The Second Vendor Nobody Was Actually Calling

# The Second Vendor Nobody Was Actually Calling Auditing an application's AI integration, I found it was configured against two separate model vendors. On paper this read as redundancy — a sensible failover story, one vendor down doesn't mean the feature is down. In practice, every single active code path routed through one vendor, and the second one existed only as dead configuration wired to an invalid credential that nothing had noticed because nothing was calling it. ## How a dead path hides A dead integration doesn't announce itself. It doesn't error, because nothing invokes it. It doesn't show up in a dependency audit, because the package is legitimately imported somewhere. It survives code review after code review because reviewers are looking at whether the diff in front of them is correct, not whether an entirely separate, unrelated vendor integration three files away is still load-bearing. The only way it surfaces is someone actually tracing which code paths are live versus which ones merely compile. ## What finding it is worth This is exactly the kind of finding that a "does it build, does it pass CI" review will never catch, because a dead path with a bad credential builds fine and passes every test that doesn't specifically exercise it. It only shows up under the question a real audit asks: for every integration this system claims to have, is anything actually calling it, and does the answer match what the documentation or the team's mental model says it should be. ## The fix, and the actual lesson The fix was mechanical — delete the dead path, consolidate onto the one vendor actually in use, remove the invalid credential from configuration. The lesson is upstream of the fix: an unused integration with a broken credential is a liability sitting quietly until someone tries to use it in an emergency and discovers, at the worst possible time, that the failover they were counting on was never actually wired up. Auditing what's real versus what merely exists in configuration is cheap. Discovering the gap during an actual outage is not.