Why your agent beats the bank at this
Bank fraud algorithms look at cohort behavior: is this transaction unusual compared to other customers' patterns? Your agent looks at your behavior specifically: is this transaction unusual compared to your past transactions?
The agent loses on speed (it doesn't run until you ask) but wins on specificity. Your bank will miss a fraudulent $30 charge at a coffee shop because that amount is common. Your agent knows you don't buy coffee in that city.
The weekly prompt
“Scan my transactions from the last 7 days. Flag anything that: (1) comes from a merchant I haven't used before, (2) is an unusual amount for a merchant I do use, (3) looks like a duplicate charge from the same vendor within a day, or (4) is from a location that doesn't match my other recent activity. Use your judgment.”
The agent calls list_transactions for the 7-day window, then compares against its baseline understanding of your spending from the 6-month context it built up during earlier tool calls in the session. If there's nothing suspicious, it says so. If there is, it surfaces each flag with context on why.
What it reliably catches
- Duplicate charges from a vendor within 24 hours (the classic “card was swiped twice” error).
- Brand-new merchants you've never transacted with before.
- A $300 charge at a merchant where you usually spend $30 (or vice versa).
- Small “test charges” (<$5) that sometimes precede fraud — someone testing a stolen card.
- A concentration of new charges in a short time window (card compromise in progress).
What it doesn’t catch
- Highly sophisticated fraud that mimics your patterns (rare for consumers, more common for business cards).
- Fraud on cards you haven't connected. Obviously. But if you have multiple cards, this is a good prompt to connect them all.
- Transaction authorization after the fact. The agent can only see posted transactions; pending ones sometimes show up late depending on the aggregator.
- ACH fraud sometimes, because ACH descriptors are especially hard to pattern-match.
Make it a Friday habit
Sunday evening or Friday afternoon, whichever fits your routine. Claude Code users can save the prompt as /fraud-detective (one of the six skills in our public skills repo). Other hosts: save to your agent's snippet system and trigger weekly.
If the agent ever flags something real, you'll save yourself a week of back-and-forth with your bank. If it never does, five minutes of peace of mind.