AI service outage

What it is

AI systems run on a stack: the model endpoint, the retrieval index, the orchestration layer, the provider's platform and the network between them. Any of these can fail. Hosted models add dependencies the organisation does not control, including the provider's capacity, rate limits and regional availability. Because many firms use the same few providers, an outage tends to be shared. The risk is unavailability, as distinct from a model that is running badly, which is covered in AI model drift and degradation.

How it happens

• The provider suffers an incident or throttles requests during peak demand.
• A regional failure removes the only deployment the organisation uses.
• A supporting component, such as the vector store, fails.
• Credentials expire or a quota is exhausted.

Why it matters

If the model sits in the path of a customer journey or a control, that journey or control stops. Queues build, customers are turned away, and where the model performs a check such as fraud screening, the choice is between halting the process and running it unchecked.

Warning signs

• A single endpoint in a single region serves a critical process.
• Nobody has asked what the process does when the model does not answer.
• Rate limit errors appear at peak times.
• The provider's availability history has not been reviewed.

Who owns it

The system owner is accountable for availability. Engineering designs redundancy. The owner of the business process decides what happens when the system is down.

In practice

A retailer's customer contact runs entirely through an assistant built on one hosted model. The provider has a four-hour outage on the busiest day of the year. There is no queue to a human team and no static fallback. The same pattern generalises to any process that was redesigned around a model without asking what happens in its absence.