The model is easy to swap. The agent layer isn’t

Here’s what 520 enterprise AI leaders told us about getting locked into frontier labs

TL;DR. The technical leaders responsible for AI adoption say vendor lock-in worries them, and many have already paid for it. 84% say avoiding lock-in matters, and 81% have delayed, scaled back or restructured a model migration because of switching costs. Yet when they choose AI infrastructure, avoiding dependence on a single vendor ranks dead last of eight criteria.

Approach

Stacklok’s CEO, Craig McLuckie, and CTO, Joe Beda, are co-creators of Kubernetes. One of the primary motivations behind the Kubernetes movement was ensuring enterprises were never locked into any one hyperscaler. That intent permeates Stacklok today; we think a lot about how we can help enterprises avoid lock-in to any one frontier lab’s agent stack. 

We wanted to know whether or not enterprises perceived the same risk, so in September 2026 we surveyed 520 technical leaders. Respondents worked at enterprises headquartered in North America with 500+ employees. And everyone had a manager (or higher) job title, with direct oversight of AI adoption. The survey was fielded by Dimensional Research. 

Multi-model is already the norm

Only 8% of respondents use a single model provider. Half use two or three, and 24% use five or more. The bigger the company, the more providers: 37% of companies with more than 10,000 employees run five or more models, versus 12% of companies in the 500 to 2,500 range.

Enterprises running five or more models look different in important ways. 76% of them have built integrations to internal systems or MCP servers, compared with 42% of organizations on one or two providers. They’re also more than twice as likely to evaluate an open alternative to closed frontier architectures. 

“Portability stops being a preference and becomes an operating requirement once you’re routing work across several models.” — Joe Beda, CTO

Lock-in is feared, but it’s a problem to solve later

The headline tension in the data is between what respondents say they worry about and what they use to inform architectural decisions. 84% agree it’s important that their AI strategy avoids vendor lock-in, and that’s because they’ve already felt the pain: 81% report that switching-cost concerns have already caused their organization to delay, scale back or restructure a migration from one model provider to another.  

But, when we asked what matters most when they select AI infrastructure, the ability to switch providers ranked sixth of eight potential criteria. Avoiding dependence on a single vendor ranked dead last. Data security, cost control and regulatory compliance are immediate issues that affect buying decisions now. Lock-in is an issue to deal with later (and it’s probably someone else’s problem anyway). 

“Nobody’s budget line is ‘avoid lock-in.’ Security and compliance have a deadline this quarter. Lock-in sends the bill in two years, usually to whoever replaced you. That’s why it ranks last right up until the day it costs you a migration.” — Craig McLuckie, CEO

The model is easy to swap. The agent layer isn’t.

Here’s the finding that enterprises need to internalize now. We asked respondents which elements of their AI system would be hardest to move. 53% named integrations to internal systems or MCP servers, and 52% highlighted multi-step agent workflows. 

In other words, it’s the layers directly atop the model that risk becoming locked inside a model provider’s walled garden. If a provider raises prices or if an alternative option offers better performance, enterprises are realizing they cannot just lift and shift their harnesses, tools, policies, memories and more. 37% of respondents specifically worry about being unable to export agent state, memory or evals, and in financial services that becomes the single most common concern at 52%.

“Prompts are text. You can diff them, port them, rewrite them in an afternoon. An MCP integration into your ticketing system that’s been in production for a year has auth, policy, audit and a dozen edge cases baked into it.” — Joe Beda, CTO

Appetite for open is broad. Readiness is narrower.

80% of respondents say they would be likely to evaluate an open source alternative to the vertically integrated architectures from frontier labs, and one in five is highly likely. 88% expect to take more control of their AI architecture as their teams gain experience, reducing reliance on current vendors. Those respondents see open source as offering cost control (62%) and sovereignty over code and data (53%).

Finally, we asked every respondent to finish the sentence “We would move more AI workloads to an open architecture if it offered…” Of the 277 respondents who gave a substantive answer, roughly 36% emphasized the need for security, compliance or governance guarantees. Open source has to clear the same security bar as the closed alternatives before it earns enterprise workloads.

“When an open architecture passes the same security bar as a closed one, the conversation about adoption changes from ‘why’ to ‘when.’” — Craig McLuckie, CEO

What we take from it

Enterprises have already decided to be multi-model. What they haven’t done is build the layer around the model so it survives a provider change. The integrations, workflows, guardrails and memory that make an agent useful are the same things respondents expect to be hardest to move, and in regulated industries like financial services, the inability to take that state with you is now the top concern about building on a frontier lab.

Stacklok has the open source DNA and the security pedigree to offer enterprises an alternative. We’re building an open architecture that delivers on enterprise requirements for security and control, and that ensures our customers rent models, but own their agent stack. I encourage you to engage with us; you can start with our open source projects, or reach out directly to our team. 

The full report, with every chart and the respondent profile, is at https://stacklok.com/resources/state-of-ai-architecture-2026/