I Sell AI.
Yesterday's Outage Still Sent My Team Home.
Yesterday the AI went down, and my office stopped. Not slowed — stopped. People tried again, then packed up and left early. I run an AI company, and an outage I didn't cause still emptied my office. That bothers me more than the lost afternoon.
What actually happened
On 23 June, Anthropic had a major outage — elevated errors across every Claude model and every surface: Claude.ai, the Console, the API, Claude Code, and Cowork. It started late morning US time, peaked at thousands of failure reports, and was resolved a few hours later. It came less than a day after another outage the night before. Anthropic was honest about the cause: demand has grown faster than their infrastructure can carry, especially at peak.
None of that is a scandal. Big systems fail; growing ones fail more. What unsettled me wasn't their outage — it was my reaction to it.
We stopped using AI and started running on it
My team didn't lose a tool yesterday. They lost the tool. Writing, drafting, coding, research, customer replies — it all routes through the same ecosystem, and when that ecosystem blinked, there was no second gear to shift into. We didn't switch to a backup. We didn't drop to a manual pace. We just… stopped.
That's the part worth sitting with. A tool is something you use to do the work. Infrastructure is something the work can't happen without. AI crossed that line in my company and I didn't notice the day it happened — until the day it wasn't there.
The lesson everyone will reach for — and why it's wrong
The easy takeaway is "we're too dependent on AI — use less of it." I don't believe that, and it would be strange coming from me. We're not going to write worse, build slower, and serve customers worse just to feel less exposed. Going backwards isn't resilience; it's decline with extra steps.
Treat AI like a utility, not a miracle
Electricity, internet, payments — we already depend totally on things that occasionally fail. The mature response was never to depend on them less. It was to engineer for the failure: backups, redundancy, and a plan for the hour the power is out. AI is now in that category for a lot of us, and most companies — mine included, at one layer — haven't caught up. Four moves I'm now holding my own company to:
- No single provider. If one model going down stops all your work, you don't have an AI strategy — you have an AI dependency. Be able to route the same work to a different model when one is down. (Ironically, that's exactly how we built our own studio at the infrastructure layer: provider-swappable on purpose. The outage showed our daily-work layer never got the same treatment — caught practising less than I preach.)
- A degraded-mode playbook. Every role should know what "AI is down" looks like for them — what they do by hand, what waits, what they pre-draft for days like this. The goal isn't full speed without AI. It's not stopping.
- Keep the humans capable. The real long-term risk isn't a four-hour outage — it's the skills underneath quietly atrophying because the machine always did them. Use AI to go faster, not to forget how.
- Own your knowledge. If your standards, data and know-how live only inside one vendor's product, you're exposed twice — to their downtime and to their lock-in. Keep the source of truth yours.
So what happens if AI is "gone" one day?
Honestly? For a few hours, it already was — and we failed the test. The company that survives the next one isn't the one that used AI the least. It's the one that could keep moving when its main provider couldn't: switch models, drop to a manual gear, and keep its people sharp enough to carry the load for an afternoon.
We're an AI company. We're going to keep betting on AI. But after yesterday, we're going to bet like a business that knows its most important supplier will, sometimes, not pick up the phone.
The bottom line
Dependence on AI isn't the problem — undefended dependence is. The answer to yesterday isn't less AI. It's a second provider, a backup plan, and a team that doesn't forget how to work without it. Build that, and an outage costs you an afternoon. Skip it, and one day it costs you a lot more.
Published by IMA AI — June 2026. On the 23 June 2026 Anthropic/Claude outage: contemporary reporting from TechRadar, TechTimes and The Next Web, and Anthropic's own status page.