I had a read-only audit run over my own mailbox looking for recurring payments. The reasoning is simple: signing up is a decision you make once; continuing to pay is a decision you make never. Money leaves on a schedule nobody re-decides, and the mailbox is the only place that remembers every one of those sign-ups — so that is where you go to get the standing costs back in front of you.
The detection rule was the easy part. A payment counts as recurring when repeated charges arrive from the same sender at a regular monthly or yearly interval, when a mail carries explicit renewal language, or when invoices form a sequential series. Variable amounts on a fixed schedule count too; pay-per-use receipts do not.
The interesting part was not the list at the end — it was what the search missed, because I made the tool test its blind spots instead of assuming them. Some billers never send receipts at all and only surface through price-change notices, so those mails became first-class evidence rather than noise. Attachments proved unreadable through the mail connector — tested, not guessed. An address I had written off as dead was still receiving bills.
Two rules survived contact. The report must state what it did not search, because the dangerous failure is believing you are covered. And read-only is enforced by never granting the write tools — not by instructing the tool to behave.