AI Won’t Find Your Bug. It’ll Confirm Your Hunch.

We brought Claude into our first incident war room — the team huddled under pressure on a live incident — to help find what was taking down our payments. It helped. Not on its own, and not the way I expected. That was the best lesson of the night.

It verifies, it doesn’t discover

The incident was a payment service falling over, and the root cause wasn’t in the payment service at all. It sat in the way several connected services were behaving together, one of them in a different project and a different account entirely. We weren’t debugging one thing. We were debugging the space between half a dozen.

We did the part that needed judgement: looking across the services and forming leads about where the fault might sit. Then the groundwork began — chasing each lead through the logs, pulling records from one service after another and reading them by hand. Hours of manual labour, and exactly what we handed over. Claude scanned the log records far faster than any of us could, pulled them together, and gave back a summary. One of our leads turned out to be true. We narrowed in on it and built a fix, and the outage cleared.

Here’s the split, distilled:

Works:  "Here are three theories for the payment failure. Which does the evidence back?"
Works:  "I think the fault is upstream, not in payments. Chase the logs and confirm it."

Fails:  "Go find the problem."   → a confident tangent through the logs
Fails:  "Just fix it."           → if you can't name the symptoms, neither can it

Give it a hypothesis and it has something to test against. It tears through logs and metrics far faster than any of us could at 2am, and comes back with a clean confirm or kill. Give it a blank “go find it” and it has nothing to aim at, so it manufactures a direction and pursues it with total confidence, which is the worst possible failure mode when you’re already under pressure.

“Just fix it” fails for the same reason. If you can’t name the symptoms, neither can it. The naming is the thinking, and the thinking is still yours.

The part we missed

Here’s the honest bit. We fixed the bug that night and went home. What we found the next day was that the outage had left a trail: other services sitting in states they should never have reached, quietly wrong. We had fixed the record that failed. We had not fixed everything the failure touched.

That was a second investigation we never ran. The same thing that chased our leads through the logs could have swept the neighbouring services for the ones that no longer looked healthy. Finding the root cause is one search. Finding everything it damaged is another — and AI is just as good at the second, if you remember to ask.

The pattern held all night, and into the next day once we saw the full picture. AI didn’t replace the thinking. It absorbed the grind.

Bring it a hypothesis, not a mystery. It’s the fastest verifier on the team and a terrible detective.