Evidence-informed decision-making combines relevant data, research, domain knowledge, stakeholder input, and explicit values to choose among alternatives. Data do not make a decision objective by themselves: measurements reflect definitions, collection processes, missingness, incentives, and uncertainty.

Make the decision auditable

Name the owner, objective, alternatives, affected people, constraints, time horizon, evidence standard, uncertainty, and review date. Distinguish description of what happened, prediction of what may happen, and a causal estimate of what an intervention will change.

QuestionDocumentCommon failure
DecisionOwner, alternatives, objective, constraintsOptimizing a proxy
MeasurementDefinitions, lineage, coverage, missingnessTreating available data as representative
InferenceClaim type, design, assumptionsTurning correlation into an intervention
RiskPrivacy, security, fairness, safety, legal limitsIgnoring harms outside the metric
ActionThreshold, authority, guardrails, rollbackA dashboard becoming an automatic rule
LearningEvaluation, review date, unintended effectsDeclaring success from a selected metric

Use a six-step decision loop

1. Frame the decision

State the deadline, feasible alternatives, objective, constraints, affected groups, and costs of false positives and false negatives. Define the outcome and observation window before inspecting results.

2. Specify the evidence plan

Record hypotheses, provenance, permissions, inclusion criteria, primary and guardrail metrics, uncertainty method, and what result would change the decision. Separate exploration from confirmation.

3. Validate measurement

Test schema, lineage, sampling, missingness, duplicates, label quality, time alignment, and subgroup coverage. A live dashboard can still be delayed, incomplete, or wrong.

4. Match method to claim

Use descriptive analysis for patterns, predictive validation for forecasts, and randomized or defensible quasi-experimental designs for causal effects. A model that predicts churn does not identify why customers leave or prove outreach will retain them.

5. Decide with guardrails

Present estimates with uncertainty, assumptions, alternatives, cost, and distributional effects. Name who may override a recommendation and predefine stop, rollback, and escalation rules.

6. Evaluate and update

Measure intended and unintended outcomes over an appropriate period. Compare against a baseline or control, record deviations, and revisit the decision when conditions change.

Apply privacy purpose and limits

Before linking store, online, loyalty, and device-location data, establish a lawful purpose, appropriate notice and choice, minimization, retention limits, access controls, and security. Hashing an identifier does not necessarily make tracking anonymous.

An association—such as mug purchases being more common among coffee buyers—does not show that an offer will increase incremental sales. Randomize where feasible and measure returns, complaints, opt-outs, accessibility, and trust as well as revenue.

Communicate results precisely

“In the prespecified experiment, the treatment estimate was two percentage points higher than control. The interval estimate, exclusions, stopping rule, and guardrail metrics are reported below; the result applies to the tested population and period.”

Report absolute and relative effects, denominator, timeframe, uncertainty, and design. Avoid causal “lift” language when the design supports only association.

Keep a decision record

  • Decision, owner, date, and alternatives.
  • Evidence, quality, assumptions, and conflicts.
  • Expected value, costs, risks, and affected groups.
  • Action, authority, controls, and rollback.
  • Evaluation plan, review date, and actual outcome.

Reliable decisions begin with data-cleaning controls, require an AI governance framework for automated systems, and depend on leakage-safe feature engineering when models are involved.

Originally published August 7, 2025; technically reviewed and substantially updated September 4, 2026.