Decision PlaybookResearch Prioritization

When to Stop Researching: The Expected Value of Information

More information is valuable only when it can change the choice. Expected value of information turns research from a comfort purchase into a decision about stakes, sensitivity, cost, and delay.

Decision summaryBefore commissioning more analysis, write what action follows from each plausible finding. If every finding produces the same action, the research has no decision value. If actions differ, estimate the stakes and buy the cheapest test capable of crossing the decision threshold.

When this framework fits

Use it when

  • A choice is live and one or more uncertain inputs could change it.
  • Research competes with the cost of waiting.
  • Several evidence gaps need to be prioritized.

Do not use it alone when

  • The work is required for safety, law, audit, or governance.
  • Learning has durable value across many future decisions.
  • The decision-maker has not stated the alternatives or risk tolerance.

The core idea

Expected value of perfect information, or EVPI, is the difference between the expected outcome when the true state is known before deciding and the expected outcome when deciding with current information. It is the ceiling on what perfect research could be worth. Real research is imperfect, so its expected value is lower.

You rarely need a complex calculation. The discipline begins with an action table. List the plausible findings, the action you would take under each one, and the outcome range. If the action column is constant, another study may increase confidence but cannot improve this decision. If the action changes, information may be valuable, especially near a threshold where a modest update flips the preferred option.

A practical five-step process

1. State the decision and alternatives

"Understand the market" is not a decision. "Enter now, run a pilot, wait, or pass" is. Include the deadline, the decision owner, and what is already committed. Research becomes easier to price when the action is concrete.

2. Build the action table

For each uncertain issue, define credible high, middle, and low findings. Write the action under each. Combine findings only where interactions matter. An open question that never changes the action belongs in the appendix, not at the top of the research plan.

3. Estimate the value spread

Estimate the expected cost of choosing wrongly compared with choosing correctly. Include downside avoided, upside captured, reversibility, and risk tolerance. A small expected-value difference can still matter when one path threatens survival, but that is a utility and risk-capacity question that should be stated explicitly.

4. Subtract research cost and delay

Include fees, management time, data acquisition, and the cost of postponing action. In a fast market, one month of delay can exceed the invoice for the study. In a reversible decision, a small pilot may create information while also creating progress.

5. Buy the cheapest decisive test and stop

Rank tests by their chance of changing the action per dollar and week. Choose the smallest evidence package that can cross the threshold. Pre-commit to a date or evidence rule for deciding. Otherwise the team can continue buying precision after the useful uncertainty has already been resolved.

Worked hypothetical example, not a client case

A product variant with a $100,000 information ceiling

Launching costs $250,000. The team estimates a 60% chance the variant contributes $600,000 and a 40% chance it contributes nothing. Launching now has an expected value of $110,000: 60% times $600,000, less the $250,000 cost. Not launching has an expected value of zero.

With perfect knowledge, the company would launch only in the good state. That produces an expected value of $210,000: 60% times the $350,000 net contribution. EVPI is therefore $100,000. A $40,000 test that resolves most of the uncertainty may be attractive. A $150,000 study cannot be justified by this decision alone, even before pricing delay.

Evidence required

InputWhat must be explicitCommon weak substitute
AlternativesActions that can actually be takenA broad research question
Current beliefProbabilities or bounded scenariosOne point forecast
PayoffsOutcome by action and stateRevenue without cost or downside
Decision thresholdThe finding that changes the action"More confidence"
Research performanceCost, time, likely accuracy, and limitationsAssuming the study is conclusive

Failure modes

Confidence laundering: analysis is commissioned to defend a decision already made. Precision theater: a tight estimate is purchased for an input the outcome barely responds to. Symmetric pricing: false positives and false negatives are treated as equally costly when they are not. Ignoring delay: the study is cheap only because the opportunity cost is omitted. Open-ended inquiry: no decision date or stopping rule exists. Perfect-test fantasy: EVPI is used as if real research will reveal the truth without error.

Research stopping checklist

  • The decision, alternatives, owner, and deadline are written.
  • Each plausible finding maps to an explicit action.
  • The most sensitive and load-bearing uncertainties are ranked.
  • The value of avoiding a wrong action is bounded.
  • Research cost includes time, delay, and management attention.
  • The test has a stated chance of resolving the decision threshold.
  • A stop date or evidence rule is agreed in advance.

Limitations

Inputs are estimates, dependencies between uncertainties can matter, and people value risk differently. Information can also improve future decisions, coordination, or compliance even when it does not change the immediate action. EVPI should frame the ceiling, not provide false precision. When the downside is existential or irreversible, use scenarios, utility, and risk capacity alongside expected value.

Price the next question before buying the answer

We can identify which uncertainty can still change the choice and what it is worth resolving.

Discuss a decision