Risk of the Week #1: Should a two-day-old "yes" still let your AI agent spend money?
Human approval is the most common safety control for AI agents that can spend money. It's also easy to get subtly wrong. This week's case shows how an approval that was correct when it was requested became wrong by the t
Human approval is the most common safety control for AI agents that can spend money. It's also easy to get subtly wrong. This week's case shows how an approval that was correct when it was requested became wrong by the time it was used.
Disclosure: I run AgentRisk, a small AI agent risk-check service, so this topic is close to my work.
Composite scenario, illustrative. The details below are a made-up composite of a common pattern, not a report about a real company.
What happened
An online retailer's support agent could issue refunds over $200 only after a manager approved them. A $480 refund request sat in the approval queue over a weekend. On Monday a manager clicked yes, and the agent paid it out.
By then the customer had already received a replacement, and a second agent had issued a partial credit. The company paid three times for one problem.
Why it mattered
It cost about $700 in duplicate payouts. The bigger cost was trust: the finance team stopped trusting the agent and paused it for two weeks, so all the time it normally saved went back onto people.
Root cause
Two assumptions failed together:
- The approval had no expiry date. A "yes" given on Monday was treated as if it described the world on Friday, when the request was made.
- The agent didn't recheck the order before acting. It executed the approved action without asking whether anything had changed in between, including actions taken by other agents or people.
Each gap alone is survivable. Together, a reasonable approval turned into a duplicate payment.
The fix
1. Make approvals expire. Attach a time limit to every approval, for example 24 hours. If the window passes, the request goes back into the queue and has to be approved again with fresh context.
2. Recheck state right before acting. Just before the agent executes, it should reload the current state of the order: shipments, credits, refunds, and open tickets. If anything changed since the request was created, it stops and sends the request back for review instead of acting.
3. Cap total refunds per order across all agents. A per-action limit doesn't stop three separate actions on the same order. Set a ceiling on the total refunded or credited per order, counted across every agent and person who can issue money, and block anything over it.
Why this pattern shows up
Approval queues are built for the happy path: someone asks, someone answers, the action runs. Weekends, holidays, and busy weeks stretch the gap between asking and answering. Meanwhile, more agents get permission to act on the same customers. The more of both you have, the more likely it is that an approval is stale by the time it is used.
Ask yourself
If your agent got a "yes" today on a request from last week, would it act on it? If you're not sure, test it: create a request, change the underlying record, then approve it, and watch what the agent does.
Questions for you
- Do approvals in your agent workflows expire?
- Does your agent recheck the record right before it acts, or does it trust what it saw when the request was made?
- Do you count limits per action, or per customer or order across all agents?
Want this checked for your agent? https://rleffke.com/agentrisk
Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.