๐๐ก๐ ๐ช๐ฎ๐๐ฌ๐ญ๐ข๐จ๐ง ๐ข๐ฌ ๐ง๐จ ๐ฅ๐จ๐ง๐ ๐๐ซ “๐ข๐ฌ ๐ญ๐ก๐ ๐๐ ๐ฎ๐ฉ?” ๐๐ญ ๐ข๐ฌ “๐ข๐ฌ ๐ญ๐ก๐ ๐๐ ๐ฌ๐ญ๐ข๐ฅ๐ฅ ๐๐จ๐ข๐ง๐ ๐ฐ๐ก๐๐ญ ๐ข๐ญ ๐ฎ๐ฌ๐๐ ๐ญ๐จ ๐๐จ?”
Last week, Anthropic published a rare engineering postmortem: Claude Code had been measurably worse for weeks, and nobody on the customer side could see it. Same model name. Same API endpoint. Same “uptime green” status page. Behavioural drop, no observability.
That is the story of vendor-managed AI in April 2026.
This is not a stable supply chain. It is a release-cycle ecosystem that occasionally introduces silent regressions, sandbox escapes, and prompt-injection paths.
For consumer products, that is a productivity headache. For aviation, it is an operational and safety question.
- An AODB that depends on a third-party model can lose a percentage of decision quality without the vendor announcing it.
- A turnaround prediction that started 87% accurate can drop to 78% across a deploy that does not ship to your tenant.
- A regulator asking “what was the AI behaviour on the day of the disruption” cannot be answered by a status page.
The question is no longer “is the AI up?” It is “is the AI still doing what it used to do?”
That is behavioural monitoring, not availability monitoring. And it is the new bar.
AirportLabs built the entire stack, including the AI, in-house for exactly this reason.
Gate allocation, stand assignment, tow planning, disruption recovery. The models live inside SkyCore and Allegra, on infrastructure we operate. We can replay any decision against any historical state. We can detect drift before a customer feels it. We can answer the regulator with a timestamped audit trail, not a vendor SLA.
It is not a marketing line. It is the operational doctrine that lets us look at an aviation customer and say: when Anthropic, OpenAI, or anyone else has a quiet bad week, your operation does not.
Buy the AI that lives inside the product. Be careful with the AI that lives next to it.