Does every decision require a complicated model?

No. A small transparent model may be sufficient when it captures the important alternatives and uncertainties.

Can expert judgment supply probabilities?

Yes, but its basis and uncertainty should be documented and tested for sensitivity.

Does expected utility describe how people always behave?

No. It is a normative decision framework under stated assumptions, not a claim that every observed human choice follows it.

AI Glossary / Graphs and symbolic AI / Intermediate

Decision analysis

Definition

Decision analysis is a structured approach to choosing among alternatives by making objectives, uncertainties, consequences and preferences explicit. It uses a model to clarify trade-offs, compare options and investigate which assumptions or additional information could change the choice.

What it means

A forecast describes what may happen; a decision requires choosing what to do. Even a very accurate forecast cannot determine the preferred action without knowing the available options and the consequences the decision maker values.

Decision analysis connects these elements. It can use decision trees, influence diagrams, probability assessments, utility functions and sensitivity analysis. The model is a way to examine a choice, not a machine that removes responsibility for its assumptions.

The quality of a decision also differs from the luck of its outcome. A well-supported choice can have an unfortunate result when uncertainty is real. Conversely, an unsupported choice can turn out well. Evaluation should consider the information and alternatives available when the decision was made.

How it works

Start by framing the decision: who chooses, what can be changed, when the choice is needed and what objectives matter. Include realistic alternatives rather than comparing only a favored proposal with an artificially weak baseline.

Describe uncertain events and the consequences of each action under those events. Probabilities express beliefs about events; utilities or other preference measures express how outcomes are valued. Keep those roles distinct so that optimism about an outcome is not confused with a strong preference for it.

Under an expected-utility model, multiply each outcome's utility by its probability and sum across mutually exclusive, exhaustive outcomes. Compare those expected utilities, subject to any required constraints.

Then vary uncertain inputs and preference assumptions. Sensitivity analysis identifies where the recommendation changes. Value-of-information analysis asks whether learning something before acting could improve the decision enough to justify its cost. A useful analysis can reveal that the current choice is robust, that a missing alternative matters, or that one unresolved assumption deserves more attention.

Three practical examples

An events team compares venues by considering weather uncertainty, participant experience and operational constraints.
A maintenance group compares immediate inspection with continued operation using a stated model of possible conditions and consequences.
An AI project team compares alternative workflows instead of assuming that the highest prediction accuracy automatically produces the best business outcome.

Think of it this way

Think of laying the choices, possible developments and priorities on a table before committing. The arrangement helps everyone see what drives the recommendation and where they disagree.

The analogy does not make judgment mechanical. A neatly organized table can still omit a crucial outcome or encode preferences that the responsible people do not actually hold.

Worked example

A weather-sensitive venue choice

Suppose an invented event has two options. An outdoor venue gives utility 100 in dry weather and 0 in wet weather. An indoor venue gives utility 60 in either case. Assume these numbers are valid utilities for this teaching model, all relevant consequences are included, and wet weather has probability 0.30.

Expected outdoor utility is 0.70×100+0.30×0=70. Expected indoor utility is 60, so the model prefers outdoors.

Let p be the probability of wet weather. The options tie when 100(1-p)=60, giving p=0.40. Above 40%, indoors is preferred; below 40%, outdoors is preferred. This threshold makes the disputed forecast actionable.

With perfect weather information available before the venue choice, the team would choose outdoors when dry and indoors when wet. Expected utility would be 0.70×100+0.30×60=88. The improvement over the best uninformed choice is 88-70=18 utility units. An imperfect forecast is not worth this amount automatically, and utility units cannot be treated as currency without an appropriate conversion model.

Illustrative case study

A fictional workshop organizer, Larch Events, initially argues about whether its preferred outdoor venue is “too risky.” A decision analysis separates the weather estimate from participants' preferences and identifies a probability threshold at which the preferred venue changes.

The discussion also reveals a missing alternative: an outdoor venue with a usable indoor backup. The team adds its capacity limits and transition arrangements to the model rather than choosing between the original two options prematurely.

After the event, the organizer reviews whether its assumptions and contingency planning were reasonable. It does not judge the decision solely by whether rain happened on that particular day.

Common misconceptions and mistakes

Decision analysis is just prediction. It also requires choices, consequences and preferences.
The largest expected monetary amount is always the correct criterion. Risk attitudes, constraints and nonmonetary consequences may require a different utility model.
A detailed spreadsheet makes subjective assumptions objective. The assumptions remain judgments that should be visible and examined.
More information is always worth obtaining. Information has value when it can improve an actionable choice enough to justify its costs and timing.

Business use

Decision analysis is useful when a choice combines uncertainty with meaningful trade-offs. It can support clearer discussion among technical teams, managers and other stakeholders by showing why a recommendation follows from specified assumptions.

For an AI-supported decision, separate the predictive model from the action policy and preference model. Test whether forecast errors near the decision threshold matter more than errors elsewhere. Record the alternatives considered and the assumptions that would trigger a review. An analysis should make the choice easier to inspect, not bury it under unnecessary numerical detail.

Check your understanding

In the venue example, a revised wet-weather probability is 0.50. Which option has higher expected utility under the unchanged utilities?

Answer: Indoors, with 60 compared with 100×0.50=50 outdoors. The choice changes because the probability crosses the 0.40 threshold.

Related terms

Learn next

Add an indoor-backup alternative to the venue example and specify its consequences before calculating. Then vary the wet-weather probability and identify the ranges in which each option is preferred. State which assumptions would need real evidence before using the model operationally.

Sources

Russell and Norvig, AIMA materials on decisions, utility and information value: https://aima.cs.berkeley.edu/4th-ed/slides-pdf/chapter16.pdf

Lawrence D. Phillips, Decision Analysis for Practitioners, discusses decision structure, judgments and preferences: https://pubsonline.informs.org/doi/full/10.1287/deca.2025.0356

The venue utilities, threshold calculation and Larch Events case are original teaching examples.

Previous: Decile Next: Decision boundary