Enterprises just handed the keys to agents they don't trust, running on tests they know are broken, calling chatbots "orchestrated workflows" to justify the spend.
The Summary
- Half of enterprises shipped agents that passed internal evals and then failed customers in production, and 57% traced confident-but-wrong answers to missing or inconsistent business context
- Only 5% of organizations fully trust automated evaluation today, yet two-thirds are engineering toward deploying agent changes on automated eval alone with no human in the loop
- 71% admit most of their deployed "agents" are single-prompt chatbot wrappers, not true multi-step orchestrated workflows
- Anthropic's Claude captured 40% of enterprise orchestration share, more than double any rival, driven by native model alignment
The Signal
VentureBeat surveyed 258 enterprises across three waves and found the same pattern everywhere: the infrastructure for autonomous agents is being deployed faster than the trust to run them safely. The gap shows up in three places at once. Context systems feed agents information they can't verify. Evaluation platforms gate production with tests nobody believes. Orchestration layers coordinate workflows that don't actually exist yet.
Start with the context gap. RAG is the default source of business context for enterprise agents, and provider-native retrieval has quietly overtaken dedicated vector databases. But 57% of enterprises watched their agents produce confident, wrong answers in the past six months, traced directly to missing or inconsistent context. More than half of those said it happened multiple times.
"A governed semantic layer is emerging as the fix, but most are still building it."
The field is converging on hybrid retrieval and a governed semantic layer as the solution, but the architecture is still under construction. Meanwhile, agents continue sounding authoritative on foundations their operators don't trust. The core issue isn't retrieval accuracy. It's that enterprises don't have a single source of truth for what their agents should know, and RAG systems are exposing that gap at scale.
The evaluation gap is worse. Half of enterprises deployed an agent or LLM feature in the past year that passed internal evals and then caused a customer-facing failure. A quarter saw it happen more than once. Trust in automated evaluation is nearly nonexistent: only 5% say they fully trust it.
The most-cited weakness: evaluations don't align with real-world outcomes. Yet two-thirds of enterprises are already allowing, or actively engineering toward, deploying agent changes to production on automated eval alone. No human review. The tests are known to be unreliable, and enterprises are removing the safety check anyway because the alternative is slowing down deployment velocity.
Key evaluation failures cited:
- Evals pass agents that fail real customers
- Tests don't measure what actually breaks in production
- Automated metrics optimize for the wrong outcomes
The orchestration gap closes the loop. Enterprises are consolidating fast onto major model platforms, with Anthropic's Claude holding 40% share, Microsoft at 18%, and OpenAI at 13%. The driver is "model gravity": native alignment with a state-of-the-art base model. Success is judged by task completion reliability and multi-step workflow management.
But 71% of enterprises admit that a quarter or fewer of their deployed "agents" are true multi-step orchestrated workflows. Most are single-prompt chatbot wrappers. The label "agent" is being applied to anything with an LLM backend, regardless of whether it exhibits autonomy, planning, or tool use. The orchestration layer exists. The orchestrated workflows mostly don't.
"Real-time fiscal control over token burn remains the exception."
What's driving this? Pressure to ship. Enterprises see competitors deploying agents and feel compelled to match the pace, even when the infrastructure underneath isn't ready. The result is a stack where every layer is compromised: agents run on context they shouldn't trust, pass evals that don't predict production behavior, and get called "orchestrated" when they're glorified autocomplete.
The Implication
If you're building agents for enterprise, the gap between what's deployed and what's trustworthy is your opening. The market wants governed semantic layers that make context verifiable, evals that align with real-world outcomes, and orchestration platforms that actually coordinate multi-step work. The current stack is provisional, and everyone knows it.
If you're buying, don't trust the label. Ask how many deployed "agents" are true multi-step workflows. Ask what percentage of evals have predicted production failures. Ask how context conflicts get resolved when RAG pulls from three different sources. The answers will tell you whether you're buying an agent platform or expensive vaporware.