Asking an AI assistant what you spent on travel last quarter means letting it read some of your transactions, so it is worth deciding what it may see before you type the question. With SpendixAI you choose a scope when you connect Claude or ChatGPT, you can start read-only, and you can disconnect at any time. The safest habit is to ask for totals first and details only when you need them.
What actually leaves your workspace?
There are two different things going on and people mix them up. The first is categorisation, which works in stages: your own rules, then your history and similar merchants, and only if AI is switched on an AI model for the leftovers. For that step the model gets the normalised merchant name and an amount band, not your name, your account numbers, the full statement text or balances.
The second is when you connect an assistant and ask it a question. Whatever the assistant reads to answer you becomes part of that conversation, handled under the assistant provider's own terms. So read those terms and the connection screen, not just ours.
The questions, from least to most revealing
| What you ask | What the assistant has to read | Exposure |
|---|---|---|
| Total travel spend for the quarter | Travel rows, or a summary of them | Low |
| Month-wise total by sub-category | Same rows, grouped | Low to medium |
| List every travel transaction | Dates, merchants, amounts | Medium |
| Which trips were with whom, and where | Merchant names and patterns | Higher |
Start at the top. If a total answers your need, stop there. You can always go down a row if the number looks wrong.
A worked example
Say the January to March travel is in three trips: ₹8,200 for a train and stay, ₹6,450 for a cab-heavy week, and ₹4,250 for a short flight. The total is ₹8,200 plus ₹6,450 plus ₹4,250, which is ₹18,900. You can ask for just that figure, and then ask for the three subtotals. You never needed to ask for the merchant names, so you never exposed them.
Settings worth checking first
- Scope. Choose the narrowest one that does the job, and begin with read-only.
- AI controls inside the app. AI can be switched off per feature, given a monthly budget, or turned off entirely; then unmatched items wait for review.
- Document-processing features. Optional features such as reading an unfamiliar statement layout process the uploaded document, so leave them off if you want less exposure.
- Retention. Owners can choose document retention and delete transactions.
Things we would not do
- Paste whole statements into a chat window. A scoped connection is more controlled than a pasted file.
- Share bank logins, card numbers or one-time codes with anything. SpendixAI does not ask for net-banking logins, and nothing in this flow needs them.
- Leave the connection on a shared or office computer.
If you change your mind
Deleted transactions go to a recycle bin for 30 days. Owners can export their data, and can schedule deletion of the entire workspace with a seven-day window to reconsider. Those are your exits. For the conversation with the assistant, check what its settings let you delete.
A realistic scenario
Say you want to know whether a planned trip is affordable, and your statements also contain transfers to a flatmate and a few friends. A question about the quarter's travel total never needs those transfers. A request to 'list all UPI payments' would pull in names of people who never agreed to be part of your question. Keep the request to the category and the dates. If a person's name must appear, such as a cab driver paid by UPI, remember it is someone else's information too. Where you can, ask for merchant-level rows and leave person-to-person transfers out of the request altogether.
A short checklist
- Name the narrowest scope that works.
- Ask for the total first.
- Look at the rows only if the total surprises you.
- Disconnect, and note the date.
Useful ideas start with a clearer picture of your own spending.
Explore your reports
