Intelligence per worker, measured honestly.

A metric arrived at the wrong moment

"Intelligence per worker" is a good idea wearing bad timing. The idea is sound: as AI absorbs routine cognitive work, the unit that matters is no longer headcount and no longer raw output, but how much useful thinking a company can field per person on the payroll. A firm with the same number of people producing more decisions, more analysis, more resolved cases — at the same or better quality — has more intelligence per worker. The frontier research is right that the gap between the companies that built this and the ones that did not is now wide, and widening.

The bad timing is that the phrase landed in the same news cycle as the cuts. When a board hears "we should raise intelligence per worker" three weeks after a competitor cut 20% of its staff and called it an AI restructuring, the metric does not arrive neutral. It arrives pre-loaded. And a metric that can be read two ways will be read, by someone in the room, the way that is easiest to action — which is almost always the denominator, because cutting people is a decision a CFO can make on a Tuesday, and growing the numerator is a programme that takes a year. We are not going to pretend that tension away. The whole value of defining the metric carefully is to make the lazy reading visibly wrong.

Intelligence per worker is a ratio. You can move a ratio by lifting the top or by cutting the bottom, and only one of those is a capability. The other is just a smaller company.

This is the same fault line we wrote about from two other angles — the ROI gap (why most AI spend never reaches the P&L) and the capture gap (where the saved hour physically leaks before it becomes margin). This piece is narrower and more forensic than either: not "did it pay back" and not "where did the hour go", but what is the number, exactly, and how do you keep it honest. Because the version of intelligence per worker that survives a board meeting is not a slogan. It is a numerator and a denominator, each of which someone can audit.

What the metric actually measures

Strip the phrase down. The numerator is useful cognitive output — work that required judgement, not just motion. The denominator is the human cost of producing it — not only salaried hours, but the things that erode when you push the ratio the wrong way: retention, institutional knowledge, exception quality, the slow tax of burnout. Most dashboards measure a thin version of the numerator (volume) divided by a thin version of the denominator (headcount), get a flattering number, and miss the point. The honest metric is harder to fake because it forces you to define useful on the top and cost on the bottom in terms a skeptic would accept.

Four components make it auditable. None of them is exotic; the discipline is in measuring all four and refusing to report one without the others. Each has a tell — a way you can see it being gamed — and that is the column that matters.

Read the four together and the metric is hard to abuse: you cannot claim a rise in intelligence per worker by cutting the denominator alone, because component 3 will show time recovered with no destination, component 4 will show exception quality falling as the experienced people leave, and component 1 — if you count outcomes rather than activity — will refuse to move just because the headcount did. The four are a system of checks. Report one in isolation and you have built a vanity number; report all four and you have built something a board can actually steer on.

The denominator is the whole metric

Here is the trap stated plainly: the numerator is where the capability lives, and the denominator is where the damage hides. A company that lifts the ratio by growing the numerator — more decisions per person, at quality — has built something durable. A company that lifts the same ratio by shrinking the denominator has booked a number that looks identical on the slide and behaves nothing alike over time. The two are indistinguishable for exactly one reporting period, which is precisely long enough to make the wrong decision look right.

The cost of the denominator move arrives on a delay, which is what makes it so easy to take. Cut the experienced operators and the ratio jumps this quarter; the exception-handling quality they carried in their heads degrades next quarter; the rework, the escalations, and the re-hiring land two quarters after that, by which point the original decision is three meetings in the past and nobody connects the line on the cost report to the line on the quality report. The numbers that would have caught it — component 4's quality delta, the tenure curve, the share of exceptions resolved on first touch — were the ones a numerator-only dashboard never carried. This is the six-months-of-gain, year-of-cost pattern, and it is structural, not a failure of anyone's good intentions.

There is also a compounding effect that the layoff reading gets exactly backwards. The companies pulling away on intelligence per worker are not the ones that cut hardest; they are the ones whose people moved up the value chain fast enough to use the freed capacity. That move — from doing the routine work to directing the agents that now do it — is the executor-to-orchestrator shift, and it only happens if the skills arrive with the automation rather than behind it. Cut first and upskill later and you have manufactured the worst denominator outcome: the routine work is gone, the people who could have grown into the new work are gone with it, and the remaining team is doing more volume at lower judgement. The ratio looks great for a quarter. The capability is lower than when you started.

Measuring it without it becoming a layoff number

If the metric is going to live at the board level — and it is — the governance around it has to make the honest reading the default and the lazy reading expensive. Four practices do most of the work.

Report the ratio with its decomposition, never as a single number

A board that sees "intelligence per worker up 22%" learns nothing it can steer on. A board that sees the same 22% broken into the four components — outcomes per FTE up, delegable-task rate up, time recovered with named destinations, quality delta flat or positive — can tell a capability gain from a headcount cut at a glance. Mandating the decomposition is the single highest-leverage rule, because it removes the ambiguity the euphemism depends on.

Pair every denominator move with its quality and tenure leading indicators

If a reorganisation is going to shrink the denominator, the same paper has to carry the leading indicators that would catch the delayed cost: first-touch exception resolution, rework rate, tenure distribution, time-to-competence for the people inheriting the redesigned roles. Not as a compliance footnote — as the gating metrics that the decision is reviewed against ninety days later. A cut that improves the ratio and degrades all four leading indicators is not a productivity win, and the only way to know is to have written them down before.

Upskill in parallel with automation, on the same timeline

The freed capacity converts to a higher numerator only if the person can do the higher-value work the moment the routine work leaves — not in a training module scheduled for next quarter. Run the upskilling on the automation's own timeline and the recovered hour becomes judgement; run it behind and the hour becomes slack, the ratio stalls, and the pressure to "fix" it falls back on the denominator. The sequencing is the whole game.

Make the numerator count outcomes, and audit the definition

Everything above fails if "output" quietly means activity. The numerator has to be outcomes a customer or a downstream team actually received — a decision approved, a case closed inside SLA, an analysis a human acted on — and the definition has to be audited the way a finance number is, because the incentive to soften it toward "tasks completed" is constant. The day the numerator drifts to activity, the metric is back to vanity and the denominator is back in charge.

What this means for the document layer

Intelligence per worker is abstract until you put it on a concrete desk, and the document desk is the cleanest one we know — high volume, rules-bound, and, usefully, the rare function where all four components of the metric are already numbers you can pull rather than estimates you have to argue about. That is not a coincidence we are being modest about: if you want to measure intelligence per worker for real before betting an org on it, a document workflow is where the measurement is honest first.

The numerator is the decision, not the page. "Documents processed" is the activity trap from component 1, dressed for the back office — it rises when the programme works and tells you nothing about useful output. The number that maps to the numerator is the decision rate: the share of documents that clear straight through, per class, inside the stated accuracy. A platform that can only report volume hands the operator the vanity numerator and hides whether intelligence per worker moved at all.

The delegable-task rate is the auto-resolution rate. Component 2 — the share of safely-delegable work actually delegated — has a direct reading on the document desk: of the cases the policy says can clear without a human, how many do, and at what error rate. That single number is the ceiling on how much judgement the desk frees, and it is honest only if it is reported next to the error rate, so "delegated more" cannot quietly mean "lowered the bar".

The quality delta is measurable per class, if you look. Component 4 lives or dies on whether you can see accuracy and rework on the same work with and without the agent. On documents you can, per class, against a held-out baseline — which means a throughput gain that costs you a quiet rise in downstream defects gets caught here instead of in a dispute three months on. That only works if the audit trail is honest enough to reconstruct why each field was decided, so the quality number is evidence and not a vibe.

And the recovered hour has to land on judgement, not re-checking. The document-desk version of capture is the processor becoming an exception handler and a threshold owner — the routine reading absorbed by the agent, the freed hour pointed at the cases it flags and the policy that decides what clears. That is the numerator growing and the denominator holding, which is the only version of a rising ratio worth defending. It needs per-field confidence and a clean escalation path rather than one all-or-nothing answer; without them the "freed" hour is spent re-checking the agent by hand, and you have moved the denominator without moving the capability.

Closing thought

The uncomfortable thing about intelligence per worker is not that it is a bad metric. It is a good one — arguably the right one for a year in which the binding constraint stopped being how many people you can hire and became how much judgement you can field per person. The risk is entirely in the reading. Measured as a single number, it is a layoff justification with a respectable name. Measured as four components a skeptic can audit — outcomes per FTE, delegable-task rate, time recovered with a destination, quality held or improved — it is a map of where a company actually built capability and where it only got smaller. The board has started asking for the number. The companies worth betting on are the ones that hand back the decomposition without being asked.

At Cogneris we build the document layer so that all four components are measurable on day one: a decision rate per class instead of a page count, an auto-resolution rate reported next to its error rate, a quality delta against a held-out baseline so a throughput gain that costs accuracy is visible before it is expensive, and per-field confidence with a clean exception path so the recovered hour becomes judgement instead of re-checking. If your board has started asking about intelligence per worker and you want the honest version — the numerator that counts decisions and the denominator that holds — talk to our team and we will help you measure it on a real workflow. The ratio is the right question. The decomposition is the honest answer.