What Happens If Your AI Supplier Loses Its AI Supplier?

Your direct supplier may look independent while depending on another company’s model for a critical part of its service. That dependency belongs in your continuity assessment even when you have no contract with the underlying provider.

Consider an outage, a retired model, an access restriction or a substantial increase in inference cost. These are scenarios to plan for, rather than predictions about a particular company. The immediate question is what the supplier can continue to deliver while it adapts.

Trace the dependency

Map the service through the agent platform, model provider, hosting infrastructure and material tool connections. Ask which components are essential and which have tested alternatives. Two model brands may still share a cloud dependency or another common point of failure.

The CMA’s foundation model update examines competition risks and connections in the foundation model market. For a customer, this provides context for asking how concentration could affect choice and dependency. It is not evidence that any named supplier is about to fail.

“We support multiple models” is a starting point

Changing an endpoint is only one part of migration. A replacement model can behave differently with the same instructions, tool descriptions or ambiguous inputs. Retrieval systems may require new embeddings and rebuilt indexes. Workflows need evaluation against known tasks and failure cases.

An AI Dependency Exit Plan should cover:

• Model performance, safety, latency and cost requirements.
• Exportable data, configuration, prompts and evaluation assets.
• Retrieval and embedding migration, where applicable.
• Tool schemas, permissions, identities and approval controls.
• Alternative capacity and recovery objectives.
• Safe degradation to conventional software or qualified people.

The most useful evidence is a recent migration exercise. Ask what failed, how long recovery took and which service commitments could not be met.

An alternative is credible when the supplier has shown that it works under realistic conditions.

Stress the economics too

As a planning exercise, test variable model costs at two, five and ten times the current level. These are deliberately chosen scenarios, not forecasts. Establish what they would mean for price, margin, usage limits and continuity.

A system that is technically portable may still be commercially dependent on one provider’s prices. A lower cost alternative that materially reduces quality may not preserve the service the customer bought.

Plan for the transition

Agree who makes the migration decision, how customers are informed and what triggers reassessment of security or data flows. Include transition assistance and access to the information needed to continue elsewhere.

Resilience improves when the supplier can explain both its normal operation and its fallback. Ask the question plainly: if the model underneath your product became unavailable, what would we still receive, when would full service return, and what evidence supports that answer?

References

• CMA: AI foundation models update paper, April 2024. Market context; the exit plan and stress scenarios are this publication’s recommendations.