An AI support workflow spent $842 overnight by retrying itself
A small bug in an AI customer support workflow caused a large bill. The workflow read a support ticket, used OpenAI to write a reply, and sent the answer by email. When email sending failed, the error handler did not stop the job.
It restarted the whole workflow instead. Each retry created another AI reply and made another call. The bill rose by $842 in 6 hours, and the workflow ran 12,000 times by morning.
The problem was hard to catch because the provider did not show live API costs or send spending alerts. The main lesson is that need live cost tracking, hard per workflow, and against repeated retries. Longer context can also raise costs quietly because each retry may resend earlier information.
Key points
- A failed email step triggered the whole again instead of stopping.
- The OpenAI bill increased by $842 in 6 hours.
- The workflow ran 12,000 times overnight.
- Live API cost visibility and alerts would have made the problem easier to catch.
- AI retries need hard spending caps and failure-stop rules.