Enterprises Now Run Three AI Orchestration Platforms at Once, and a Fifth Can't Stop a Runaway Agent From Spending

Enterprise IT departments were sold on the idea that they'd pick one AI platform and build around it. According to a recent survey, that's not happening.

According to VB Pulse survey data covering 107 enterprises, the median company now runs three separate AI agent orchestration platforms at the same time. Eighty-five percent use two or more. Sixty-four percent use three or more. Only 15% have committed to a single platform.

This reflects deliberate strategy, not indecision.

Why enterprises refuse to pick one vendor

Microsoft AI Foundry and Copilot Studio show up in 70% of enterprise stacks, the most of any platform, according to VentureBeat. OpenAI's Agents SDK is close behind at 68%, and Anthropic's Claude Platform sits at 47%. Google's Enterprise Agent Platform, LangChain/LangGraph, Salesforce Agentforce, Amazon Bedrock, and LlamaIndex all show up too, and 22% of enterprises have built their own in-house orchestration on top of all of it.

A companion analysis from novalogiq, drawing on the same survey, breaks down why. Flexibility across models and tools was the top purchase driver at 29%, nearly three times the 10% who said they picked their platform because it plays nicely with a specific state-of-the-art model. Security and permissions concerns drove 17% of decisions, production reliability another 15%.

Companies aren't buying based on which AI is smartest. They're buying based on which platform lets them escape if they need to.

The trust problem is about security, not lock-in

Novalogiq's analysis found the risk enterprises most associate with letting a provider control the orchestration layer isn't vendor lock-in. It's security and permissioning limitations, cited by 37% of respondents, well ahead of lock-in itself at 23% and limited visibility at 22%.

Enterprises don't fully trust Microsoft, OpenAI, Anthropic, or anyone else to police what their own AI agents are allowed to do. So they're building their own guardrails on top, or running multiple platforms so no single vendor's blind spot becomes the company's blind spot.

That distrust shows up in where the money goes. Agent monitoring and debugging draws the largest share of investment at 31%, according to novalogiq, with security and permissions enforcement close behind at 30%. Workflow tooling, the stuff that actually makes agents do useful work, gets only 19%. Enterprises are spending more to watch their AI than to build with it.

One in five can't stop the bill in real time

One in five enterprises has no real-time mechanism to halt a runaway AI agent's spending before the invoice lands. An agent that goes sideways, calling the wrong APIs, looping on token-heavy tasks, or simply misfiring, can burn through budget with nobody able to pull the plug until after the damage is done.

This represents a genuine governance gap. Novalogiq also found that most companies overstate how "agentic" their systems actually are. A plurality of 47% of respondents said only 26% to 50% of what they call "agents" are genuinely orchestrated systems making autonomous decisions. Another 37% put that figure at a quarter or less. Much of what enterprises are marketing internally as agentic AI is still, functionally, a chatbot wearing an agent label.

What comes next

The industry isn't settling down. More than two-thirds of surveyed enterprises plan to change orchestration platforms within the next year, according to VentureBeat: 15% within three months, 24% within three to six months, and 28% within six to twelve months.

Anthropic's Claude Agent SDK is the platform most enterprises say they're evaluating next, drawing interest from 43% of respondents, well ahead of Microsoft's own forward consideration. Roughly a third are looking at Google's Enterprise Agent Platform, another 31% at building more in-house, and 25% at OpenAI's tools.

By the end of 2026, 53% of enterprises expect their primary control plane to be hybrid, mixing provider-native tools with external, model-agnostic orchestration, according to novalogiq's read of the same data. Only 14% expect to rely on a single provider-managed service.

The unresolved question is whether any vendor, Microsoft, Anthropic, OpenAI, or Google, closes the security and metering gap fast enough to earn the trust enterprises are currently withholding. Until then, the default enterprise strategy is to run several platforms at once and hope the redundancy catches what any one vendor misses. Whether that redundancy actually stops the next runaway agent's bill, or just spreads the risk across three dashboards instead of one, is not something this survey answers.