The Adoption Gap No One's Measuring: Using an AI Platform Isn't the Same as Betting On It

The argument over which model provider is "winning" the enterprise has been running mostly on spend data — corporate card platforms, model-routing traffic. That measure has a blind spot: it can't tell a company piloting a platform from one that has made it the backbone of its agent stack. An August survey of enterprise agent deployments across company sizes and industries gets at that distinction directly, and the numbers are stark. Among enterprises that have OpenAI's agent tooling anywhere in their stack, roughly seven in ten have made it their primary orchestration layer. Among those running Anthropic's Claude Platform, fewer than four in ten have — the lowest conversion rate of any platform in the survey, and one that holds regardless of company size or industry.

For a regulated enterprise, that distinction isn't academic. Naming a primary orchestration platform means deciding whose infrastructure decides what an agent is allowed to touch, what happens when a step fails, and who's accountable for the outcome. That's a materially bigger commitment than adding a model to an existing stack, and it shows up in audit scope, vendor-risk review, and exit planning long before it shows up in a spend line.

The same data complicates a simple "Anthropic is losing" reading. Among enterprises naming platforms they're considering adopting in the next year, Anthropic drew the highest ratio of prospective interest relative to its current installed base of any provider with meaningful scale in the survey — ahead of OpenAI, Google, and Microsoft on that specific measure. Enterprises not yet committed to Claude are, proportionally, more likely to be evaluating it than non-users of other platforms are evaluating theirs. Present caution and future interest are coexisting in the same numbers.

What's actually being decided underneath all of this is where the agent control plane lives — the coordination layer that sits above any single agent's runtime and decides which agents exist, what each may touch, and who answers for them when something goes wrong. Roughly a quarter of the enterprises surveyed expect a model provider to own that layer outright by the end of 2026; a larger group is planning for a hybrid of provider-native and external orchestration rather than committing fully either way.

Tellingly, the leading objection to letting a provider hold that control plane isn't vendor lock-in — it's inflexibility across models and tools, cited well ahead of lock-in as the top concern. The spend-based figures that dominate this debate, from corporate card platform Ramp and model-routing service OpenRouter, measure how much is flowing through a platform, not whether an enterprise has committed its control plane to it — which is exactly the distinction that matters for regulated institutions. For BFSI and other regulated institutions, control-plane ownership deserves the same vendor-risk scrutiny already applied to model selection under frameworks like TRUST framework and CLEAR framework, not an inherited default from whichever platform a team happened to pilot first.