The Invisible Organization

A manager walks past the service desk and counts the queue. An engineer notices that the same ticket has come back for a third time. Someone overhears a complaint about a workaround that everybody uses and nobody authorized. These are terrible measurements: noisy, partial, distributed unevenly across whoever happens to be paying attention. They are also how an organization finds out what is actually happening, as distinct from what its reporting says is happening.

Agents absorb this kind of work, and the signals don't improve. They disappear. The queue is short because requests resolve on their own. The ticket stops recurring because an agent handles it, or handles something near enough to it. Nothing surfaces about the workaround, because agents don't complain about the tools they are given. What remains is a dashboard, and a dashboard shows what it was built to show.

Three recent studies, in different domains and by different researchers, give some shape to the disappearance. A study of AI coding tools found commits up 240 percent where autonomous agents were in use, while releases — the point at which code reaches actual users — rose 30 percent. An evaluation of enterprise agents found that 92 percent of runs producing correct-looking final output contained at least one process deviation along the way: wrong scope, a skipped step, a write to the wrong target. And a survey of nearly 6,000 executives across four countries put active AI use at 69 percent of firms, while those same executives reported essentially no change in labor productivity over the prior three years.

Put them together and a pattern comes into view worth naming: closure inflation. Activity rises, recorded completions rise, and the gap between what the organization has closed and what it has resolved widens in places the instruments were never pointed.

There is precedent for confusing the instrument with the condition. In January 2012, JPMorgan's chief investment office adopted a new risk model that cut its reported value-at-risk from $132 million to $66 million overnight, ending a formal limit breach without anyone touching the positions that caused it. Under the old model the breach would have run continuously into May. Under the new one there was no breach. The underlying exposure roughly tripled. That was a deliberate change, made by people who understood what it would do, which is not the situation most organizations are in now. The shape of the error carries anyway.

What makes the current version harder is that nobody decided to stop watching. The queue stopped being visible because the work that used to sit in it moves through a pipeline that reports completion. Observation loss as a byproduct is more difficult to notice than observation loss as a choice, and much more difficult to attribute to anyone in particular. It also lasts longer. The measurements in use were designed for human-mediated work and are now pointed at work that is no longer human-mediated, and nothing in the current adoption pattern is pressing anyone to re-instrument.

Organizations absorb the consequences of completed work — reviews, returns, complaints, reconciliations — at a pace set by human attention. Production can accelerate past that pace. Absorption mostly cannot. The difference accumulates somewhere, and the somewhere is rarely a place anyone built a chart for.

One way to check your own organization: find a stretch where volume went up and complaints went down at the same time. Sometimes that is exactly what it looks like, and the work got better. Losing sight of the work produces the same picture, and nothing in the picture tells you which one you are looking at.