Companies swap AI models easily, but the connections built around them stick

Companies are getting good at switching AI models. They are not good at moving everything built around them. A survey of 520 technical leaders by Stacklok, a software company, found that the real lock-in sits in the connections to internal systems and in the multi-step workflows, not in the model itself.

What the survey found

Only 8% of the companies use one AI model provider. A quarter use five or more. Every respondent holds a manager-or-above title at a North American company with more than 500 employees that runs AI models or agents in pilot or production.

Asked what is hardest to move, 53% named links to internal systems or MCP servers. MCP is a common standard for connecting AI tools to other software. Multi-step agent workflows came next at 52%. Guardrails and access-control settings followed. When a company changes provider, these pieces have to be rebuilt.

The cost of switching

More than 80% of the companies have delayed, scaled back or restructured a switch because it looked too costly. For 29%, that happens often. When asked how long a move off their current provider would take, 57% said four months or more. Also, 81% agree that big AI labs and cloud giants make it hard to change vendors.

Yet the ability to switch without rebuilding ranked sixth of eight buying criteria. Avoiding dependence on one vendor ranked last. Security, privacy, cost control and compliance came first. The author notes that buyers are judged on this quarter's bill, not on a rebuild two years away.

What heavy multi-model users do differently

Companies with five or more providers look different from those with one or two. Among them, 76% have built internal-system links, versus 42%. And 45% use open source agent frameworks, versus 22%. Also, 67% call portability very or critically important, versus 47%.

Many also worry about getting trapped on a closed platform. Thirty-nine percent mentioned products or APIs being retired. And 37% fear being unable to export an agent's state, memory or test results.

What to do

The author, a co-creator of Kubernetes, offers advice for anyone building AI tools now:

• Keep tool definitions on the open protocol, not a vendor's extension of it.
• Put the orchestration, meaning the steps agents follow, in code you own.
• Make sure tests can run against any model.
• Build your own harness for identity, memory and tools instead of renting one.

The survey was run by a company that sells in this space, so its framing is worth weighing. The numbers still point to a plain lesson: choose carefully where your AI connections live.