Your stack is not as portable as the contract suggests
A model provider changes its pricing. A feature you depend on moves to a higher tier. You open the migration plan and find the work is mostly eighteen months of prompts, stored context, evaluation sets and permission rules.
None of that appears in the contract. Vendor lock-in now lives in your code, your data and your governance.
What buyers say, and what they actually do
According to the State of AI Architecture 2026, 81% of respondents agree that the proprietary platforms of frontier labs and hyperscalers make it difficult to move between vendors, and 84% say it is important that their AI strategy avoids vendor lock-in.
In the same research, 81% report that concern about switching costs has caused their organization to delay, scale back or restructure a migration from one model provider to another, and 29% say it happens frequently. Among financial services respondents, 92% have seen it at least once.
The survey also found that the ability to switch providers without rebuilding ranked sixth out of eight factors at 24%, while avoiding dependence on a single vendor ranked last. Data security and privacy led at 64%, ahead of cost control at 43%.
Zapier’s survey of 500 U.S. enterprise executives found that 89% believe they could switch AI vendors within a month, and 41% estimate two to five business days. The same poll found that 47% say ending AI services would disrupt key business functions and 27% are completely reliant on their AI vendors.
Teams that have done the work report slower numbers. A developer survey of 19 organizations with production AI deployments found that simple use cases move in a median of three to eight hours, while complex pipelines take a median of two to seven days, with 24% reporting that switching is not feasible. Complex-pipeline teams expect to rewrite 21 to 50% of their prompts.
Dependence often starts below the model. Research for the Open Cloud Coalition, covering 1,250 UK and EU enterprise AI customers, found that 91% say their cloud provider influences which enterprise AI tools they use, and 56% feel restricted in using enterprise AI products not sold by their current cloud provider.
The portability audit to run on Monday
Omdia’s Lian Jye Su told TechTarget that the biggest risk lies in context engineering, memory management and harness engineering. Portability is a per-layer decision, and renting a foundation model is usually fine. The layers around it decide whether you can leave.
- Prompts and schemas: can another model run this workflow with prompt rewrites only? Keep prompt templates and response definitions versioned in your own repo.
- Context and memory: does the retained information move with your workflow? Keep it in a store you can export.
- Evaluation sets: do your tests live outside the vendor dashboard? Keep a golden test set and scoring code you own.
- Identity and permissions: are agent identities defined by your directory? Keep policy in your own identity layer.
- Integrations and business logic: do connectors call your interfaces? Keep an adapter layer between model and workflow.
- Audit and governance records: can you reconstruct a decision without the vendor? Keep logs and traces in your storage.
Then test the theory on one workflow.
- Pick one live workflow with real traffic.
- List its dependencies by layer, prompts, context, tools, identity, evaluations, logs.
- Run it on a second provider behind your own adapter.
- Score the output against your evaluation set and record the hours spent rewriting prompts.
- Write down, per layer, whether you own it, abstract it or leave it with the vendor.
What this costs, and where it fails
Portability is not free. IBM’s calculus of AI sovereignty report found that 72% of executives would accept a 20% cost increase to maintain multiple AI vendors for strategic flexibility, while only 9% say they have an excellent understanding of their dependencies today and 71% say switching their primary AI vendor or model would be difficult if required today.
Renting a model is cheap and rebuilding a memory layer is not, so not every layer deserves the same investment. A second vendor does not remove dependence by itself. BCG found that 64% of companies use more than one digital platform, yet using multiple platforms often shifts dependence rather than removing it.
The hidden cost is usually governance. In the developer survey, prompt templates and governance sign-off were each cited by nine of nineteen organizations as where switching hurts most.
Customer-facing layers are usually where fragmentation shows up first, as we described in fixing fragmented customer journeys.
If your situation looks like one provider running customer workflows you could not move in a month, we can help. Data Innovation has worked in CRM and data from Barcelona for more than 20 years, and we build and operate intelligent systems around the models you choose so the portable parts stay yours. Tell us what you are running.