Decision intelligence is the structured design, support and review of choices. It connects objectives, options, evidence, uncertainty, trade-offs, authority, commitment and outcome monitoring.

No. Analysis produces bounded findings about observed evidence. Decision intelligence uses those findings alongside objectives, constraints, values and consequences to support a choice.

Yes. Scores depend on criteria, scales, weights and evidence. They summarize a declared model; they do not remove uncertainty, hidden dependencies or judgment.

Reversibility depends on whether resources, relationships, rights, time and future options can be restored at acceptable cost after commitment.

When the expected value of additional information is lower than its cost, delay and opportunity cost, or when the decision deadline makes further acquisition impractical.

It is the consequence of choosing incorrectly, acting too early, failing to act or delaying beyond the useful decision window. Different errors can have very different costs.

Urgency can justify a faster or precautionary response, but it should change the commitment size, safeguards and review cadence rather than erase uncertainty from the record.

Compare observed outcomes with the original objectives, assumptions and expectations while keeping implementation quality, environmental change and luck analytically separate.

Do not combine objectives until the conflict is visible. First identify whether each objective is mandatory, directional or aspirational; then specify who has authority to trade one against another. Safety, legality or solvency may operate as non-compensable constraints, while speed and cost can remain trade-off criteria. If weights are used, test whether reasonable changes reverse the selected option. A decision is not robust when its winner depends on one unexplained preference. The final record should state which objective was prioritized, what value was knowingly sacrificed and why that trade-off was acceptable in the declared context.

A constraint defines the feasible set: an option that violates it is excluded. A criterion compares options that remain feasible. Confusing the two allows attractive performance elsewhere to compensate for an unacceptable condition. For example, a supplier may have the lowest price but fail a mandatory security requirement; the price cannot repair infeasibility. Constraints should have explicit tests, owners and exception authority. Criteria need direction, scale and evidence. When a constraint can be waived, the waiver itself is a separate decision with its own authority, consequence and documentation rather than an invisible score adjustment.

Precise probabilities are not always necessary or honest. The analysis can use ordered scenarios, plausible ranges, threshold tests and dominance relationships. Ask which option remains acceptable across several defensible states, where the preferred choice changes and which unknown has the greatest decision value. Qualitative uncertainty must still be structured: “unknown” should identify what is missing and how it affects options. Scenario labels cannot masquerade as measured likelihoods. When probability cannot be supported, the record should preserve that limitation and favor options that control downside, retain reversibility or create informative feedback.

Yes, when every feasible option has an expected loss and the choice minimizes unavoidable harm, protects a critical constraint or preserves future options. Incident containment, legal response and failure recovery often have this form. The comparison must include non-action because doing nothing also produces consequences. A negative expected outcome does not remove the need for proportionality: the selected option should still be tested against lower-commitment alternatives, distribution of loss and the possibility that short-term harm prevents larger irreversible damage. The record should state that the decision manages loss rather than creates benefit.

Stakeholder disagreement is not a data-quality problem. Identify affected groups, decision rights, exposure and the basis on which competing values enter the choice. Do not average incompatible values into one anonymous score without governance. Some interests may create obligations; others may inform preferences. The process should expose distribution: who receives the benefit, who bears the downside and who can reverse the action. A technically efficient option can remain unacceptable when costs are imposed on parties without authority or remedy. The final decision should document the governing principle used to resolve the conflict.

An optimal option performs best under one specified model and input set. A robust option remains acceptable across several plausible models, estimates and future states. Robustness matters when evidence is unstable, downside is asymmetric or the environment can change before commitment pays off. Test alternative weights, thresholds, scenarios, time windows and implementation assumptions. The robust choice may not produce the highest modeled upside, but it avoids catastrophic failure and retains performance under uncertainty. Report both the best-case option and the stable option when they differ so the decision-maker can see the price of robustness.

A model contributes a conditional output based on its inputs, objective function, training or rule structure and available data. It does not own the objective, accept consequences or possess decision authority. Use the recommendation as evidence whose relevance depends on domain fit, calibration, error distribution and current conditions. Inspect cases where the model is least reliable, not only average accuracy. Human override should require a reason, but model agreement should also require justification when stakes are high. Preserve the recommendation, final choice and rationale separately so later review can distinguish model error from governance or execution failure.

Escalation is justified when the choice exceeds delegated mandate, crosses a consequence threshold, creates irreversible exposure, conflicts with policy or contains unresolved disagreement that the current owner cannot legitimately resolve. Escalation should not be a way to transfer discomfort or avoid responsibility. The package sent upward must state the decision question, viable options, evidence, critical unknowns, recommendation and exact issue requiring authority. A higher title does not improve evidence automatically. It changes who is permitted and accountable to accept the trade-off.

Separate terminal outcomes from leading indicators and assumption checks. A five-year strategy cannot wait five years for its first review, but short-term metrics must not silently replace the intended objective. Define milestones that test causal prerequisites: access, adoption, unit economics, capability, retention or regulatory stability. Each milestone should specify expected range, observation date and response. Review whether the underlying assumptions remain valid and whether implementation matches the chosen option. Continue, adapt or stop based on the evidence the decision said would matter, not on whichever metric currently looks favorable.

A stop condition is observable, decision-relevant and set before commitment distorts judgment. It can be a maximum loss, failed milestone, unresolved safety issue, evidence contradiction, time limit or disappearance of the original opportunity. It should name who monitors it and whether crossing the threshold triggers automatic termination, review or escalation. Avoid vague wording such as “if performance is poor.” A useful condition identifies the measure, boundary, duration and permitted response. Stop rules protect resources and learning, but they also need exceptions for confirmed measurement failure or a formally authorized change in objective.