When AI Goes Down, Customer Success Stays Up

When AI Goes Down, Customer Success Stays Up

On September 3, Claude, ChatGPT, and Grok experienced outages during overlapping periods. At first glance, the timing made it tempting to connect them. Was there a shared provider? A routing problem? A data center failure? A broader infrastructure event?

Those were reasonable questions, but they were still questions.

Anthropic reported elevated errors across multiple Claude models and services. Its status updates moved from investigating, to identifying a cause, to deploying a fix, and finally resolving the incident. Reporting later said Anthropic described the disruption as an infrastructure issue. xAI connected Grok's outage to its Memphis compute center. OpenAI's problems occurred later in the same general window. Public reporting did not establish that all three outages shared one cause.

That distinction matters because this is exactly what a Customer Success Manager faces during a major incident: customers want answers before the company has all of them.

Imagine you are the CSM for a large customer whose team depends on Claude throughout the workday. Their users cannot complete normal tasks, leadership is asking what happened, and someone wants to know when the service will be back. Engineering is investigating. Your status page is changing. Social media is filling the gaps with theories.

What does the CSM say?

The answer is not to disappear until Engineering has solved the problem. It is also not to turn the latest theory into a fact.

This is where Customer Success becomes part of the incident-response infrastructure.

The CSM can establish the customer point of contact, understand the business impact, make sure the correct internal teams know the severity, create a consistent communication channel, coordinate updates, and separate what is confirmed from what is still being investigated. Leadership may need one level of detail, technical teams another, and the customer something different. The CSM helps translate between all of them.

During our outage exercise, I kept coming back to four groups surrounding the CSM: the customer, internal technical teams, company leadership, and third parties. Information is moving between all four while pressure increases. The CSM is not repairing servers or rewriting code, but that does not mean the CSM is standing on the sidelines.

When a major customer is deciding whether to stay, the product may be offline, but the relationship is still running.

AI can help here too. An AI Operator could compile status updates, organize a timeline, summarize technical information, compare reports, track open questions, and flag inconsistencies. That can make the CSM faster. But speed creates another responsibility: judgment.

During the September outage, it was easy to develop plausible theories from the timing. Claude and Grok overlapped. Grok had a confirmed Memphis compute-center problem. ChatGPT also experienced issues. That did not make a shared failure a confirmed fact. A CSM repeating an unverified theory to a customer could create a second problem while the technical team is trying to solve the first one.

This is where T.R.U.S.T. becomes practical:
-Talk to the customer.
-Research what is actually known.
-Understand the business impact.
-Serve by keeping communication moving.
-Teach by translating confirmed information into something the customer can use.

An outage will test the technology team, but it also tests the relationship around the technology. Customers remember whether they had to chase for information. They remember whether different departments gave them different answers. They remember whether their business impact was understood.

Customer Success cannot prevent every outage, and it cannot always explain the cause while the incident is happening. But, it can reduce confusion, keep the customer connected to the company, and protect trust while everyone else works on getting the technology back online.

Sometimes the most important thing a CSM can keep running during an outage is the relationship.

Sources and incident record

• Anthropic Status, September 3, 2026 incident history: https://status.anthropic.com/

• The Verge, September 2026 coverage of the overlapping Claude, ChatGPT, and Grok outages.

• WIRED, September 3, 2026 coverage of the unresolved questions surrounding the overlapping outages.