The numbers are striking. The AI agent market crossed $7.6bn in 2025 and is projected to exceed $50bn by 2030. Gartner predicts that 40% of enterprise applications will include task-specific AI agents by 2026. Inquiries about multi-agent systems surged 1,445% between the first quarter of 2024 and the second quarter of 2025. If you believed the headlines, every boardroom in the world would already be humming with autonomous digital workers.

Only 11% of companies have AI agents in production. That figure — buried in the same research that generated all the excitement — is the one worth keeping in mind.

The gap between interest and action

The breakdown is instructive. About 30% of organisations are still at the exploration stage, 38% are running pilots and 14% describe themselves as deployment-ready. That leaves just over one in ten that have actually crossed the line. The pattern is familiar: a technology generates genuine enthusiasm, boards demand a strategy, teams run experiments and then discover that the distance between a working prototype and a reliable production system is longer than anyone budgeted for.

This is not scepticism about AI agents as a category. The technology is real and the use cases are compelling. Agents that can browse the web, write and execute code, query databases and hand tasks off to specialist colleagues represent a qualitative leap beyond the chatbots of a few years ago. The question is not whether they work in controlled conditions but whether organisations can deploy them safely, consistently and at scale.

What is actually changing

The architecture of AI systems is shifting in a way that matters. Early deployments leaned on a single general-purpose agent asked to do everything. The emerging model is different: an orchestration layer that coordinates a team of specialist agents, each with a defined role and a limited scope. One agent retrieves data; another analyses it; a third drafts a report; a fourth checks it for errors before sending. The whole is more reliable than any one part would be alone, and failures are easier to isolate.

Two new standards are accelerating this shift. The Model Context Protocol (MCP) gives agents a common way to connect with external tools and data sources, removing the need to build bespoke integrations for every system. The Agent-to-Agent (A2A) protocol allows agents built on different platforms to communicate and collaborate without human mediation. Together they are doing for AI agents something like what HTTP did for the web: making interoperability the default rather than the exception.

Low-code platforms are pushing further in the same direction. Vendors now promise that a non-technical user can deploy a working agent in 15 to 60 minutes. That claim deserves scrutiny — deploying something and deploying something that works reliably in a live environment are different things — but the trajectory is real. The barrier to experimentation has fallen sharply.

Why production remains hard

If the tools are improving and the interest is genuine, why are only 11% of companies in production? Several reasons converge.

Trust is the first. An agent that autonomously sends emails, updates records or executes transactions needs to be demonstrably reliable before anyone will let it run unsupervised. A single high-profile failure — a wrong decision made at machine speed, at scale — can set back an entire programme. The organisations that have reached production tend to be those that have invested in testing, monitoring and the ability to intervene quickly when something goes wrong.

Integration is the second. Most enterprises run on systems built over decades, with data scattered across platforms that were never designed to talk to each other. Getting an agent to act usefully in that environment requires connecting it to the right sources of information and the right points of action. MCP helps, but it does not eliminate the underlying complexity.

Governance is the third, and perhaps the most underrated. Agents that make consequential decisions raise questions that go beyond the technology: who is responsible when something goes wrong, how decisions are audited, what guardrails exist and who sets them. Large organisations move slowly on these questions, and sensibly so.

Where this leaves us

The gap between the hype and the 11% is not a reason to dismiss AI agents. It is a reason to think clearly about what deployment actually requires. The companies that will benefit most from this technology in the next few years are unlikely to be those that moved fastest in 2025; they are more likely to be those that spent that time building the infrastructure, governance and institutional knowledge needed to run agents safely at scale.

The market will grow. The standards will mature. The tools will improve. But the distance between a pilot that impresses a steering committee and a system that runs reliably in production has always been where technology projects succeed or fail. AI agents are not exempt from that rule, whatever the projections say.

Eleven percent is not a ceiling. It is a starting point. The question for most organisations is not whether to get there but how long it will take — and how much it will cost to find out.