What Does It Actually Cost to Switch Your LLM Provider? 

Every enterprise can tell you what it paid for its AI model. Almost none can tell you what it would cost to leave it. 

Choosing a model gets a procurement cycle, a benchmark bake-off, a signature, the full weight of enterprise process. Living with that choice gets none of it, because nobody put "cost of leaving" on the roadmap. The true cost of a model decision was never the contract. It is everything you quietly wired to that model while nobody was tracking it as a liability. 

I have sat in enough of these conversations to know the instinct. A CIO tells me switching providers is "just an API change." I do not disagree that the API call is trivial. But an API call is not a system, and the system is what actually breaks. 

The six things nobody line-items

Ask an enterprise what it costs to change primary AI vendor, and you get a number that only covers the contract. Ask what it costs to change everything the model touches, and the honest answer is: nobody has priced it, because nobody has mapped it. 

Start with the knowledge base. Its embeddings were generated in a vector space specific to the old model, and a new model does not speak that geometry, so the corpus must be re-embedded from scratch. Next comes the evaluation suite. Its pass/fail thresholds were tuned to the old model's failure patterns, and a benchmark built for one model rarely measures another fairly without substantial rework. Prompts do not transfer seamlessly either. They were hand-shaped around one model's quirks, so swapping means rewriting and re-testing them line by line. Everything that has already cleared compliance needs attention as well. The approval workflows, audit trails, and explainability reports were certified against the old model's behaviour, so a regulator or an internal risk committee must sign off again from the beginning. Underneath it all, the integrations assume the old model's latency, context window, and output format. This mismatch is why downstream systems often break in ways that only become apparent weeks later, not on day one. 

None of that shows up on the vendor invoice. All of it shows up on your roadmap, disguised as unrelated delays. 

The number the industry has started measuring

IBM's Institute for Business Value put a figure on the instinct enterprises already had. In a global study of 1,000 senior executives across 16 countries, published in June 2026, 71% said switching their primary AI vendor or model would be difficult. That is not a surprising number. What is more revealing is this one: 72% said they would accept a 20% cost increase just to preserve the flexibility to change vendors later. That is not a cost of switching. That is an enterprise pre-paying an insurance premium against a dependency it cannot otherwise price. 

The same study found that 91% of organisations do not fully understand their own dependencies across AI vendors, models, and infrastructure, and 81% said a seven-day outage from a single AI vendor would cause severe or critical disruption. Only 7% of the organisations surveyed operate at what IBM classifies as the most advanced level of AI control, and those organisations protect 55% more of their operating profit when disruption hits. 

Read those together and the shape of the problem changes. This is not a story about switching being expensive. It is a story about enterprises not knowing what they are locked into until the lock turns. 

The cost compounds even if you never switch

Here is the part that should worry you more. Gartner's prediction, published in August 2026, is that inference costs per agentic workflow will increase more than fivefold through 2028, driven by workflows that consume far more tokens than a simple chatbot ever did. Gartner calls this the inference paradox: better unit economics on the model itself, escalating total cost of ownership overall, because the workloads running on top of it keep getting heavier. 

McKinsey's own research into agentic AI economics explains where that overspend actually goes. About 60% of an agentic task's cost is tied up in refining the answer, not producing the first draft of it. And that cost is not even stable: identical tasks can see token consumption vary by a factor of 30 between runs. If you cannot predict what a task costs today, you certainly cannot predict what re-tuning that task for a different model will cost tomorrow. 

The FinOps discipline has caught up with the spend side of this fast. 98% of FinOps teams now manage AI spend directly, up from just 31% two years ago. What almost none of them are asked is whether that spend, and everything built to justify it, would survive a change of model underneath it. Visibility into cost is not the same as readiness to move. 

Here is the fast-forward, because a statistic will not make you feel it the way a timeline does. Three agents wired to a single model today. Three hundred within the year, because that is what every agent programme does once the first ones work. The chaos you are not planning for now is the chaos you will be firefighting then, at ten times the scale and with none of the mapping done. 

This is the compounding I want you to sit with. Every month you do not decouple your workflows, evaluations, and governance from a specific model, the volume of what is welded to that model grows. The retrofit does not get cheaper by waiting. It gets more expensive, quietly, on a schedule nobody put in the budget. 

I will name what is actually happening here, because vague language protects nobody. It is not in a frontier lab's commercial interest for your switching cost to fall. The deeper your embeddings, your evaluations, and your workflows are threaded through one model, the less leverage you have at renewal, and the more of your operational knowledge sits inside a system you do not control. This is not malice. It is incentive. But an incentive you do not name is one you cannot manage. 

Where replaceability actually lives

This is exactly why we built the Enterprise AI Harness the way we did. A harness sits above the model, not inside it: the governed context, the evaluation logic, the audit trail, the workflow orchestration, all built to survive a model change instead of being rebuilt by one. Own that layer, and swapping the model underneath it becomes what it should have been from the start: a procurement decision, not an archaeology project. 

That is what we mean when we say build and reason anywhere. Not that switching becomes free. Switching a model will always carry some cost. But it should be the cost of a decision, not the cost of an excavation. 

I would rather tell you what still needs building than claim the switch is already painless everywhere, for everyone, today. Say less than you could. The vendor who tells you honestly what is still on the roadmap earns more trust than the one who insists it is all already built, and finds that out the day you actually need it. 

Ask your own organisation a harder question than "what would it cost to switch": can you say, right now, who touched which data, through which model, and why? If you cannot answer that in the time it takes to ask it, you do not have an AI strategy. You have exposure, and exposure is the most expensive thing to price later. If the honest answer takes longer to construct than the switch itself, you have already paid the toll. You just have not been billed for it yet. 

You can renegotiate a contract in a quarter. You cannot renegotiate a dependency you never mapped. 

Build anywhere. Govern through Jeen. This is AI on your terms.