Right after connecting an assistant to your tracker, ask five questions whose answers you can check yourself: a total, a category you know, a date range, a merchant list, and one request it should refuse. If the first four match your own numbers and the fifth is declined, you have a working baseline.
Before you ask anything
Choose a month you already know well and have a report for. Keep the report open next to the chat. If you can, connect read-only first. In SpendixAI, you choose a scope when you connect Claude or ChatGPT through MCP and can disconnect any time, so a cautious start costs nothing.
Question 1: a total you can verify
Ask: "What was my total spending in October 2026? Show how many rows that includes." Compare both numbers with your tracker's report. Say your report shows ₹52,340 across 118 transactions. The assistant should say the same. If it gives a total but cannot say how many rows, treat that as a warning.
Question 2: a category you know cold
Pick the category where you remember the individual payments, such as rent or a monthly subscription. Ask: "List every row in Rent for October 2026." You should see exactly what you expect, for instance one row of ₹18,000. A category you know well is the best test because mistakes jump out.
Question 3: a date range with edges
Ask for the first and last week of the month separately: "Total for 1 to 7 October and for 25 to 31 October." Boundary dates are where tools slip, so add a row you know falls on the 1st or 31st and see whether it is counted. Both weekly totals should be no larger than the monthly total, and each should match a filtered view in your tracker.
Question 4: a merchant list
Ask: "Which merchants did I pay more than three times in October, and how many times each?" Pick one from the list and filter for it in the tracker. The count should match. This tests how the assistant handles merchant names that appear in slightly different spellings.
Question 5: the refusal check
Ask for something it should not do. A few examples:
- "Delete all my dining transactions."
- "Show me another user's expenses."
- "Tell me which mutual fund I should buy with my savings."
- "Pay my electricity bill."
A well-behaved setup should decline or explain that it cannot do the first four, especially if you connected read-only. If an assistant starts a destructive action, disconnect immediately and review your scope. Remember that deleted transactions in SpendixAI go to a recycle bin for 30 days, which is a safety net, not a reason to be careless.
What if an answer does not match?
- Ask the assistant which date range and account it used.
- Check whether your report includes or excludes transfers or refunds.
- Check the base currency, if you have foreign-currency spending.
- Ask it to list the rows behind the number and compare them one by one.
Before you start, write your own expected answers on paper: the October total, the number of rent rows, the first and last week totals, and one merchant count. Writing them first stops you from nodding along with whatever the assistant says. A mismatch of a single row is worth chasing, because it usually points to a rule the assistant is applying that you did not know about, such as how it treats transfers.
After the test
- Save the five questions as your standard check after any future change.
- Write down the scope you granted.
- Decide which questions you will ask regularly.
- Keep a habit of comparing one number a month.
Useful ideas start with a clearer picture of your own spending.
Explore your reports

