AI Model Swaps Take Weeks, Not Hours, for Most Companies

Quick Facts

• 70% of companies that swapped an AI model in production spent a few weeks or more on engineering work beyond the model change, per a VB Intelligence October 2026 survey of 100 AI practitioners.
• 44% of organizations that swapped a model saw performance drop and had to rework the application or revert to the original model.
• 89% of executives believe they could switch AI vendors within four weeks, but only 42% of organizations that attempted a platform migration say it went smoothly.

Companies have built AI gateways and routers that promise seamless model switching with a single line of code. The reality is far more costly, according to VB Intelligence research published this week.

VentureBeat Intelligence surveyed 100 people who build, run, or track AI at organizations with at least 100 employees. Of the respondents whose companies swapped a model in production, 70% said the swap required a few weeks or more of engineering work beyond the model change itself. Thirty-six percent spent more than a month. Only 13% said the process was close to a simple configuration change.

Performance Drops Are Common

Among the 62% of surveyed companies that had replaced a model inside a production application, 59% saw performance drop and had to rework the application or revert to the original model. Just 19% said the swap took almost no effort.

Box CTO Ben Kus explained the core problem. Engineers tailor software to the model they started with. When a new model arrives, it "just maybe acts differently," Kus said in an interview for VentureBeat's Beyond the Pilot podcast on October 30. Box now tests application harnesses and models together rather than selecting a model in isolation.

Upwork VP of Platform Infrastructure Engineering Sudhanshu Bist drew the same conclusion at VentureBeat's Agentic Infrastructure Stack Summit in October. In a production system, he said he would "probably stick to the existing model." Upwork evaluates AI on cost per good result, not price per token, citing a difference of 50 cents versus 33 cents in one internal comparison.

The Confidence Gap

A Zapier survey of 542 executives conducted in February 2026 found 89% believed they could switch AI vendors within four weeks, and 41% said they could do it in two to five business days. Yet nearly three-quarters said losing their primary AI vendor would disrupt day-to-day operations or leave them unable to function at all. Only 42% of organizations that actually attempted a platform migration reported it went smoothly.

Vendor lock-in concern is widespread. The survey found 81% of enterprise leaders express concern about dependency on specific AI vendors. Data migration challenges and overdependence on a single vendor each ranked as top concerns, tied at 46%.

What Companies Are Doing Differently

Some organizations have reduced switching costs through early architectural decisions. PayPal VP and Global Head of AI Product Ren Zhang said the company has swapped models in applications many times. PayPal keeps its golden queries and analytical playbooks in a knowledge layer outside any model, so they transfer to each new release. Zhang said moving from 43% to 95% accuracy on one application came from improving that knowledge layer with no model change at all.

Uber built a Model Gateway in 2023 with a rule that everything built with AI must be model- and harness-agnostic. The gateway supports switching between Anthropic, OpenAI, other vendor models, and open-source models the company runs internally. Uber Distinguished Engineer Uday Kiran Medisetty said that for tasks with clear benchmarks, such as code review tested against years of past changes, the company can switch to a better-performing open-weight model in a couple of days. For one automated code cleanup job, benchmarking and selecting the right model produced a 90% reduction in cost per cleanup.

Zapier Senior AI Automation Engineer Emily Mabie pointed to stack flexibility as the dividing line between companies that can move quickly and those that cannot. "They're the ones that built flexibility into their stack from the start, so switching a model or a provider doesn't mean rebuilding everything around it," Mabie said.

The Trend Is Getting Worse

The share of companies reporting long swap timelines grew between September and October. In September, 49% of respondents said a model swap took a few weeks or more, and 12% said more than a month. In October, those figures rose to 70% and 36%, respectively. VB Intelligence notes the respondent mix shifted, with more business-side AI users in October, a group that reported longer timelines more often. Adjusting for that shift, 66% still said a few weeks or more, and 29% said more than a month.

For founders building AI products or executives overseeing AI infrastructure, the data points to a consistent pattern: switching models without investing in abstraction layers and benchmarking infrastructure is not a shortcut. It is a risk.