An enterprise can approve an AI system, validate its outputs, restrict its permissions, and still overlook a dependency that could undermine the entire business case: the economics of the supplier delivering it.
The Economist’s headline, “The AI industry’s gigantic debts are making investors nervous,” raises a question that deserves attention beyond financial markets. What happens to an organization’s operations when the AI capabilities it depends on become more expensive, less available, or commercially unsustainable?
A supplier’s debt does not establish that it will fail. Nor does investor anxiety establish an imminent disruption. But enterprise leaders should examine how a change in supplier economics could reach their own costs, commitments, and critical services.
AI governance should include evidence that the business can withstand a material change in its AI supply chain.
How supplier economics become an enterprise dependency
Consider an illustrative insurance company that integrates an external AI service into claims intake. The system summarizes documents, identifies missing information, and prepares cases for human review. Its initial business case assumes a particular price, processing capacity, and level of service.
Now stress those assumptions. Suppose prices rise materially at renewal, access to the preferred model is withdrawn, or capacity becomes constrained. None of these scenarios requires the supplier to become insolvent.
The insurer may still have a technically functional system. Yet its original cost assumptions may no longer hold. Its human review team may lack the capacity to absorb a backlog. Its alternative model may perform differently on the documents that matter most.
The risk crosses organizational boundaries. Procurement owns the contract. Technology owns the integration. Operations owns the queue. Finance owns the expected savings. Risk owns the assessment.
Who owns the decision to continue when the assumptions supporting approval change?
That question should have an answer before the service becomes indispensable.
Model assurance must connect to business continuity
Accuracy, privacy, security, fairness, and human oversight remain essential. They should sit alongside a second set of questions about commercial and operational dependence.
Can the organization keep delivering its critical service if the model changes? Can it afford the service under a materially different pricing structure? Can it recover its information and operating context? Can it replace the supplier within a time frame the business can tolerate?
These questions require more than a model inventory. They require a map connecting AI services to business capabilities, contractual commitments, data flows, operational owners, and recovery arrangements.
An inventory may tell an executive that the company uses several AI vendors. A dependency map may reveal that supposedly separate services rely on the same underlying model or infrastructure provider.
Vendor count alone does not demonstrate resilience. The organization needs to understand which failures or commercial changes could affect several services at once.
Five questions boards should require management to answer
1. Which critical outcomes depend on each AI supplier?
Begin with the business service: processing a claim, responding to a customer, reviewing a transaction, or supporting an employee. Identify the consequence of degraded AI support and the maximum disruption the service can tolerate. Assign an accountable business owner.
2. Which economic assumptions could invalidate the business case?
Document the assumptions behind expected value: usage, price, review effort, error handling, integration support, and switching costs. Test a range of plausible changes. A scenario is a planning tool, not a forecast of what a particular supplier will do.
3. What can the enterprise actually recover and move?
Data export is only part of the answer. The organization may also need retrieval configurations, evaluation datasets, workflow definitions, permissions, decision records, and integration documentation. Procurement and architecture teams should establish what is contractually available and technically usable.
4. Has the fallback been demonstrated?
A second supplier listed in a presentation is an option to investigate. A tested fallback has evidence: task performance, capacity, security review, migration effort, and a clear activation procedure. Where a human fallback is proposed, its staffing and throughput must be credible.
5. What change triggers a new approval decision?
Define thresholds for reassessment. These might include a material price change, model retirement, repeated service degradation, acquisition, or a contractual change affecting access to information. Give an owner the authority to reduce scope, suspend a workflow, or fund a transition.
The role of enterprise architecture
Enterprise architecture can make these decisions concrete by showing how a commercial dependency reaches an operating outcome.
For each critical AI workflow, architects should help connect the business capability to the application, model, data sources, identity controls, infrastructure, contract, and recovery process. They should also identify dependencies shared by proposed alternatives.
This work becomes more consequential when an AI agent can take action. An assistant that drafts text and an agent that updates records have different recovery needs. If execution stops midway, the organization may need to reconstruct completed actions, reconcile inconsistent states, prevent duplicate transactions, and assign unfinished work.
Replacing the model does not automatically resolve those problems. Recovery must address the workflow and its effects.
Architecture earns its place in governance when it enables leaders to answer practical questions: what breaks, who is affected, how long recovery takes, and what evidence supports that estimate.
A practical exercise for the next governance meeting
Choose one critical AI-enabled process and run a tabletop exercise around a significant supplier change.
Ask operations to describe how work continues. Ask finance to recalculate the business case. Ask procurement to explain the applicable contract terms. Ask technology to demonstrate what can be exported and what would need rebuilding. Ask risk to identify which approvals must be revisited.
Record the gaps, assign owners, and agree on the evidence needed to close them. Then test the most consequential assumption. If the plan depends on moving to another model within a week, establish whether that can actually be done for the relevant workload.
This exercise does not require a prediction about an AI market crash. It requires a clear view of the enterprise’s own dependence.
The decision executives cannot outsource
Financial leverage can support infrastructure investment. Its implications depend on the supplier’s obligations, resources, revenues, and circumstances. Enterprises should avoid treating debt as a simple proxy for failure.
They should be equally careful about treating impressive model performance as evidence of a durable operating arrangement.
An enterprise controls neither every supplier’s financing decisions nor the future price of AI. It does control how much dependence it accepts, what evidence it requires, and whether its recovery plan has been tested.
Before asking how much more of the business AI can run, ask what the business can still run if its AI arrangements change.
That is a governance decision worth putting in front of the board.
Source and editorial scope
Original commentary prompted by The Economist’s article linked through Apple News. The headline and publisher were verified through the Apple News landing page; the full article was not available for review. This commentary does not summarize its reporting or attribute financial figures or company-specific conclusions to it. The insurance example is illustrative, and the governance measures are the author’s proposed approach.
