One agent per employee — and almost nobody is watching.

The agent boom has produced a number nobody would have believed two years ago: enterprises now average one AI agent per employee. The governance behind them has not kept up — and that gap is where the next wave of incidents comes from.

The numbers, and they are wild

An August 2026 industry report found enterprise agent interactions grew 14x in the first half of the year, that organizations now average roughly one agent per employee — and that about 60% of those agents are over-permissioned, granted broad access by default (Opsin Labs, State of Agentic Adoption). Meanwhile WRITER’s 2026 enterprise survey reports 79% of organizations face adoption challenges — up double digits from last year. Deployment is sprinting. Oversight is walking.

How sprawl happens

Most of these agents are not built by engineering. They come from sales ops, customer success, finance — teams solving a real problem with a no-code builder on a Tuesday afternoon. Each agent individually looks harmless. Collectively you get hundreds of semi-autonomous processes with production credentials, no owner of record, no evaluation set, and no answer to the only question that matters in a regulated business: when this thing makes a decision, who can explain it afterwards?

What ungoverned agents cost

Three bills come due. Security: an over-permissioned agent is a standing credential waiting for a prompt injection. Accuracy: without an eval set, nobody notices drift until a customer or a regulator does. Audit: in healthcare, legal, lending and education, “the agent did it” is not a defense — you need the input, the rule applied, the output, and the human who reviewed it, on file. This is exactly the discipline gap that separates the 12% of AI projects that ship from the 88% that quietly get turned off.

Fewer agents, operated properly

The correction is not an agent freeze — it is portfolio discipline. Pick the workflows where an agent carries real economic weight, and run those few properly: least-privilege access, an evaluation harness that runs on every change, drift alarms, audit trails, and a named owner. That operating layer is precisely what our Agent Operations retainer exists to provide. One well-governed agent that owns a $200K workflow beats two hundred ungoverned ones that own nothing — and it is the only version a compliance officer will ever sign off on.