If Your AI Provider Disappeared Tomorrow, Could Your Business Keep Working?
What if your AI provider vanished? Discover the risks of dependency on AI technology and how it can impact your business operations.
Article
It's Monday morning at a growing professional-services firm, and employees are discovering that the AI platform supporting several everyday activities has stopped working. The customer service team can no longer retrieve internal guidance through its usual assistant, while document-processing workflows have stalled and several client reports are due later that day.
The supplier's status page confirms a service problem but provides no reliable indication of when access will be restored. Staff can still reach some of their original business applications, although the connections that made information easy to retrieve and organise are unavailable.
The managing director wants to know whether employees can continue operating and how quickly the company could move to another provider if necessary.
Eighteen months earlier, the firm began with a relatively modest AI project designed to reduce the time employees spent searching internal documents. The application worked well, so management gradually connected additional information sources and introduced it into customer service, administration and reporting. Each extension appeared commercially sensible because employees already understood the product and the supplier offered attractive terms.
Over time, however, established working practices changed and some manual procedures were abandoned as the platform became central to operations. The business didn't deliberately decide to become highly dependent on one AI supplier; the dependency developed through a series of individually reasonable decisions.
I was thinking about this while reading Oliver Wyman Forum's 2026 report, *The Reality Check: Beyond the AI Hype*. Its discussion of AI sovereignty raises a practical question that extends beyond debates about countries developing their own foundation models.
Businesses need to understand which technology dependencies they're prepared to accept, what it would cost to replace them and how essential operations would continue if a provider became unavailable. Buying AI capability can be relatively straightforward, but the consequences of relying on it may become apparent only after the technology has been embedded across several processes.
The important issue isn't complete independence from suppliers; it's whether the organisation has realistic options when a dependency that matters stops working.
## The surrounding system may be harder to replace than the model
When the outage begins, somebody suggests switching to another AI model. That could be manageable for a simple application that sends a question to a language model and displays the answer, although the replacement would still need testing. Our professional-services firm has developed something more complicated because the platform retrieves internal documents, applies user permissions and supports specialised workflows.
Its behaviour may also depend on carefully constructed instructions, particular integrations and features provided by the original supplier. Changing the model is therefore only one part of restoring the complete service.
An alternative may interpret instructions differently, produce responses in another format or handle connected tools in ways that require modification. The firm also needs to establish whether the replacement retrieves the right information and respects the permissions governing confidential client records.
Existing evaluation procedures may need updating, and employees may require training before relying on the new process for important work. Oliver Wyman reports that technology leaders in one survey estimated approximately 21 to 60 engineer-days to switch their primary foundation model for a main use case.
That range reflects surveyed leaders' estimates rather than a universal switching cost, but it illustrates why the work around a model can be as significant as the model itself.
I'd therefore be cautious about interpreting a supplier's claim of model flexibility as evidence that the complete workflow is portable. An application might connect to several foundation-model services while relying heavily on proprietary document handling, permissions or orchestration features. Those dependencies can become difficult to replace when the business has invested in specialised workflows and training.
This doesn't make the original supplier unsuitable, but it does change the commercial meaning of the relationship. Management should understand whether replacing the provider would involve a limited configuration change or a substantial implementation project.