Suppose a repair shop can pay for a report before committing to a job. The report contains only one answer: whether an expensive replacement will be needed. If that answer changes a costly decision, a tiny message can be worth more than a warehouse of unrelated records.
This is an invented decision problem. Its numbers are chosen to make the accounting visible, not drawn from a business trial. It lets us ask what “information creates wealth” would have to mean in a particular situation.
Put a decision behind the claim
The shop has two choices: accept the job or decline it. Declining produces a payoff of zero in this simplified model. Accepting produces $100 if the inexpensive repair is sufficient and loses $300 if the replacement is required. Assume each outcome has probability one-half.
Without further information, accepting has an expected payoff of minus $100: half of $100 plus half of minus $300. Declining therefore has the better expected payoff, zero.
Now suppose a perfectly reliable report arrives before the decision. The shop accepts the inexpensive jobs and declines the others. Its expected payoff becomes $50: half the time it earns $100, and half the time it earns zero. The gross expected value of this perfect information is $50 relative to the original best choice.
If obtaining and using the report costs $20, the net expected improvement is $30. If it costs $60, the net improvement is negative. The report’s physical length has not changed. Its cost and its role in the decision have.
This illustrates the standard decision-analysis idea of comparing the best expected outcome with additional information against the best expected outcome without it. MIT’s Risk and Decision Analysis course gives the general formulation. Our repair-shop numbers are original and deliberately simplified. Course transcript on expected value of sample information.
Expected payoff is an average under the assumptions. It is not a promised result on the next job. The model omits capacity constraints, customer relationships and many other things a real shop would have to consider. Its purpose is to show the comparison, not substitute for those facts.
Reliability changes the answer
A real report can be wrong. Keep the same prior probabilities and payoffs, but suppose the report identifies either condition correctly 80% of the time. Assume those error rates are symmetric and known, and that the report arrives in time to be used.
A favorable report then gives an 80% probability of the inexpensive repair. Accepting after that report has an expected payoff of $20: 0.8 times $100 plus 0.2 times minus $300. An unfavorable report makes accepting unattractive, so the shop declines.
In this model, favorable reports occur half the time. The overall expected payoff is therefore $10 before the report’s cost. A $20 report now costs more than the improvement it provides.
At 70% symmetric accuracy, even a favorable report gives an expected acceptance payoff of minus $20. The shop declines regardless of the report. If it remains free to ignore the information, the gross decision value is zero. It could still pay a fee and become worse off after costs.
The same one-answer format has produced three different results: a $50 gross improvement with perfect reliability, $10 with 80% reliability and zero with 70% reliability. What matters is the combination of accuracy, prior probability, available actions and consequences.
More accurate prediction is useful here because it sometimes changes the chosen action. A prediction tool that raises an accuracy score but never changes an actionable decision may provide no benefit under this particular payoff model. A different job, different losses or a different prior can change that conclusion.
The cost belongs to the whole process
A report has to arrive before the choice becomes irreversible. Someone must understand it. The shop must have authority and capacity to act. If the report comes after the replacement has already been ordered, its excellent accuracy cannot improve the earlier decision.
This makes information value partly organizational. A technically capable system may fail because its output reaches the wrong person, conflicts with another record or cannot be checked in time. Adding records would not necessarily remove any of those obstacles.
Include acquisition, integration, review and correction costs. A system that saves five minutes of search but creates ten minutes of checking has not demonstrated a time saving. If it occasionally produces an expensive mistake, count that consequence alongside its routine successes.
Costs and benefits also have owners. A gain for one party can impose work or harm on another. A business case should identify that distribution rather than treating every transferred cost as new wealth. Privacy and permission constrain which information can legitimately be acquired or reused, even when a calculation suggests it would be profitable.
Propose a comparison that could fail
Consider a team evaluating a new information tool. The following is a proposed evaluation, not an experiment performed for this series.
First, define one narrow recurring decision and the current process. Preserve a set of representative cases that the tool’s developers cannot use to tune their system. Include awkward cases: conflicting records, missing inputs, stale facts and circumstances in which declining to answer is appropriate.
Choose success criteria before seeing the results. For example, the tool might need to reduce total handling time while keeping consequential errors below an independently reviewed limit. Count verification time, escalations and failed cases. A speed improvement alone would not satisfy that two-part requirement.
Compare the existing process, a simpler improvement such as a cleaned reference sheet, and the proposed tool on comparable cases. Keep access to evidence and task difficulty comparable. Where possible, randomize assignment and have reviewers assess outputs without knowing which process produced them. Protect the final cases from repeated tuning.
Bound the work with a fixed sample and budget. Stop if a predeclared serious-error condition is met. If the tool misses the threshold, report the miss instead of changing the definition of success. If it passes, confirm the result on new cases before extending the claim.
An offline comparison can justify a further trial. It cannot establish effects on actual customer behavior or production revenue. A live evaluation needs its own permission, safeguards and measurement plan. Differences between user groups or changes over time can also limit how far the result travels.
The useful finding would be scoped: this tool improved this process under these conditions. Recheck it when the data, workflow or model changes. That is enough to support a real decision without declaring a universal law of value.
Scientific excitement does not establish an investment return
A profitable application, a valuable company and an attractive investment are different propositions. Even successful technology can be bought at a price that leaves an investor with a poor outcome. A scientific paper contains neither the entire business nor the terms of the investment.
The joint SEC staff, NASAA and FINRA alert on AI investment fraud warns about purported AI trading systems and claims of high guaranteed returns with little or no risk. It also explains that AI-generated information can be inaccurate, outdated or fabricated. Investor.gov alert.
Quantum language does not resolve those evidentiary problems. A promoter still needs to explain the offering, the costs, the risks and the basis for its performance claims. A reference to entropy or emergent spacetime cannot do that work.
Return to the small message
The repair shop’s report could fit in a single binary answer. Its value came from an opportunity to choose differently before paying for a mistake. The numbers changed when reliability and cost changed. Nothing in the model required a relation between monetary value and the thermodynamic cost of storing that answer.
That is the useful connection across this series. Information theory specifies uncertainty. Physical theory studies how information is embodied and how quantum relationships can enter geometric descriptions. Engineering asks what a system can reliably do. Economic value appears when a capability improves a consequential choice enough to justify its whole cost.
The frontier can be exciting while those connections remain unfinished. To evaluate a claim, follow the evidence across each step. The place where the evidence stops is the place where the next question begins.
Loading comments…