The news
VentureBeat's latest Pulse Research surveyed 157 enterprises on how they measure AI agent performance — and the headline number is hard to ignore: 50% have deployed an agent that passed internal evaluations and then caused a customer-facing failure in production. Only 5% fully trust automated evaluation today, yet two-thirds are already allowing or actively engineering toward zero-human-in-the-loop deployments.
Our take
This research is about enterprise AI infrastructure, but the signal matters directly to GTM and marketing teams — because your team is almost certainly sitting downstream of exactly this problem.
Here's the mechanism: when you adopt an AI-powered sales tool, a marketing automation layer with "AI features," or an agent your team has wired into your CRM or outbound sequence, you're inheriting someone else's evaluation gap. The vendor ran their evals. The vendor's evals don't model your specific data quality, your ICP, your pipeline stages, or the actual sentences your prospects send back. A passing benchmark in a controlled environment is not the same thing as a working agent inside your GTM motion.
The VentureBeat data makes one thing explicit that most enterprise AI teams are quietly experiencing: the autonomy is arriving faster than the assurance. Two-thirds of organizations are marching toward fully automated deployments while admitting their testing frameworks don't reflect real-world outcomes. That gap isn't going to close on a vendor's roadmap timeline — it closes through production observation, tight feedback loops, and someone with domain knowledge actually watching what the agent does when things get weird.
For GTM teams, this is the part that matters: you are the ones with the domain knowledge. You know when an AI-generated follow-up email sounds off. You know when a lead score doesn't match what the rep is seeing. You know when the "qualified" account the agent surfaced is a company that went dark two years ago. That knowledge has to be in the loop — not as a backup plan, but as a design constraint. An agent with no human review isn't more efficient; it's just failing faster and quieter.
The instinct to automate everything is right. The instinct to set it and forget it is where things go wrong.
The so-what
The takeaway isn't "slow down on agents." It's "know what you're inheriting when you deploy one."
Before you expand any agent's autonomy in your GTM stack — automated outreach, lead routing, pipeline updates, content generation — ask two questions: What does failure actually look like in this workflow, and who will catch it? If the answer to the second question is "the eval passed, so we're good," you have your evaluation gap.
The teams that get this right aren't the ones running the most sophisticated testing infrastructure. They're the ones who treat production as the real eval — and build enough review into their workflows that a human sees the edge cases before they become customer-facing failures.
---