The fear of AI lock-in has moved. Companies swap large language models all the time now, but a new survey of 520 technical leaders says the parts that really trap you are the integrations, agent workflows and guardrails you build around the model.
The research comes from Stacklok and was written up by its CEO, Kubernetes co-creator Craig McLuckie, on the Stack Overflow blog on 21 October 2026. Here is what the numbers say, where they fit with other surveys, and what you should keep portable.
Key takeaways
- Only 8% of the companies surveyed rely on a single model provider. 25% use five or more.
- 53% put integrations with internal systems or MCP servers among the two hardest things to migrate. 52% said the same about multi-step agent workflows.
- More than 80% have delayed, scaled back or restructured a migration because of switching costs.
- Even so, “switch providers without rebuilding” ranked only sixth of eight buying criteria.
Who did Stacklok survey, and when?
Stacklok ran the survey in early October 2026. Every respondent held a manager-level title or above at a North American company with more than 500 employees, and every company was already running LLMs or AI agents in a pilot or in production.
Keep the source in mind. Stacklok sells an open “harness” (the runtime that wires a model to tools, memory and identity) along with MCP and LLM gateways, so a finding that the agent layer is the new trap suits its pitch. The numbers are still useful. They just aren’t neutral.
Why is the model the easy part to swap?
Most providers now accept roughly the same request: a prompt goes in and text or a tool call comes out. Routing layers turn changing models into little more than a config edit. We covered one of them in our piece on Vercel AI Gateway session affinity.
Other research shows the same multi-model habit. A Dataiku and Harris Poll study of 600 CIOs, reported in April 2026, found that 81% expected to use two or more LLM providers in 2026 and 55% had already switched models at least once, mostly to cut costs.
Where does AI lock-in really live now?
Think of the model as the engine and everything around it as the car. You can drop in a new engine, but the wiring loom was cut to fit the old one. Here is how respondents rated the different layers:
| Layer | What it is | How hard to move |
|---|---|---|
| The model | The LLM you call through an API | Easy. 92% already use more than one provider |
| Integrations and MCP servers | Connectors to your databases, tickets and internal tools | Hardest. Named by 53% |
| Multi-step agent workflows | Chains of plans, tool calls and hand-offs | Named by 52% |
| Guardrails and access control | Policies on what an agent may see and do | Next most cited |
| Agent state, memory and evals | What your agents remember, plus your test suites | 37% fear they can’t export it |
MCP, the Model Context Protocol, is the open standard that lets an AI app call outside tools. Anthropic donated it to the Linux Foundation’s Agentic AI Foundation in December 2025. The protocol is open, but the way a vendor extends it often isn’t, and those extensions are where integrations get stuck.
How much does switching actually cost?
The answer is a lot of time. 57% of respondents estimated that leaving their current provider would take four months or more, and 29% said switching costs derail migrations frequently. 39% worry about product or API deprecation, which can force a move on someone else’s timetable.
Teams also tend to underestimate the job. In a Zapier survey of 542 US executives, covered by The Register, nearly 90% thought they could switch AI vendors within four weeks. Only 42% of those who had tried said the migration went smoothly.
Why do buyers say they care but rank it low?
This is the strangest result in the survey. 84% said avoiding lock-in matters to their AI strategy, and 81% agreed that frontier labs and hyperscalers make switching hard. Yet when they ranked what drives a purchase, security and privacy came first, then cost control and compliance. “Avoiding dependence on a single vendor” came last.
Our opinion: this is how lock-in usually happens. Portability never feels urgent on the day you sign. You notice it later, when a price changes or an API is retired.
What do the five-provider teams do differently?
Stacklok compared organisations using five or more providers with those using one or two. The heavy multi-model users were more likely to:
- have built integrations to internal systems or MCP servers (76% vs 42%);
- use open-source agent frameworks (45% vs 22%);
- rate provider portability as critical or very important (67% vs 47%);
- be highly likely to evaluate open alternatives to closed frontier-lab stacks (35% vs 13%).
McLuckie’s advice follows from this. Keep tool definitions on the open protocol rather than a vendor’s version of it, own your orchestration code, and make sure your evals run against any model endpoint. If you are choosing a framework, our LangChain vs LangGraph guide is a good starting point.
What happens next
Respondents expect to take more control as they gain experience. 88% said they will own more of their AI architecture over time, and 80% would likely evaluate an open alternative to closed frontier stacks. What would win them over is security and compliance guarantees, performance on par with frontier models, and proper enterprise support.
They don’t want to give up the frontier models. They want to keep calling them through a harness they control. Expect more vendors to sell exactly that, and expect the frontier labs to keep bundling agents, memory and tools more tightly into their own platforms. For a sense of how mature that layer needs to be, see our production-grade AI agents maturity model.
FAQ
What is AI lock-in?
AI lock-in is when leaving an AI vendor costs so much time or money that you stay even if a better or cheaper option exists. The Stacklok survey says the cost now sits mainly in integrations, workflows and policies, not in the model.
How many companies use more than one AI model provider?
In Stacklok’s October 2026 survey, 92% used more than one provider and 25% used five or more. Dataiku’s 2026 CIO poll found 81% expected to rely on two or more.
How can my team avoid AI lock-in?
Build tools on standard MCP rather than vendor extensions, keep orchestration code in your own repo, store agent memory and evals in formats you can export, and test regularly against a second model endpoint.
Your next step: list every integration your agents touch and mark which ones would break if you changed provider tomorrow. That list shows where you are locked in.
2 comments