New research from Cisco puts a number on something business leaders already sense: there is a massive gap between experimenting with AI agents and actually deploying them.

Published on March 23, 2026 at RSA Conference, Cisco’s “Agent Trust Gap” report surveyed senior IT and security leaders across major enterprises. The headline finding: 85% of organisations are already experimenting with, piloting, or deploying AI agents. But only 5% have moved AI agents into broad production.

That’s an 80-point gap. Cisco named it the “agent trust gap.”

The contrast is stark. Nearly every large organisation is running some kind of AI agent initiative. Barely any of them have pushed it past the pilot stage.

Why the Gap Exists

The most common barrier isn’t technology. It’s security and governance.

Nearly 60% of security leaders said security concerns are their primary obstacle to broader agentic AI adoption. The specific fears named were:

• Agents acting beyond their intended scope
• Agents being tricked or manipulated (prompt injection, social engineering)
• Supply chain risks in the tooling and infrastructure agents rely on

Here’s the part that should concern every business leader: only 29% of security leaders rank securing agentic AI among their top three priorities for the coming year. They’re worried about it, but they’re not actively preparing for it.

Cisco’s 2025 AI Readiness Index adds another layer. Only 24% of organisations that have already deployed AI say they have the controls in place to govern agent actions with proper guardrails and live monitoring. Most enterprises are deploying agents into environments that are not ready to manage them.

The Production Gap Is Closing Fast

The current 5% production figure is almost certainly temporary. Gartner projects that by the end of 2026, 40% of enterprises will be running AI agents in production, up from under 5% today.

That’s a near-tenfold increase in less than a year.

The business implication is clear: organisations that are still “evaluating” in mid-2026 will find themselves behind. The window to set up proper governance, test deployments, and build internal capability is right now.

What This Means for Business

The agent trust gap is not just a security problem. It is a competitive problem.

If 40% of your sector will have AI agents running real workflows by the end of 2026, the question isn’t whether to deploy. It’s whether your organisation will be in the 40% or watching from the outside.

The research identifies the specific blockers: governance, security controls, and the institutional knowledge to manage non-human workers who have real access to real systems. These aren’t technical problems that a software purchase solves. They require strategic decisions about how AI fits into your organisation’s operations.

For most businesses, the conversation needs to shift from “should we pilot AI agents” (almost everyone already is) to “how do we move from pilot to production without creating risk we can’t manage.”

That’s a harder question, and it’s where most organisations are stuck.

The Right Way to Cross the Gap

The enterprises that will close this gap first share a few traits. They have a clear inventory of which workflows they’re automating and what authority each agent has. They have defined escalation paths when an agent encounters a decision it shouldn’t make alone. They have someone accountable for AI governance who isn’t just an IT function.

Most importantly, they started with a real business problem rather than a technology exploration. “We want to see what AI agents can do” is a pilot. “We want to reduce the time our team spends on X by 60%” is a deployment.

The 85% figure means your competitors are already in this conversation. The 5% figure means most of them are still figuring out how to cross to the other side.

Part of what keeps agents in the pilot stage is infrastructure complexity — sandboxing, session state, error recovery, tool orchestration. Anthropic’s Managed Agents, launched in public beta in April, addresses that directly by handling the infrastructure layer so teams can focus on defining what agents do, not how to keep them running.

---

If this is the kind of problem agents can help with, the free Working With Claude field guide is the practical next step. Thirty-two pages, no fluff. Get the free guide.