Value of Information

The team had data but no decision rule

A product team plans a nationwide launch on Monday. The main uncertainty is whether an important flow works reliably on low-end phones and weak connections.

The decision is real:

• If performance is acceptable, launch broadly.
• If it's poor, start with a small rollout and fix the problem first.

A focused test on representative devices can answer that question by Friday.

Instead, requests accumulate for more retention cuts, market reports, competitor comparisons and another executive dashboard. That work may be useful elsewhere, but nobody can say which result would change Monday's choice. An engineer spends the weekend assembling a more complete picture while the decision remains exactly where it began.

The team had mistaken having more information for making a better decision.

Value of Information asks whether learning something is likely to improve a decision enough to justify the time, delay and risk of finding out.

For this launch decision, information is useful only if one possible result leads to a different rollout.

Start with the decision

Suppose you need to choose now. With what you know today, one option looks better than the others.

New information might show that a different action is better. Before paying for it, consider the answers you might get, how likely they are, and how each would change your choice.

Put simply:

Value of information = the expected result after learning and choosing again − the expected result from choosing now − the cost of learning

The word expected matters because you pay for the information before you know which answer it will give. You are weighing the possible results, rather than looking back at a result you already know.

The cost includes more than money:

• Time spent collecting and interpreting data
• Delay while an opportunity closes or harm continues
• Privacy, safety or legal exposure created by collection
• The chance of acting on a misleading test
• Attention taken from the work itself

If every possible result leaves you doing the same thing, that information is not worth delaying this decision for.

The information may still matter for another decision, but it doesn't justify delaying this one.

Write down what each result would change

Before commissioning analysis, complete two sentences:

If the result is A, we will ___.

If the result is B, we will ___.

If both blanks contain the same action, ask why the information is needed now.

For the launch team, the focused device test passes this test:

• Good performance leads to broad rollout.
• Poor performance leads to staged rollout and targeted repair.

Many of the extra reports don't. Whether competitor sentiment is slightly higher or lower, the team still needs to resolve the same reliability risk.

The focused test is worth doing because a good result leads to a broad launch, a bad result leads to a smaller launch, and the team still has time to choose. Information that arrives after a decision cannot be undone can be accurate and still too late to help.

Keep the decision reversible long enough to use the result

The test matters because the team can still stage, delay or reverse the rollout when the result arrives.

A staged rollout, one delayed market or a reversible configuration gives the team a response to new evidence. Once every choice is locked in, the same evidence can only explain the loss afterwards.

That gives the team four steps:

1. Preserve the decisions that may need to change.
2. Identify the uncertainty that separates those decisions.
3. Buy the smallest reliable piece of information that resolves it.
4. Commit only as far as the evidence supports.

Sometimes the team should first make the rollout easier to pause or reverse. Better evidence has little practical value when there's no remaining action it can change.

How this works across fields

Economics and mathematics. This way of thinking compares the best option now with the best option after you learn something new. Perfect knowledge gives a ceiling: a flawed study should never be worth more than knowing the relevant fact for certain. That stops a company spending more on research than even a flawless answer could return.

Medicine. A diagnostic test is valuable when its possible results change treatment, follow-up or the need for another test. Testing can add little when both results lead to the same care. It can make things worse when false reassurance, an invasive sample or treatment delay causes more harm than the answer prevents. A patient’s right to know can still matter even when the medical plan does not change.

Software, product and strategy. When an operational decision has to happen now, logs, experiments, prototypes and small rollouts help when their results point to different next moves: undo the release, reduce traffic, switch to a backup or leave things alone. A dashboard that cannot help choose between those moves may show more without helping anyone act. Feature switches and limited releases make information more valuable because the team can still respond when it arrives.

When more information is worth buying

Use this model when deciding whether to:

• Run an experiment or prototype
• Add tracking that shows what happened
• Delay a commitment for more research
• Pay for a formal review before an investment or deal
• Order a diagnostic or confirmatory test
• Interview more customers or candidates
• Investigate an incident before acting

Ask:

1. What decision must be made?
2. What would we do with current information?
3. Which possible situation could make another action better?
4. What observation would distinguish those states?
5. How reliable will that observation be?
6. Can we still act differently when it arrives?
7. Is the expected improvement larger than the cost and delay?

Run the smallest reliable test that can separate the actions you are choosing between.

What this model can miss

The result-to-action test can undervalue exploratory learning. Basic research, customer conversations and broad observation sometimes reveal options nobody knew to list beforehand. Information may also be valuable for accountability, consent, a record people can use later or future decisions even when it doesn't alter today's action.

It is easy to make this look more precise than it is. The probabilities may be guesses, a test may be biased and the information can arrive after the useful moment has passed. People can also change their behaviour once they know what you are measuring.

There's also a moral limit. “The result will not change our decision” may reveal that the decision-maker has ignored another person's rights or consequences, not that the information is worthless.

Use the model to stop analysis that can't affect the decision. Don't use it to dismiss research, consent, accountability or dissent just because their value isn't immediate profit.

Nearby models

Bayesian Updating asks how evidence should change the relative weight of explanations. Value of Information asks whether obtaining that evidence is worth doing given the decisions it could change. An update can be correct while the evidence was too costly or arrived too late to be valuable.

Real Options preserve the ability to delay, expand, abandon or reverse a commitment. Value of Information measures the benefit of learning before exercising that ability. Real Options keeps several actions available; Value of Information asks whether learning more will help you choose among them before the option expires.

Before asking for another dashboard, write the decision and the two results that would make you choose differently. If no plausible result changes the action, decide whether the blocker is missing authority, unwillingness to commit or the absence of a stopping rule.