Skip to content
finalstate.aiThe AI Wins You Can Measure Are Table Stakes
004Atlas

004THE MEASURED WIN

The AI Wins You Can Measure Are Table Stakes

A study of 5,172 support agents estimated roughly 15% more issues resolved per hour. That operating result leaves costs, competitive response and retained value to be tested.

A support ticket connects to an operating measurement and an unanswered retained-value field.A ticket reaches a clock representing resolutions per hour. A separate route opens an empty retained-value field. No ticket count, time saving or profit is invented.SupportMeasuredRetained?
Support example; no retained-value estimate is supplied. Costs and competitive response remain open.

THE FIELD RESULT

One firm resolved more issues per hour

A fixed bar shows an estimated increase of roughly 15 percent in issues resolved per hour.The bar runs from zero to 15 percent on a zero-to-20-percent axis. It is the rounded overall estimate in the final 2025 study of 5,172 agents, not a profit measure or subgroup comparison.Resolutions/hour~15%0%5%10%15%20%
Rounded overall productivity estimate. Section V.A, Table II, column 3 reports +0.301 resolutions per hour (SE 0.0329), equivalent to 15.2% of the pretreatment mean in the text; 12,295 agent-month observations in that specification.Brynjolfsson, Li & Raymond, QJE 140(2), 2025, pp. 889–942; DOI 10.1093/qje/qjae044
  • 5,172 customer-support agents at one firm
  • Roughly 15% higher issues resolved per hour
  • Staggered rollout; access was not randomly assigned

Brynjolfsson, Li and Raymond's final 2025 study estimates the effect using difference-in-differences. Agents could edit or ignore suggestions. The study supplies no estimate of rival adoption, price pass-through or durable competitive advantage.

A shared gain can still be worth buying

Faster support might reduce a queue, lower a price or raise the service customers expect. Those are possible responses, not findings of the study. A profitable gain need not be exclusive. But throughput alone does not justify deployment: quality, tool fees, integration, checking and upkeep belong in the operating case. The retention test asks who benefits over the chosen horizon.

THE SEAM

Better prediction changes which judgment pays

A proposed support response and a cost record meet before an unresolved action branch.A prediction route approaches the choice between a reply and escalation. A separate cost record joins before that choice. Both action paths remain open; the illustration does not make judgment automatically human, scarce or correct.PredictionCostReplyEscalate
Illustrative support application. The source is a conditional economic model, not a field test of this workflow or a current boundary on machine ability.Agrawal, Gans & Goldfarb, NBER Working Paper 24626, May 2018, §3, pp. 4–6, Proposition 1

In a May 2018 working paper model, Agrawal, Gans and Goldfarb make better prediction a complement to judgment about hidden costs and a substitute for judgment about hidden opportunities, under specified payoff and information assumptions. It does not make every form of judgment more valuable. In the support workflow, predicting a likely resolution leaves the cost of a mistaken commitment to be assessed. The prediction–decision distinction identifies that question; it supplies no proof of scarce expertise.

TWO CORNERS

Ownership is only the start of a scarcity claim

A common assistant feeds two support workflows whose resolution histories have different access conditions.The same assistant connects to both firms. One history is openly reproducible; the other has restricted access and a question mark about scarcity. Neither workflow receives a profit figure or an advantage verdict.AssistantReproducibleRestricted?
Hypothetical comparison. Restricted access is not evidence of usefulness, lasting scarcity or captured value.

Suppose rivals acquire the same support assistant. One firm uses a resolution history that competitors can reproduce; another claims exclusive access to a useful history. The second is a candidate advantage only if the input improves results, stays hard to replace and lets the firm retain a benefit after costs. These are hypothetical cases. Owning the history proves none of those conditions.

Rare and consequential is not the same as defensible

Kahneman and Klein (2009) identify two conditions for skilled intuition: an environment with learnable regularities and an adequate opportunity to learn them through practice and feedback (pp. 520, 524–525). Confidence is an unreliable test. Before treating an escalation team's judgment as an advantage, examine the decisions it has learned from and the outcomes it actually sees.

RARE FOR YOU

A retrieved precedent still needs a comparison

A traceable precedent approaches a support escalation but stops at an unresolved comparison boundary.The cited case moves toward the current ticket. It remains on the source side of a comparison boundary marked with a question. No reference class, event rate or scarce human contribution is certified.PrecedentEscalationComparable?

An unfamiliar support escalation may have precedents elsewhere. Traceable cases can supply candidates for comparison; model familiarity establishes no frequency. Check inclusion, outcomes, observation periods and relevant differences. A usable reference class may be missing. Rare for you, common for the corpus keeps that boundary explicit. Private context and responsibility for the decision establish neither expertise nor scarcity.

TWO MAPS

Gross value and retained value are different maps

A support productivity record stays intact while costs, competition and the horizon enter an unresolved value estimate.A ticket represents the measured operating result. Separate inputs enter a retained-value worksheet, whose result remains unknown. No productivity percentage is converted into profit and no deployment is ranked or certified.MeasuredRetained?CostsCompetitionHorizon
Conceptual workflow, not an empirical heatmap. The field estimate stays fixed; retained value has not been measured here.

Keep the support productivity result fixed. To estimate retained value, add operating costs and assumptions about competitors, customer terms and the horizon. These inputs require evidence; they are not percentages to subtract from throughput. The maps need not invert. Differentiated value develops the accounting; selection must carry its uncertainty into the funding decision.

POINT IT THERE

Make the proposal answer four questions

A support proposal reaches four open questions: metric, diffusion, complement and capture.The ticket joins a selection memo with four unanswered fields. The memo remains pending review; it does not route a discovered candidate directly into approved funding.Metric?Diffusion?Complement?Capture?Pending
  • Metric: what moves, for whom, and over what horizon?
  • Diffusion: how soon can rivals match the capability?
  • Complement: what do we control, and what evidence shows it remains scarce?
  • Capture: who receives the benefit after competition, verification, and upkeep?

Mark each answer as measured, assumed or unknown. A sound operating case can justify a shared capability; an exclusive advantage needs further evidence. Test the cost of verifying the work before scaling.

Read the transcript

01 · THE MEASURED WIN

An assistant helps a support team resolve tickets. Start with the operating record: issues resolved per hour. That is a useful measurement, but it leaves another field open: what value does the firm keep? Tool fees, integration, checking and upkeep have to be paid. Customers and competitors may respond. The operating result and the retained benefit need different evidence. The support team needs both to judge its funding request.

02 · THE FIELD RESULT

Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied five thousand one hundred and seventy-two support agents at one firm. Their final twenty twenty-five article estimates roughly fifteen percent more issues resolved per hour with access to an AI assistant. Agents could edit or ignore its suggestions. Access arrived in stages, without random assignment; the researchers used difference-in-differences to estimate the effect. The result does not measure rival adoption, changes in prices or a durable competitive advantage.

03 · WHERE THE SAVING GOES

If rivals buy the same assistant, the gain need not disappear. A firm might retain a profitable saving while competitors do too. Faster support might also become a lower price or the service customers expect. None of these responses follows automatically from the productivity estimate. Compare benefits with quality and the full operating cost before deployment. Then ask who receives the benefit over the horizon that matters. A shared gain can be useful without being an exclusive edge.

04 · THE SEAM

Prediction and judgment do not have one fixed economic relationship. A May twenty eighteen working paper by Agrawal, Gans and Goldfarb models better prediction as increasing the value of judgment about hidden costs while reducing its value for hidden opportunities, under its assumptions. Return to the support case. A predicted resolution approaches the choice between a reply and escalation. A cost record joins before that choice: what would a mistaken commitment cost? That is a question to investigate, not proof that human judgment is scarce or that its value must rise.

05 · TWO CORNERS

Suppose the same assistant reaches two competing support teams. One uses a resolution history rivals can reproduce. The other claims exclusive access to a useful history. Restricted access could matter, but owning a record does not establish an advantage. Does the history improve the result? Can rivals replace it? Who keeps the benefit after costs? Both operating cases may be worthwhile. The second remains a candidate for an exclusive edge until those additional claims have evidence.

06 · LEARNING CONDITIONS

Rare and consequential is not the same as defensible. Daniel Kahneman and Gary Klein identify conditions for skilled intuition: an environment with learnable regularities, and an adequate opportunity to learn them through practice and feedback. Confidence is an unreliable test. For the escalation team, inspect which decisions generated feedback and which outcomes it could observe. An experienced person can have skill in one part of the job without having it in an unfamiliar case. Responsibility for that case does not settle the question.

07 · Advertisement · ElevenLabs Studio

The last mile of content is often sound: the right voice, the clean take, the version that works in another market. ElevenLabs brings AI voice, music, dubbing, and sound tools into one production flow, so teams can shape audio with the same care they bring to the story. Create the version people can actually hear.

08 · RARE FOR YOU

An unusual escalation may have precedents beyond this team. Bring a traceable case toward the current ticket. Stop at the comparison: do the inclusion rules, outcomes, observation periods and relevant circumstances fit? Retrieved cases are candidates. Model familiarity supplies no event frequency, and a usable reference class may be missing. Private context may help, but neither owning it nor carrying responsibility proves expertise or scarcity. The comparison and the decision remain open.

09 · TWO MAPS

Keep the measured support result in place. Bring operating costs into the value estimate, then add competitive response and the horizon. These inputs need evidence; they are not discount percentages to subtract from issues resolved per hour. The retained-value field stays unresolved until the benefits, costs and assumptions can be defended. The maps need not invert. They answer different questions. A saving may be worth taking while the claim to an exclusive advantage fails.

10 · POINT IT THERE

The support proposal now carries four questions. Metric: what moves, for whom, and over what horizon? Diffusion: how soon can rivals match the capability? Complement: what do we control, and what evidence shows it remains scarce? Capture: who receives the benefit after competition, verification, and upkeep? Mark which answers are measured, which are assumptions, and which remain unknown. A sound operating case can justify a shared capability. A claim to an exclusive edge needs more evidence, and naming a candidate does not establish its value. Price the work of verification before scaling.

01 / 10 · THE MEASURED WIN0:00 / 5:48