A quiet threshold has been crossed in how businesses use AI. A year or two ago, an organization with an AI agent - a piece of software that does not just answer questions but carries out multi-step work on its own - typically had one, as a pilot, with a team watching it. Recent industry surveys suggest the average organization now runs more than a dozen, nearly triple the count of a year earlier, and building a new one has become fast enough that teams spin them up the way they once created spreadsheets. The capability story here is genuinely positive. The management story is the one arriving late: many companies can no longer say with confidence how many agents they are running, what each one touches, or who is responsible for it.
This is a familiar problem wearing new clothes. Businesses have been through it with unofficial software, with cloud subscriptions signed up on company cards, with automation scripts written by someone who later left. The pattern is always the same: when creating a thing becomes easy, the population of things grows faster than anyone's list of them, and the gap between what exists and what is known to exist becomes a risk in itself. Agents sharpen the pattern because they act. A forgotten document is inert; a forgotten agent is still doing something - sending messages, updating records, making small decisions - according to instructions that made sense when someone was watching and may not make sense now. The failure mode is not dramatic malfunction but quiet drift: an agent faithfully executing a process the business changed two quarters ago.
The specific hazard worth naming is the orphaned agent - one whose creator has moved teams or left the company, whose purpose is half-remembered, and which keeps running because nothing is broken enough to force the question. Orphaned automation has always existed, but an orphaned agent typically holds access to real systems and acts with a latitude that scripts never had. When something does go wrong, the first hour is spent discovering what the agent even was; when an audit or a customer question arrives, nobody can explain why the agent did what it did, because the person who could is gone and nothing was written down. None of this requires the AI to be faulty. It only requires the organization to have treated deployment as the finish line rather than the start of an ownership obligation.
The remedy is unglamorous administration, applied early. Keep an inventory: every agent, what it does, what it can reach, and - non-negotiably - a named current owner, a role rather than a memory. Give agents lifecycles, not just launches: a review date, and a presumption that an agent nobody will claim gets switched off rather than left running out of caution. Fold new agents into this at creation, when the details are fresh, because reconstructing them later is archaeology. And treat the rising count itself as a management signal: a dozen agents is a dozen sets of standing instructions acting on the business's behalf. Companies learned - expensively - to keep track of who has keys to the building and who can sign on the account. Software that acts has joined that category, and the habit of knowing what is running is much cheaper to build at thirteen agents than at a hundred.