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

What happened

A survey of 520 enterprise AI leaders revealed significant concerns about vendor lock-in, with many having delayed or restructured model migrations due to switching costs. While avoiding dependence on a single vendor is a priority, it ranks low in decision-making criteria. The article highlights that the agent layer, not the model itself, is the main barrier to portability, with integrations and workflows being the hardest to move.

Why it matters

For AI developers and enterprise leaders, this highlights the critical need to design systems that are portable and vendor-agnostic. The article underscores the importance of building flexible agent stacks that can survive provider changes, especially in regulated industries where data portability is a top concern.

• Avoiding vendor lock-in is a priority for enterprises, but it is often deprioritized in decision-making.
• Integrations and agent workflows are the main barriers to portability, not the model itself.
• Open source alternatives are appealing but require meeting the same security and compliance standards as closed systems.
• Enterprises are moving toward more control over their AI architecture as they gain experience.
• Security and compliance are immediate concerns, while lock-in is a long-term risk.

The bigger picture

The article emphasizes that enterprises are already multi-model but lack the infrastructure to support seamless transitions between providers. This has significant implications for AI infrastructure design, as the agent layer becomes a critical point of focus for portability and control.

The article concludes that enterprises are already multi-model but need to focus on building portable agent stacks. Stacklok is positioned to offer an open architecture that provides security and control, helping enterprises avoid vendor lock-in and maintain flexibility in their AI strategies.

Technical details

• Multi-model adoption is common, with 8% using a single provider, 50% using two to three, and 24% using five or more.
• Integrations to internal systems or MCP servers are critical for enterprise AI workflows.
• Agent workflows and state management are key challenges in portability.
• Open source alternatives are seen as offering cost control and data sovereignty.
• Security and compliance are immediate concerns, while lock-in is a long-term issue.

FAQ