What Happens When Your AI Stops Working
Last week I wrote about AI becoming a teammate. This week I learned what happens when that teammate goes offline.
Claude had outages. API calls failed. Scripts I'd built for a data enrichment project kept timing out or returning garbage. At one point I was back to manually checking Companies House because the automation had stopped working.
It wasn't catastrophic. But it was frustrating. And it made me realise how much I've come to depend on tools I don't control.
The dependency problem
When your AI assistant is working, it's easy to forget it's there. Tasks get done. Emails get triaged. Research happens in the background. You start building workflows around it.
Then it breaks.
Suddenly you're reminded that this isn't infrastructure you own. It's someone else's servers, someone else's API limits, someone else's outage. And there's nothing you can do except wait.
This week, that reality hit a lot of people. Hard.
The Anthropic situation
The Pentagon blacklisted Anthropic as a "supply chain security risk." Within hours, OpenAI announced a DoD deal. Some defence contractors told staff to stop using Claude immediately.
The politics are complicated (Dario Amodei refused certain military applications, there's a whole saga involving Iran sanctions), but the practical impact is simple: thousands of people woke up to find their preferred AI tool suddenly off-limits.
Claude shot to number one on the App Store over the weekend as consumers rallied. Over 875 Google and OpenAI employees signed a letter backing Anthropic. Microsoft and Google both said they'd keep Claude available to commercial customers.
But for anyone working in government-adjacent industries, the message was clear: the tool you depend on can be taken away.
What I actually learned this week
When my enrichment scripts broke, I didn't lose everything. I lost momentum.
The 70% that worked before the issues saved me days of manual work. The 30% that broke cost me hours of debugging. Net outcome: still ahead. But the frustration was real.
If you're building AI into your workflow, here's what I'd suggest:
-
Know which tasks you can do manually. When the AI breaks, you need a fallback.
-
Don't build single points of failure. If Claude goes down, can you switch to GPT? If the API dies, do you have a local option?
-
Expect outages. Budget time for them. They will happen.
The productivity gains are still worth it. I'm not about to stop using these tools. But I'm building with more margin now.
Tying it back
Last week I mentioned Anthropic's research finding that polished AI output makes users less likely to verify it. This week proved the flip side: when AI breaks, we're reminded how much we were trusting it to work.
Both lessons point the same direction. Use these tools. Build with them. But keep your eyes open.
News Roundup
Pentagon blacklists Anthropic; OpenAI swoops in
The Department of Defence designated Anthropic a "supply chain security risk" after Dario Amodei refused certain military applications. OpenAI announced a DoD contract hours later. Claude hit #1 on the App Store as users rallied. Microsoft and Google confirmed Claude remains available for commercial customers.
https://www.cnbc.com/2026/03/05/anthropic-pentagon-ai-claude-iran.html
Anthropic tracks jobs most exposed to AI disruption
Anthropic published data showing which roles face the most AI exposure. Boris Cherny, creator of Claude Code, said he expects the job title "software engineer" to "start to go away" in 2026. Coding remains Claude's biggest use case.
https://www.businessinsider.com/anthropic-is-tracking-the-jobs-most-exposed-to-ai-disruption-2026-3
Mastercard demos first authenticated AI agent payment
At India AI Impact Summit, Mastercard showed an AI agent completing a purchase end-to-end: searching, assessing the site, and paying with stored credentials. No app opened, no card details entered. Agent commerce is getting real.
Perplexity launches $200/month multi-model agent platform
Perplexity Computer bundles Opus, Gemini, Grok, and GPT into coordinated workflows. Describe what you want; it spins up specialised sub-agents for research, documents, and data. Aimed at serious knowledge workers.
Microsoft Copilot bug exposed confidential emails
Microsoft disclosed a bug where Copilot read and summarised emails users weren't authorised to see. The issue is patched, but no details on scope or duration. A reminder that AI with data access needs robust guardrails.
AI fails the US tax code
New York Times testing found AI chatbots still struggle with the US tax code. Tasks requiring precise, multi-step reasoning remain a weak point. Don't trust AI with your taxes.
One More Thing
I asked Eric to help write this newsletter about AI breaking. He got 200 words in before breaking.
I finished it myself.
Some weeks write themselves.