AI model and vendor selection in 2026: switching cost, residency, and the portability test nobody runs

90% of executives believed they could switch AI vendors within four weeks. Of those who tried, 58% found the migration failed or took far more effort than expected. Here is what to test during selection, including the thing nobody tests.

Model selection is usually treated as reversible, and it is not. Of the enterprises that actually attempted a migration between AI providers, 58% found it failed or took significantly more effort than expected, against 90% who had believed they could switch inside four weeks.

That gap is the most useful thing in the public record on this question.

The expectation, and the reality

How hard executives think switching is, and how hard it was

• Believed they could switch within four weeks. 90%. 41% said two to five business days. 13% said one day.
• Of those who actually migrated, found it smooth. 42%. 58% reported the migration failed or required significantly more effort than expected.

74% would face operational disruption if they lost their primary AI vendor. 47% would have at least one broken business function. Only 6% would see none.

Source: Survey of 542 US decision-makers with active paid AI contracts, fielded by Centiment 30 January to 6 February 2026, margin of error 4% at 95%.

The survey behind those figures covered 542 US decision-makers with active paid AI contracts, fielded in early 2026 with a stated margin of error of four percentage points. It is the cleanest evidence available on switching cost, and it is unusually well documented for this category.

The dependency picture is worse than the switching picture. 74% said they would face operational disruption if they lost their primary AI vendor, 47% would have at least one broken business function, and only 6% would see none.

Your migration schedule is partly not yours

Even organisations with no intention of switching get moved. Model providers publish deprecation policies, and they are shorter than most procurement cycles.

Your migration schedule is partly set by your vendor

Published deprecation policies and observed cycle times, from vendor documentation.

Source: Anthropic and OpenAI published deprecation documentation, retrieved September 2026.

Anthropic commits to at least 60 days of notice before retiring a publicly released model, and the observed cycles through 2025 and 2026 ran close to that. OpenAI has a series of platform retirements scheduled through 2026 and 2027. None of this is hidden, it is on the vendors' own deprecation pages, but it rarely appears in a selection business case.

Plan on the assumption that the specific model you select will not be the one you are running in two years, and choose for the ease of that transition as much as for today's benchmark score.

What actually decides the choice

The selection tests, in the order that eliminates fastest

Residency first, because it removes options before you spend time on capability.

• Data residency and inference residency, separately. Where data is stored and where requests are processed are different guarantees, and they differ by region and purchase route.
• Your own task, on your own data, at your own quality bar. Public benchmark performance does not transfer reliably to a specific enterprise task.
• Total cost at real volume, not list price per million tokens. Include caching behaviour and batch discounts, which change the answer materially.
• Deprecation policy and notice period. You are choosing a migration cadence as much as a model.
• The portability test. Run the same evaluation set against a second provider and record what it took. Half a day now, against what 58% learned the hard way.

77% of organisations factor country of origin into AI vendor selection and 83% view sovereign AI as strategically important.

Source: Deloitte State of AI in the Enterprise 2026, 3,235 leaders; Thinklytics selection practice.

Apply residency first, because it eliminates options fastest. Providers distinguish between where data is stored and where inference is processed, and the guarantees differ by region and by whether you buy direct or through a cloud platform. A global endpoint that routes to the nearest capacity is a different compliance posture from a regional endpoint that guarantees both storage and processing stay in region.

Deloitte's January 2026 survey of 3,235 leaders found 77% factor country of origin into AI vendor selection and 83% view sovereign AI as strategically important. That is a procurement reality now, not a European edge case.

The thing most evaluations skip

Capability gets tested well. Portability does not get tested at all.

Before signing, run the same evaluation set against a second provider and record what it took. Not to switch, just to know. The organisations that discovered a migration was hard discovered it during the migration, which is the expensive moment to find out. Half a day during selection tells you what 58% of them learned the hard way.

What we would do first

Write down the three things that would make you leave: a price change, a deprecation you cannot absorb, a compliance requirement the provider cannot meet. Then check what each one would actually cost you today. If the answer to any of them is "we could not", you have a single point of failure rather than a vendor.

Our AI governance and managed operations practice covers the policy and monitoring side of this, and AI readiness covers whether your data layer can support any provider at all. For the related question of whether customisation is the answer, see fine tuning does not fix hallucination.

Frequently asked questions

How hard is it to switch AI model providers?

Much harder than executives expect. A survey of 542 US decision-makers with active paid AI contracts, fielded January and February 2026, found 90% believed they could switch vendors within four weeks and 41% said within two to five business days. Of the 66% who actually attempted a migration, only 42% found it smooth. 58% reported the migration failed or required significantly more effort than expected.

What is the real cost of vendor lock-in?

Operational rather than financial, and it shows up as disruption. In the same 2026 survey, 74% said they would face operational disruption if they lost their primary AI vendor and 47% would have at least one broken business function. Only 6% would see no disruption. 81% reported concern about vendor dependency, but far fewer had done anything structural about it.

How often do enterprises actually change model provider?

Rarely, and usually within the same provider. Menlo Ventures found 66% of enterprises upgraded to a newer model from their existing provider, 23% made no change, and only 11% switched vendors. Performance rather than price drove the decision. The practical implication is that your first provider choice tends to persist, which is an argument for taking it seriously rather than treating it as reversible.

What forces a model migration even when you do not want one?

Model retirement. Anthropic commits to at least 60 days of notice before retiring a publicly released model, and observed 2025 to 2026 cycles ran roughly that long from deprecation to retirement. OpenAI has a series of platform retirements running through 2026 and 2027. These are documented on vendor deprecation pages and they mean your migration schedule is partly set by someone else, regardless of your roadmap.

How should data residency affect the choice?

It narrows the field faster than any other criterion, so apply it first. Providers distinguish between where data is stored and where inference is processed, and the options differ by region and by whether you buy direct or through a cloud platform. Some regional guarantees cover both storage and processing while global endpoints do not. Deloitte's 2026 survey of 3,235 leaders found 77% factor country of origin into AI vendor selection and 83% view sovereign AI as strategically important.

Should we run more than one model provider?

It is the most common mitigation, and it is not free. In the 2026 survey, 44% of organisations ran multiple vendors, 35% kept open-source alternatives available, 34% designed for data portability and standard APIs, and 29% negotiated shorter contracts. Multi-provider architecture buys optionality at the cost of complexity, and it only pays if you actually maintain the second path rather than letting it rot.

What should the evaluation actually test?

Your own task, on your own data, with your own quality bar, and the total cost at your real volume rather than a list price per million tokens. Add the operational questions that decide the next two years: what the deprecation policy is, what the notice period is, whether regional processing is guaranteed or best-effort, and what it would take to run the same workload somewhere else. Most selection processes test capability well and portability not at all.