The Three Measures: An AI ROI Scorecard That Survives the J-Curve

Every serious technology adoption dips before it rises. Economists call it the productivity J-curve: the investment that makes a new capability pay — redesigned workflows, retrained people, rewired data, new controls — is expensed long before the payoff is booked, so the measured return goes down first. Electricity did it. Enterprise software did it. AI does it harder, because the model is the cheap part and the workflow around the model is the expensive part.

The dip is not the problem. The dip is normal. The problem is the scorecard most programs carry into it. It measures one thing. Either it measures usage, in which case the program looks wonderful right up to the moment Finance asks where the money is; or it measures the P&L, in which case a program that is doing exactly what a good program does in month four — costing more than it returns while the work is being redesigned — gets cancelled by the number it was told to hit.

The fix is not a better number. It is three measures, run simultaneously, read together, on one page. Each answers a different question, moves on a different clock, and fails in a different way. Together they tell you where you are on the curve — which is the only thing a board actually needs to know to make the hold-or-stop decision well.

Measure one — Inputs

The question: are people and agents actually using it, and is it costing what we said?

Inputs are the leading measure. They move first and they move fast: weekly active users as a share of the target population (not the company, the population the workflow was built for), tasks routed through the new path, agent actions taken, and spend — seats, tokens, vendor fees — against the limits written in the mandate. The spend line comes straight from the inventory; if it cannot, the inventory is not finished.

Two rules keep this measure honest. First, telemetry, never surveys: what people say they use and what the logs say they use diverge by a factor of two in every deployment I have watched. Second, persistence, never launch: report week twelve, not week one. Adoption that is still there after the novelty is the only adoption that counts.

Measure two — Outcomes

The question: is the instrumented workflow measurably better, in units Finance will sign?

Outcomes are the lagging measure, and the one the J-curve punishes. Cycle time, throughput, rework or error rate, cost per unit — baselined before the program on the same workflow and the same population, then measured after. Not company-wide productivity, not a satisfaction score, not an estimate of hours saved multiplied by a loaded rate that nobody agreed to. One workflow, one baseline, one delta.

Then two columns, never merged. The telemetry-verified column is what the systems show: minutes removed from the cycle, tickets closed per person, defects avoided. The Finance-verified column is what the CFO will book: cost taken out, capacity redeployed to named work, revenue that would not otherwise have arrived. The distance between the two columns is the realization gap, and closing it is the program’s job. Reporting the first column as if it were the second is how AI programs lose the CFO.

Measure three — Organization

The question: has the work itself changed?

This is the measure almost nobody keeps, and it is the one that predicts whether the dip ends. The J-curve is paying for intangible capital: the share of the workflow that has actually been redesigned rather than decorated with a chatbot; the roles whose job descriptions changed; the decisions that moved — from a person to an agent, or from a committee to a person with a mandate; the skills built; and the controls in place, which in this series means every agent on the inventory, every mandate signed, every evidence feed running.

Organization is scored, not measured, and that is fine. Five yes-or-no questions, answered by the workflow owner and checked by whoever owns the inventory, are enough. A rising organization score with flat outcomes means the payoff is coming. A flat organization score with rising inputs means you have bought a tool and changed nothing, and the outcomes will never arrive.

How to read the three together

The scorecard earns its keep in the reading, not the collecting. Four patterns cover almost every program I have seen.

Building it in a week

Day one, pick one workflow. Not the program — one workflow inside it, with an owner, a population, and a unit of work you can count.

Day two, baseline it. Cycle time, throughput, rework, cost per unit, from the last full quarter. If the baseline does not exist, the program starts with a measurement sprint, not a model.

Day three, wire the inputs. Active users, tasks routed, agent actions, and spend, from the systems, on a weekly pull. Survey nothing.

Day four, agree the Finance rule. Sit with the controller and write down, in advance, what evidence turns a telemetry number into a booked number. This one conversation prevents the month-nine argument.

Day five, set the cadence. Inputs weekly, outcomes monthly, organization quarterly, all three on one page in the same review. The workbook below is that page.

What to do now

Put three numbers on the board page, not one. If the program is currently reported by a single metric, the next review is the moment to add the other two.

Separate the columns. Telemetry-verified and Finance-verified value never share a cell. Report the gap between them as a number.

Baseline before you buy. No pilot starts without last quarter’s cycle time on the workflow it claims to improve.

Score the organization quarterly. Five questions, answered by the workflow owner. It is the earliest signal that the dip will end.

Tie authority to the reading. Only pattern four widens a mandate. Patterns one and two hold it; pattern three withdraws it.

This is the third piece in the Operator Edition of Field Notes on the Agentic Enterprise. Previous: Write the Mandate. Start of the series: The Agentic Economy Runs on Authority.