On Sep 3 2026, I was reminded of something we tend to forget when a technology becomes too useful.

It can still disappear.

ChatGPT, Claude and Grok experienced overlapping service disruptions that day. OpenAI confirmed elevated errors across ChatGPT and Codex. Anthropic reported elevated errors across several Claude models. Grok also experienced an outage, which xAI later attributed to a problem at its Memphis compute centre.

For most people, it was an irritation. For a founder who has quietly built half the business around AI, it can be something much more serious. And that is the part I found interesting.

The problem isn’t that AI went down.

The problem is when your business goes down with it.

Think about a typical lean startup today – The founder uses ChatGPT to structure the investor update. The marketer uses Claude to develop campaign ideas. The developer uses an AI coding assistant to debug a feature. Customer support uses an AI layer to handle first-level queries. The sales team uses AI to research prospects. The operations team uses it to analyse data and prepare reports.

And somewhere in the middle of all this, the founder is probably using AI to think through the next decision.

None of these activities necessarily looks critical on its own. But put them together and something interesting happens:

AI stops being a tool and starts becoming infrastructure.

That distinction matters. Because when a tool goes down, you find another tool. When infrastructure goes down, you need a fallback.

---

I think many startups are making the same mistake with AI that businesses made with cloud software years ago.

We are optimising for productivity, but not yet for resilience.

We ask:

“Which AI model is smartest?”

“Which one gives me the best output?”

“Which one is cheapest?”

“Should I use ChatGPT or Claude?”

Those are useful questions.

But there is another question that I think founders should start asking:

“What happens to my business if this disappears tomorrow?”

Not permanently. Just for three hours. That is enough.

Imagine your customer-support AI stops responding during your busiest period. Or your developer is in the middle of a production issue and their AI coding workflow disappears. Or your marketing team has a campaign going live at 6 PM and the entire content workflow depends on an AI platform that suddenly isn’t responding. Or you’re a solo founder preparing an investor presentation at 11 PM and the conversation containing three hours of work simply isn’t accessible.

The individual outage may be temporary.

The business disruption isn’t necessarily temporary.

---

The interesting part about September 3 wasn’t that three AI companies had problems.

It was the reminder that AI dependency has become systemic.

The three outages happened within a relatively tight window, which naturally led to speculation about whether there was a common infrastructure issue.

But the more useful lesson doesn’t actually require knowing whether there was one common technical cause.

In fact, the providers’ own explanations point to different issues.

OpenAI described its ChatGPT/Codex incident as a routing error. xAI later said Grok’s disruption came from an outage at its Memphis compute centre. Anthropic reported an infrastructure issue but did not establish the same shared cause.

That’s actually more interesting to me.

Because whether the failure happens at the AI model, the network, the data centre, the API, the cloud layer or somewhere in between doesn’t matter to the founder sitting in front of a frozen screen.

Your dependency has failed.

And your customer doesn’t care whose infrastructure failed either.

---

So what should a founder actually do?

I don’t think the answer is to reduce your use of AI.

Quite the opposite. AI is one of the biggest leverage opportunities available to a small team today.

The answer is to stop treating your AI stack as something that is always available. Here are the five things I would build into an AI-dependent business.

1. Have a second AI route

If your entire workflow depends on one provider, you have created a single point of failure.

You don’t necessarily need to maintain two identical systems.

But your team should know:

“If this doesn’t work, we go here.”

It could be another major model, a specialised AI tool, an API from another provider, or even a smaller local model for certain tasks.

The objective isn’t redundancy for the sake of redundancy.

It’s continuity.

---

2. Don’t let the AI hold the only copy of your work

This one sounds obvious.

Yet I suspect many of us are guilty of it.

We have a brilliant conversation with an AI.

We refine a strategy for two hours.

We build a campaign.

We debug a piece of code.

And then we treat the AI conversation itself as the repository.

That’s risky.

Important outputs should move into your actual systems of record:

Drive. Notion. GitHub. CRM. Project management software. Your own database.

AI should help create the work.

It shouldn’t be the only place where the work exists.

---

3. Create a “no-AI” version of critical workflows

This is perhaps the most important one.

For every AI-dependent business process, ask: If AI disappeared for four hours, could my team still operate?

Not as efficiently. Not as elegantly. Just operate.

For customer support, perhaps there’s a human escalation route. For sales, perhaps the CRM still contains the necessary customer information. For development, perhaps the team can still access the codebase, documentation and deployment process. For marketing, perhaps campaign assets and calendars aren’t sitting exclusively inside an AI conversation.

You don’t need to eliminate AI.

You need to ensure that AI failure doesn’t equal operational failure.

---

4. Teach your team to diagnose dependency failures quickly

There’s another hidden cost of outages:

people waste time assuming the problem is theirs.

“Is my Wi-Fi broken?”

“Is the API key expired?”

“Did I break something?”

“Why isn’t this prompt working?”

Before spending 30 minutes debugging your own system, check whether the provider is having an incident.

OpenAI, Anthropic and other providers publish status information, while services such as Downdetector can provide an additional signal from user reports.

That simple habit can save a surprising amount of time.

---

5. Stop putting AI-dependent work at the edge of the deadline

This is a management problem, not a technology problem.

If a campaign must go live at 6 PM and your team plans to use AI at 5:45 PM to generate half the assets, you’re effectively making the AI provider part of your deadline.

That’s unnecessary risk.

If AI saves you two hours, don’t use those two hours to move the deadline closer.

Use some of them as resilience buffer.

The productivity gain should create breathing room, not tighter dependency.

---

There’s a bigger lesson here.

I’ve spent a lot of time thinking about how AI changes the way businesses operate.

And increasingly, I think the competitive advantage won’t simply come from using AI more.

It will come from designing the organisation around AI intelligently.

There is a difference.

A company where everyone has access to ChatGPT is not necessarily an AI-first company.

A company where AI is embedded into processes, data flows, decision-making and customer journeys may be.

But an organisation where all of those processes collapse when one AI service goes down?

That’s not AI-first.

That’s AI-dependent without resilience.

And I think we’re going to see more businesses discover this distinction over the next few years.

---

My takeaway

I don’t see the ChatGPT-Claude-Grok outages as an argument against AI.

I see them as an argument for better operating design.

We spent decades teaching businesses to have backup servers, backup suppliers, disaster-recovery plans and business-continuity processes.

Now we’re adding another layer to that conversation:

AI continuity.

So the next time you’re evaluating an AI tool for your startup, don’t ask only:

“How much productivity will this give us?”

Also ask:

“How badly will we suffer if we lose it?”

That second question may tell you more about whether you’ve built an AI-enabled business…

or simply built a business that happens to use AI.

And there is a big difference between the two.