Stacklok survey of 520 AI leaders says model swaps take four months

A reader whose company has planned a model swap for next quarter should know that the model is the fast part, and the other code around it is the slow part. A survey of 520 technical leaders in North America says so in large numbers, and reports in the same findings that the people buying this software are weighting the fast half when they pick a vendor.

The piece appeared on the Stack Overflow Blog on 21 October, written by Craig McLuckie, chief executive of Stacklok. Every respondent was at a company with more than 500 employees, held a manager-or-above title, and had an LLM or an agent running in pilot or production. Stacklok ran the survey, so the vendor publishing the finding has an interest in the finding, and the item is written on that basis.

What the numbers say

Only eight percent of the companies use one model provider, McLuckie reports, and twenty-five percent use five or more. Swapping which lab answers a given call is already a routine routing decision for most of them.

Fifty-three percent of the leaders name "integrations to internal systems or MCP servers" as one of the two hardest elements to migrate. Fifty-two percent name multi-step agent workflows alongside them. Fifty-seven percent say a migration takes four months or more. More than eighty percent of the companies have delayed, scaled back, or restructured a migration already.

The gap the piece draws out

The leaders know this. Eighty-one percent of them agree that frontier labs and hyperscalers make it hard to move between vendors. And then, when the same leaders are asked what they weight most when they pick a tool, the "ability to switch model providers without rebuilding" lands sixth out of eight criteria, and "avoiding vendor dependence" lands last. Thirty-nine percent say product or API deprecation is on their mind, so somebody has thought about this.

Eighty percent of the respondents also say they would be likely to evaluate an open alternative, which is the opportunity Stacklok writes toward.

The honest reading of the two findings together is that the cost of switching is paid later and the convenience of the current vendor is paid now. The leaders are quietly accepting that trade while telling a survey they do not like it. The expensive part of a model change sits in the integration code, in the MCP servers, and in the chained agent steps, and a procurement form that scores a vendor on model price alone is pricing the half of the problem that was going to move easily anyway.

For a team planning its own stack, the useful move is to list the MCP servers, the connectors to internal systems, and the agent workflows at the start, and to ask of each one how many weeks it would take to rebuild against another provider. That list is where the four months in the survey will be spent.

Source

Stack Overflow Blog, "Survey Says: You'll swap the model, but everything you build around it is staying put," by Craig McLuckie.