AI and MCP

Creating categorisation rules by chatting: what it means for privacy

Chatting to create categorisation rules can reduce how much of your data an AI ever sees. Here is why, what to share, and how to stay in control.

Editorial illustration for Creating categorisation rules by chatting: what it means for privacy

Creating a categorisation rule is one of the more privacy-friendly things you can ask an assistant to do. A rule is a fixed instruction stored in your own workspace. Once it exists, matching transactions are tagged without any assistant or AI looking at them again. So the privacy question is how little you can share to create the rule in the first place.

Why rules shrink your exposure

SpendixAI categorises in stages. Your own rules come first, then your history and similar merchants, and only if AI is switched on does a model look at what is left, using a normalised merchant name and an amount band, not your name, account numbers, full statement text or balances. Every merchant a rule covers is one that never needs the AI stage.

Say a month has 60 transactions and your rules cover 41 of them. Only 19 go on to history matching and, if enabled, AI. If you add rules for five more regular payees covering 9 more rows, only 10 remain.

What to share when you ask for a rule

  • The merchant wording as it appears in the description, nothing else.
  • The category you want.
  • Not your name, not an account number, not a balance, not a photograph of a statement.

If the assistant needs to see examples, give two or three description lines, not the whole statement. You can also build the rule in the app and skip the assistant entirely.

Controls you have

  1. Connect Claude or ChatGPT through MCP only if you want to. Pick the scope, start read-only if you like, and disconnect whenever you want.
  2. Switch AI categorisation off per feature, set a monthly AI budget, or turn it off entirely. Unmatched rows are then held for review rather than guessed.
  3. Keep records in your own workspace with role-based access.
  4. Export your data, choose document retention, delete transactions, or schedule workspace deletion with a 7-day window to change your mind.

What a privacy-minded rule list looks like

Rule wordingCategoryWhy it is safe
contains SOCIETY MAINTHousingSpecific, no personal detail needed
contains MOBILE RECHARGEMobile and internetSpecific enough to avoid FASTag rows
contains PHARMACYHealthCategory only, no medicine names

Things to avoid

  • Putting a person's name in a rule, which then sits in your workspace and in any export.
  • Pasting a full statement into a chat to find merchant names. Read the list in the tracker instead.
  • Giving a connected assistant a wider scope than the task needs, or leaving it connected for months.
  • Sharing logins, OTPs or card details with anyone or any tool. A rule does not need them.

A note on screenshots and sharing

People often screenshot their rules list to ask for help online. Rule wording can reveal the names of a landlord, a school or a clinic. Before you post or send a screenshot, replace those names with a placeholder such as the word landlord. The same goes for exports: the categories-and-tax-codes file carries your category setup, so look at it before it travels.

  • Use generic words in rules wherever the description allows.
  • Blur or crop anything that is not the rule itself.
  • Ask yourself who the screenshot will reach after the first recipient.

A last check

After the rule is made, import a month and look at the rows it touched. Count the rows, scan the categories and confirm nothing surprising was caught. If a rule misfires, edit it or delete it. A rule is easy to change, and that is part of why it is a good first step for a privacy-conscious user.

SPENDIXAI JOURNAL

Useful ideas start with a clearer picture of your own spending.

Explore your reports