Skip to content
finalstate.aiPrediction vs Decision: The Machine Forecasts, You Judge
117Atlas

117THE UMBRELLA

Prediction vs Decision: The Machine Forecasts, You Judge

A forecast estimates what may happen. Acting on it also requires an objective, an error tradeoff and authority. Systems can execute a configured rule; the forecast alone does not supply those conditions.

A fixed 70% rain forecast points toward an umbrella still standing beside a doorway.The forecast reaches a decision point marked by a question. The umbrella remains in place throughout. This is an illustrative probability, not a measured weather forecast.70%Rain?
Illustrative 70% rain forecast, held fixed. The umbrella stays by the door; no choice is made.

CLAIM AND COMMITMENT

Evidence and authority answer different questions.

A sourced invoice flag and an acting rule approach an unsigned dispute record.The source stays attached to the forecast. Adding a rule supplies another decision input; it does not sign the pending dispute or establish that the mismatch is true.Invoice?SourcePendingRule
Illustrative invoice review. A rule reaches the decision record, but the dispute remains pending and unsigned.

An invoice may appear inconsistent with its contract. Disputing it requires a rule and authority to act. Decision premises include factual claims as well as values: the alleged mismatch is contestable evidence; the relative cost of overlooking an error or raising a false dispute is a value judgment. The forecast also needs a source and someone responsible for its use.

PREDICTION AND JUDGMENT

Better prediction can change which judgment pays.

An opportunity can change a safe action to a risky one; a hidden cost can reverse a risky action after a favorable forecast.Two schematic routes illustrate different information conditions in the authors' model. They do not compare observed values or imply that all judgment complements prediction.Little informationSafeRiskyOpportunityFavorable forecastSafeRiskyCost
Agrawal, Gans and Goldfarb, Exploring the Impact of Artificial Intelligence: Prediction versus Judgment, NBER WP 24626 (May 2018), §3, printed pp. 4–6, Proposition 1. Schematic action changes; no economic time series or measured payoffs.

In Agrawal, Gans and Goldfarb's May 2018 NBER model, better prediction complements judgment about hidden costs but substitutes for judgment about hidden opportunities. The result depends on the model's payoff assumptions. It supplies a conditional explanation for value moving between inputs, not a guarantee that human judgment becomes more valuable.

A probability does not specify what the outcomes are worth.

Forecast construction involves choices about data, targets and model assumptions. Kahneman and Tversky's 1982 publisher excerpt makes the role of judgment in mathematical forecasts explicit. Those choices remain open to examination. A probability alone cannot assign the objective, error tradeoff or authority for acting on it.

INSIDE AND OUTSIDE

A project plan can miss what happened to comparable projects.

A current project plan faces past project records with the comparison still questioned.Consulting past records adds a possible outside view. The question between the plan and records remains unresolved: unfamiliarity to this team does not guarantee comparable cases elsewhere.PlanPast cases?
Kahneman, Thinking, Fast and Slow (2011), McKinsey book excerpt, opening and “A shocking disconnect.” Inside/outside definitions also appear in the excerpt’s final section. Records below are conceptual, not a sample or histogram.

In Kahneman's 2011 book excerpt, a curriculum team estimated about two more years of work. A colleague then recalled comparable teams: some never finished; those that did needed seven to ten more years from a similar stage. This was a recollected comparison, not a reported cohort dataset. An outside view seeks relevant outcomes beyond the current plan. A corpus may contain precedents; their existence and comparability still need checking.

BUILDING THE COMPARISON

Retrieved cases still need an inclusion decision.

Three candidate records receive separate include, exclude and unknown review statuses while the class remains pending.A is included after assumed source and other checks, including the same starting stage; B is excluded for a different starting stage; C remains unknown because provenance is missing. Individual review statuses do not produce a validated class or numerical base rate.ReviewAIncludeBExcludeCUnknown?Class pending
Flyvbjerg, From Nobel Prize to Project Management (2006), author manuscript, pp. 6–10 of the 32-page PDF, especially the three-step procedure on p. 7. A/B/C are illustrative excerpts from a review, not observed project data.

Flyvbjerg's 2006 method identifies a relevant class, establishes its outcome distribution and positions the current project within it. The class needs credible data and enough comparability to support the estimate. In the illustrative review, assume A passes the source and other inclusion checks. It also shares the chosen starting stage. B uses a different stage; C lacks a traceable source. Recording those differences supports inclusion, exclusion or an unresolved status. It does not yet certify a whole class or base rate. Knightian uncertainty remains relevant when a defensible comparison is unavailable.

EVIDENCE FOR APPRAISAL

Even a supported comparison leaves the commitment open.

An evidence dossier and an acting rule enter separate sides of an unsigned decision record.Even assuming that the evidence has passed review, combining it with the rule leaves the decision pending. No autograph denotes either approval or ownership.EvidenceRulePending
Green Book 2026, §1 “What is the Green Book?” and §6 “Optimism bias”; supplementary guidance, §§1.1–1.2 and 3.1–3.5. The drawing shows a hypothetical supported evidence dossier, separate rule and pending commitment; no adjustment percentage is asserted.

The UK Green Book 2026, §6, requires explicit optimism-bias adjustments at the outset of appraisal: increase estimated costs and durations and decrease estimated benefits. It directs practitioners to historical forecast errors and similar proposals. The supplementary guidance supplies generic cost/time adjustments when stronger evidence is absent. This guidance concerns government appraisal; it neither chooses policy objectives nor validates arbitrary retrieved cases. The value premise and commitment authority require their own record.

Naming the owner does not approve the action.

A system can retrieve records, calculate an estimate and execute a delegated rule. The organization remains responsible for adopting the objective and defining intervention rights. A signature identifying who owns the review must not be mistaken for approval of the proposed action.

WHERE THE VALUE WENT

Keep the evidence and the commitment traceable.

A sourced forecast dossier reaches a review owner while a separate commitment record stays unsigned.The final handoff transfers evidence for review. It preserves an unresolved commitment field and makes no claim that cheaper prediction automatically increases the value of human judgment.Review ownerSourcePending
The forecast source reaches the review owner. The action remains pending and unsigned; ownership is not approval.

Use prediction and retrieval where their performance justifies it. Preserve sources and unresolved comparisons, then record the objective, error tradeoff and authority for any action. Return to 004's measurable wins: a shift in the value of judgment has to be demonstrated in the work.

Read the transcript

01 · THE UMBRELLA

Suppose the forecast gives a seventy percent chance of rain. That number alone does not tell you whether to carry the umbrella. The action depends on how much you mind getting wet, weighed against the nuisance of carrying something you may not need. Two people can hear the same forecast and reasonably make opposite choices. Leave the umbrella by the door while that choice is open. A system could apply a configured rule, but someone still has to adopt the objective and authorize the action.

02 · CLAIM AND COMMITMENT

An invoice flag says the charge may differ from the contract. That is a factual claim to investigate. The forecast needs a traceable source and someone responsible for how it is used. Deciding to dispute the invoice adds a different question: how should the organization weigh a missed error against a false dispute? Bring that acting rule to the decision record. Software may apply the rule under delegated authority. Here the dispute stays pending. Supplying an input does not itself authorize a commitment.

03 · PREDICTION AND JUDGMENT

Agrawal, Gans and Goldfarb examined this distinction in their May 2018 paper, Exploring the Impact of Artificial Intelligence: Prediction versus Judgment. Their model separates learning the state of the world from discovering what an action would be worth. With little information, uncovering a hidden opportunity can move the decision from safe to risky. After a favorable prediction, discovering a hidden cost can move it back to safe. Under their assumptions, better prediction increases the return to finding hidden costs and reduces the return to finding hidden opportunities. When judgment serves both tasks equally, complementarity requires the downside to be sufficiently larger. These are conditional model results, not a measurement of today's workforce or a permanent boundary around human ability.

04 · CHOICES WITHIN THE FORECAST

A forecast is not free of judgment in its construction. People choose what to measure, which records to include and what the model should predict. Those choices can be challenged. The resulting probability still does not specify what each outcome is worth. An automated decision combines evidence with an acting rule. Responsibility includes deciding who adopts that rule, who can revise it and who may stop its execution.

05 · INSIDE AND OUTSIDE

Kahneman's curriculum team had worked for about a year when he asked how much longer a finished draft would take. Their estimates centered on two more years. Then a colleague recalled other teams at a similar stage. Some had never finished; among those that did, he recalled seven to ten more years. The same expert had initially joined the optimistic estimate. Consult the past cases alongside the current plan. The comparison might expose problems the plan leaves out, but the class must fit. A search that returns nothing does not establish that the case is unprecedented.

06 · BUILDING THE COMPARISON

Bent Flyvbjerg describes a sequence: identify a relevant class of projects, establish its outcome distribution, then compare the proposed project with that distribution. Retrieval can help find records, but it does not complete those steps. Review three illustrative candidates. Assume A passes its source and other inclusion checks. Its starting stage matches too, so include A. Exclude B because it measures from a different stage. Keep C unresolved because its source is missing. The overall class remains pending. A usable estimate also needs an outcome definition, coverage and a defensible calculation.

07 · Advertisement · Twain GTM Agents

A good outbound message should know why it is being sent. Twain GTM agents research accounts, qualify contacts, and turn live signals into relevant outreach. They work with your CRM, personas, and GTM stack, so the message is not just personalized. It is backed by a specific reason to start the conversation.

08 · EVIDENCE FOR APPRAISAL

The United Kingdom's Green Book makes historical comparison part of government appraisal. Its 2026 guidance requires explicit adjustments for optimism bias at the outset, informed by past forecast errors and similar proposals. Generic cost and time adjustments are available when stronger evidence is absent. Suppose the evidence dossier has passed the relevant review. Bring it to the decision record alongside the acting rule. That still leaves commitment pending. The evidence may improve an estimate; objectives, error costs and authority determine how the organization uses it.

09 · OWNERSHIP AND APPROVAL

Naming someone responsible for a review and approving the proposed action are different acts. A signature can record the first without authorizing the second. Software can carry out a delegated rule, but a forecast or a review assignment does not create the delegation. The organization must specify who can commit, revise the rule or stop the action.

10 · WHERE THE VALUE WENT

Keep the evidence traceable as it reaches the review owner. Record where the cases came from, why they belong and what remains uncertain. Then keep the commitment separate: the objective, the cost of error and the authority to act. If the reference class fails, preserve that gap instead of manufacturing odds. If prediction becomes cheaper, examine which decisions still benefit from further judgment. Here the decision record remains unsigned.

01 / 10 · THE UMBRELLA0:00 / 5:54