When AI Spending Runs Ahead of Your Controls, Three Platforms Won't Help

Enterprise AI teams have stopped betting on a single orchestration platform, and that is not a failure of strategy, it is the strategy. The median enterprise now runs three at once, not by accident, but because none of them fully trusts a single vendor to run the show, according to VB Pulse data. This is a deliberate hedge against the exact kind of lock-in that defined the early cloud era, and it is the smartest thing these teams are doing. But here is the uncomfortable truth buried in the same data: one in five enterprises still cannot stop a runaway agent's spending in real time. You can run three platforms, you can rate them 4.17 out of 5 for satisfaction, and you can still be flying blind when a token bill spirals past your budget before lunch. That is not a technology gap. That is a governance gap, and it is the one that will actually hurt. The plural-platform reality is not just about flexibility, although 29% of respondents cite that as their top buying factor. It is about distrust. Security and permissioning limitations top the list of builder concerns at 37%, and 22% specifically call out limited visibility and observability. Enterprises are not choosing multiple platforms because they are spoiled for choice; they are choosing them because they do not believe any single vendor will let them see what their agents are actually doing. This connects directly to what we wrote about in Explore Agent Harnesses: Architecting AI for Financial Efficiency, where the operational layer, not the model, is what determines whether an agent is an asset or a liability. The harness is where you enforce cost controls, set permissions, and audit behavior. If you are betting on a vendor's native controls alone, you are betting that their priorities align with yours. The data suggests they do not. Only 30% of builders rely on native platform controls, and a quarter have built custom gateway plumbing to intercept runaway agents. That is not a vote of confidence in the platforms; it is a vote of no confidence. What is striking is how little this changes with company size. You might expect the 10,000-person enterprises to have mature fiscal controls, but 18% of them still rely solely on reactive monitoring, meaning post-hoc logs and no real-time kill switch. That is nearly the same as the 23% among smaller firms. The takeaway here is blunt: size does not buy you discipline. What buys you discipline is instrumenting your stack before you need to, not after. Builders are already investing 31% of their orchestration spend on agent monitoring and debugging, and 30% on security and permissions enforcement. That is the right instinct, and it is also the clearest signal that the market is still early. When you spend more on watching the agents than on building new ones, you are still in the containment phase, not the expansion phase. And the fact that only 7% of respondents name end-user experience as a top priority tells you everything you need to know about where the industry is: still obsessing over plumbing, not polish. The most useful lens for understanding this moment comes from the honest self-assessment in the data. Only 2% of builders say 76% or more of their systems are truly autonomous, and 35% admit that just 1-25% of their "agents" are real orchestration. Most are still chatbots in an agent costume. That is not a criticism, it is a reality check. The companies that succeed here will be the ones that treat agentic AI as an operational discipline, not a feature rollout. They will build the custom harnesses, the proxy middleware, and the dynamic routing that offloads heavy work to low-cost models, because they know the vendor is not going to save them from themselves. The open question to watch is whether Microsoft, which leads in primary usage today, can close the visibility gap that is pushing 43% of builders toward Anthropic's Claude Agent SDK next. If they cannot, the migration will not be a trickle, it will be a stampede. For now, the smart money is on the teams that assume nothing from their vendors and verify everything themselves.