FAA ground stop: the AI tool got blamed first, then cleared

The AI tool was the first suspect. The 1980s copper was the cause. Friday's nationwide FAA ground stop halted U.S. departures for more than four hours. By midday, federal officials said they were checking whether the FAA's new AI traffic-planning tool had a part in it. By evening they had cleared it and blamed aging telecom switches, GeekWire reported. There's a lesson in that for anyone who puts AI into a live operation.

What actually happened

• The failure. Transportation Secretary Sean Duffy said facilities began to see unstable radar and communications data at 12:36 p.m. Pacific. Old switches and routers "got overloaded with data." The FAA reopened the airspace at 5:15 p.m. Pacific. More than 3,500 planes were in the air when the stop began.
• The suspect. SMART is the FAA's new AI system. It pulls in about 200 data streams and recommends schedules and routes. It's part of a 12-year, $875 million contract with Air Space Intelligence. The FAA said on Tuesday that it would expand SMART to all 22 en route centers. Three days later, the network fell over. Homeland Security Secretary Markwayne Mullin said CISA was looking at it.
• The verdict. Duffy said SMART isn't connected to the core systems controllers use to track flights. "There was no possible way SMART would have impacted what happened today." The vendor denied a link too.
• The trigger. Mullin said the overload appeared to come from new equipment and patches the FAA was installing. The new gear overwhelmed the old gear. Duffy said the new IP switches stayed up.

Mullin also said nothing was "completely certain" yet. This is a first-day account.

Why it matters for your business

When something breaks the week you ship AI, the AI gets the blame. Customers, staff and your own boss will make that jump. The FAA could rule SMART out in hours for two reasons, and you can copy both.

1. Keep AI off the critical path. SMART recommends. FAA specialists review, and local facilities decide. It doesn't touch the systems that keep planes apart. If your AI agent drafts the quote, a person sends it. If it's down, the business still runs.
2. Be able to prove what it did. Log every agent action with a timestamp. When the outage call starts, you want an answer in minutes, not a debate.
3. Watch the migration, not the model. The real cause was new equipment that flooded old equipment. That's the classic cutover failure. Stage it. Load test the old side. Have a rollback.
4. Change one thing at a time. A big launch on Tuesday and network patches on Friday made two suspects. Space your changes so one failure has one cause.

Sources: GeekWire, News4JAX.