The Ten-Slide Problem: What Do We Measure When Output Is Free?

The Ten-Slide Problem: What Do We Measure When Output Is Free?

Status
Draft
Author
Created
Nov 1, 2026
Tags
Human Judgment
Workforce Transformation
Opinion
Text
When an agent can produce a month of analysis overnight, volume stops being a signal. Judgment is the new KPI.
Last quarter, a consulting team I know delivered a sixty-page market-entry analysis in two days. A year ago the same work would have taken six people three weeks. The partner was delighted. The client was impressed. And in the review meeting, someone asked a question that nobody could answer cleanly: who, exactly, did this?
Not who typed it. Who decided it. Who looked at page forty-one, saw a confident claim about regulatory timelines, and thought, "that's wrong, and here is why"? In this case, the honest answer was: one senior associate, for about forty minutes, late on the second night. Everything else was generated, assembled and formatted by agents. The forty minutes were the engagement.
That is the ten-slide problem. When producing ten slides took a day, ten slides were a reasonable proxy for effort, skill and commitment. We built our performance systems on that proxy: hours billed, deliverables shipped, decks produced. Now the proxy is free, and most organisations are still measuring it.

Goodhart arrives at the office

Goodhart's Law says that when a measure becomes a target, it ceases to be a good measure. AI agents are the most efficient Goodhart accelerant ever built. Point them at "documents per week" and you will get documents. Point a team at "output per person" and the number will rise tenfold while the thing the number was supposed to represent, useful thinking, quietly stays flat or declines.
The uncomfortable corollary is that volume is no longer evidence of anything. A person who ships twenty artifacts a week may be a brilliant orchestrator or a very fast conduit for unreviewed machine output. From the dashboard, they look identical.

The contribution moved upstream and downstream

If you watch where human value actually shows up in agent-augmented work, it clusters at the two ends of the process, and almost never in the middle.
Upstream is framing: deciding which question is worth asking, what "good" looks like, which constraints are real and which are inherited habit. An agent will answer any question with equal enthusiasm, including the wrong one. The person who frames the problem well has done most of the work before the first token is generated.
Downstream is verification and ownership: noticing the plausible-but-wrong, deciding what is safe to put in front of a client, and putting your name on it. Lisanne Bainbridge wrote about this in 1983 in her paper on the ironies of automation. The more reliable the machine, the harder the human's residual job becomes, because the human is left with the exceptions, and without the daily practice that builds the instinct to catch them. Pilots now know this tension intimately. Knowledge workers are only beginning to.
The middle, the drafting, the formatting, the first-pass synthesis, is where agents are strongest and where humans add the least marginal value. Which means that most of our current KPIs are measuring the one segment of the work that has become cheap.

What to measure instead

I do not think the answer is a cleverer dashboard. But a few principles seem to hold up in the engagements I have seen go well.
First, measure decisions, not documents. How many consequential calls did this person make, and how many survived contact with reality? A short log of judgment calls (what was rejected, what was overridden, what was caught) tells you more than a folder of deliverables.
Second, measure catch rate. When an agent's output contained an error, did the human find it before the client did? This is awkward to track and impossible to game cheaply, which is exactly why it is worth tracking.
Third, make ownership explicit at the point of delivery. Every artifact that leaves the building should have a named human who can defend any line in it without opening the chat history. If nobody can, the artifact is not finished, however polished it looks.
Finally, protect the apprenticeship. If agents do all the junior work, the next generation of seniors never builds the judgment we are now claiming is irreplaceable. Firms that treat this as a training-design problem, deliberately giving juniors review-and-defend reps, will have partners in ten years. The rest will have very expensive prompt operators.

The forty minutes

Go back to that associate and her forty minutes. In the old model, she would have been invisible in the metrics: a modest contributor to a large pile of output. In an AI-first model, she is the entire reason the firm can still sign its name to the work.
The organisations that thrive will be the ones that learn to see forty minutes of judgment as clearly as they once saw three weeks of effort. Because when output is free, the only scarce thing left in the building is someone willing to be accountable for it.