When you connect ChatGPT to an expense tracker, the assistant can read what your chosen scope allows and that content becomes part of a conversation with a third party. To keep privacy intact, start read-only, ask narrow questions, avoid sensitive categories, and know how to disconnect.
Compare that with copy-paste, where the whole pasted block lands in your chat history whether or not it was needed. A connector can actually be more private if you use a tight scope and ask for totals.
What exactly leaves the tracker?
Whatever the assistant requests within the scope you granted, at the time you ask. In SpendixAI you choose the scope, can begin read-only and can disconnect any time. This is separate from in-app AI categorisation, which sends only a normalised merchant name and an amount band, not your name, account numbers or balances.
A worked comparison
Say you want to know your total on medical spends for the year. With copy-paste you might paste 12 months of transactions, perhaps 1,800 rows, to get one number. With a connector you ask for the category total, and the assistant retrieves the figure it needs. The less you hand over, the less sits in the conversation.
| Approach | Rows exposed to the chat | Answer wanted |
|---|---|---|
| Paste the whole year | about 1,800 | one total |
| Paste only the medical rows | about 40 | one total |
| Connector, narrow question | only what the assistant retrieves | one total |
The row counts are an illustration, not a measurement.
The privacy checklist
- Choose read-only first.
- Keep sensitive rows (medical, loans, family transfers) out of questions you do not need.
- Check the connections list and disconnect idle ones.
- Read the assistant provider's data terms; the tracker cannot control them.
- Use the tracker's export and delete options if you later change your mind.
SpendixAI keeps records in your own workspace with role-based access, says it does not sell personal information, and lets owners export data, choose document retention, delete transactions and schedule deletion of the whole workspace with a seven-day window to change their mind.
Which things should never be in the chat?
- A net-banking login, OTP, PIN or card number. The app never asks for these and does not read SMS.
- Other people's private details you have no right to share.
- Anything you would be uncomfortable seeing in a screenshot.
What can I do if the answer includes more than I wanted?
Tighten the question and, if needed, the scope. Rephrase the request so it asks for a total or a count rather than a list. If the assistant keeps listing named people or vendors you would rather keep private, disconnect, reconnect with a narrower scope, and rewrite the prompt. You can also ask it to refer to categories only.
If you already shared something you regret, disconnect first so nothing more can be read, then delete the chat from your history. Remember that deleting a chat in the assistant does not change anything in the tracker, and the other way round.
Is read-only enough for peace of mind?
It removes the risk of edits, which is the main one, but it does not remove the reading itself. Read-only means the assistant cannot change your records; it can still see what your scope exposes. That is why narrow questions and sensible exclusions matter even with read-only access. Think of it as two separate dials, one for what it can touch and one for what you ask.
Closing routine
When you finish, disconnect, then ask a question to confirm the assistant can no longer answer from your data. Repeat the check monthly if you leave a connection open.
Useful ideas start with a clearer picture of your own spending.
Explore your reports


