchants, so a cryptic bank string becomes a clean name you chose before anything else sees it.
SpendixAI offers all of these: AI can be switched off per feature, capped with a monthly budget, or turned off entirely, with unmatched items held for review. AI credit packs are sold separately from Pro, so check your plan page for what is included before assuming AI is free.
Where do I look for the cost itself?
Look in the app's AI settings, billing or credits area, and on the page that explains what the credits cover. If you cannot find a count of AI calls or a balance, treat that as a gap and ask support. A hypothetical check: if one AI-categorised row cost ₹0.40 of credit and 23 rows went to the model, that is 23 x 0.40 = ₹9.20 for the import. The per-row figure is made up; use the real one from your own plan.
A ten-minute audit
- Open AI settings and note which features are on.
- Read what the app says it sends for each feature.
- Set a monthly AI budget lower than you think you need.
- Add rules for your ten most common merchants so they never reach a model.
- Next month, count how many rows were held or sent, and adjust.
If anything in the wording is vague, ask before importing a sensitive statement. More detail sits on the AI and privacy page of the site.
Useful ideas start with a clearer picture of your own spending.
Explore your reports

