I was on a customer call last fall when an enterprise architect said something I haven't been able to shake. Her team had just spent four months trying to swap one AI vendor for another. The original plan said three weeks. "We didn't switch vendors," she told me. "We rebuilt half our integrations and discovered what we'd actually been depending on."
Most enterprise leaders don't expect that to be the experience.
A recent Zapier survey of 500 US enterprise C-suite executives and decision-makers asked leaders how fast they could move from their primary AI vendor to a different one. 89% said they could complete the switch within four weeks. 41% said two to five business days. 13% said they could do it in a single day.
Two-thirds of leaders said their organization had already attempted a migration between AI platforms. Among that group, only 42% reported a smooth transition. The remaining 58% said the migration either failed or took significantly more effort than expected.
That's not a small gap. That's a planning failure waiting to happen.
Why the Math Doesn't Work
Switching a vendor sounds like changing a setting. In practice it's closer to retrofitting a building. By the time a team is migrating, the AI has been wired into a long list of things nobody planned to count: prompts that someone tuned over weeks, downstream systems that expect a specific output format, an internal tool that depends on a feature only one model supports, retrieval pipelines sized for a particular context window, monitoring that quietly assumes a particular response time.
None of that is in the contract. None of it is in the original adoption plan. Most of it isn't documented at all. It accumulates because that's what happens when a tool gets useful and people build on it.
The four-week estimate assumes a clean swap of one model for another. The actual job is finding everything that depends on the old one and deciding, piece by piece, what to do about it. That's a systems problem, and procurement isn't equipped to solve it on its own.
What the Rest of the Data Is Saying
The same survey shows enterprises starting to see the problem. 81% of leaders say they're at least a little concerned about their organization's dependency on specific AI vendors. 29% say they're very concerned. The two top concerns, tied at 46%, are data migration challenges and overdependence on a single vendor. Limited flexibility to integrate AI with existing tools comes in at 42%.
Almost half of organizations (47%) now have a dedicated internal team evaluating and managing AI vendors. 44% use multiple vendors at the same time to spread risk. 34% are deliberately designing around data portability and standard APIs. A third are using third-party orchestration to coordinate AI workflows across systems. These are all variations of the same instinct: keep the option to change your mind.
What strikes me about that list is how operational it is. Five years ago, vendor risk was mostly a procurement conversation. Today it's an architecture conversation, an integration conversation, and increasingly an observability conversation. The teams doing this well are treating AI vendor flexibility as a property of their systems, not a clause in a contract.
The Observability Piece Nobody Is Talking About
If you can't see what your AI is doing today, you can't tell what would break if it disappeared.
That sounds obvious, but it's the part of the migration problem that most often gets skipped. Teams have logs of API calls. They have cost dashboards. What they often don't have is a clear picture of which workflows depend on which model, what those workflows actually expect from the model's output, and which of those expectations are load-bearing versus incidental.
When that visibility is missing, every migration starts with discovery. People build a list of integrations from memory. Something gets missed. A scheduled job runs three weeks later and quietly produces wrong output, because the new model phrases things slightly differently and the parser downstream expects the old phrasing.
The teams that migrate well have done this work in advance. They can show, on demand, where AI is in their stack, what each call is doing, what the expected output looks like, and what depends on it. That's an inventory plus a behavioral baseline, and most organizations don't have either.
What "Designing for Change" Actually Looks Like
A few patterns hold up.
Treat the AI model as a replaceable component. Workflows should be structured so the model is one step in a longer process, with clear inputs and clear outputs that other systems agree on. When the contract between systems lives outside the model, swapping the model gets easier.
Keep your data portable from day one. The survey lists data migration as the single biggest concern, and the teams who handle it best aren't pulling their data out at migration time. They've kept it in their own systems all along, with the vendor processing it rather than owning it.
Run more than one vendor on purpose. The 44% using multiple vendors aren't doing it because they can't decide. They're doing it because routing different workloads to different providers builds the operational muscle for switching. The first migration is hard. The fifth is routine.
Watch the outputs. API call volume tells you how often a model gets used. It doesn't tell you whether the answers it produces are still good, drifting, or quietly degrading. The organizations that catch a vendor's quality slipping early are the ones with output-level monitoring in place before they need it.
Where This Leaves Things
Almost three-quarters of enterprise leaders say losing their primary AI vendor would either disrupt operations or stop key business functions outright. That dependency will keep deepening. The leaders taking it seriously aren't trying to avoid lock-in by avoiding adoption. They're embedding AI deeply, and building the surrounding systems so that the model itself stays optional.
That's a longer planning horizon than four weeks. It's also the only version of the work that actually holds up when a migration arrives.