Most companies claim AI agents, few run them

Two numbers from this year's industry research describe the same market. Ninety-seven percent of executives say their organisation has deployed AI agents. Fewer than ten percent of companies have agents running in production. Both are real, and they cannot both be true unless "deployed" has stopped meaning what everyone thinks it means.

It has, and the rest of this piece sets out where the gap comes from and how to find out which side of it your organisation is on.

What "deployed" has come to mean

The gap comes from definition drift. In most organisations, "we've deployed agents" now covers any of the following:

• Licenses purchased for the engineering organisation
• A copilot suggesting code in the IDE
• A hackathon or innovation-week prototype
• A proof of concept summarising tickets in a sandbox

Each is a reasonable first step. None of them is an agent in production, which is a different thing with a much higher bar.

What production actually requires

An AI agent in production is a system that takes actions in live business systems, under scoped permissions, with evaluation gates and a named owner who answers when it fails. That definition filters out almost everything currently reported upward as deployment, and every word of it is load-bearing:

• Live actions, against real data, where a failure has a cost
• Scoped permissions, not an admin service account
• Evaluation gates that check outputs before anything irreversible happens
• A named owner with rollback authority and a pager

Cost has joined integration as the reason agents get pulled. KPMG's Global AI Pulse for Q2 2026, a survey of 2,145 senior leaders in 20 countries, found that 49% had scaled back agent deployments because operating costs outweighed the benefits. The agent is the easy part now. The API surface your legacy systems never had, the permission model, the eval harness: that's the project.

Why the gap persists

Buying licenses is procurement. It's fast, visible, and reportable to the board by Friday. Building integration, evals, and ownership is delivery work: slow, invisible, and reported only when something ships. Organisations under pressure to show AI progress choose the visible motion, and the ninety-seven percent number measures exactly that. Motion.

What the under-ten-percent do differently

The companies running agents in production treat them as a delivery problem from day one:

• Pick one workflow with measurable value, and scope it tightly
• Build the API access and the eval set before scaling anything
• Give the agent a named owner with rollback authority
• Expand only after the first workflow holds in production

Unremarkable steps. It's the same discipline that ships any production system, applied to a technology most of the market still treats as a procurement category. And the advantage compounds: production teaches integration debt, eval discipline, and failure modes that license-holders never encounter, and all of it carries to the second agent and the third.

One question to ask

Take this into your next AI status update: name one agent action that ran in production last week, what it touched, and who owns it. A concrete answer puts you ahead of roughly ninety percent of the market. Silence means you've been reporting procurement as progress, and the fix starts with scope, not spend.

The same discipline applies to the engineers who build these systems. Our assessment tests whether someone notices when a model is confidently wrong, and whether they review generated output under pressure. See how we vet.