Decision intelligence
Test the explanation before committing to the action.
Decision intelligence connects a defined choice to evidence, competing hypotheses, constraints and feedback. It is analysis designed around an action, not a report looking for an audience.
The actual problem
A metric moved. The organisation has several stories and no test.
Most consequential decisions are made with incomplete evidence. The discipline is not to remove uncertainty, but to expose the assumptions connecting a signal to a proposed action and test the alternatives that would lead to a different choice.
Without that structure, a dashboard variance becomes a debate between functions. Each team selects a plausible explanation, and the most senior or confident narrative wins.
- Explanation by instinctA performance change is attributed to price, product, market or execution before competing causes are tested.
- Aggregate-only analysisA total moves, but cohort, segment, channel and mix effects remain hidden.
- False precisionA forecast presents a single number without a range, sensitivity or explicit assumptions.
- Insight without actionThe analysis ends with observations rather than options, thresholds and owners.
- Decision without feedbackThe organisation acts but does not define how it will know whether the intervention worked.
- Repeated re-litigationThe same choice returns because the reasoning and evidence were never recorded.
Method
Design the analysis around the decision boundary.
A useful analysis changes what a decision-maker believes about the available options. That requires more than describing a trend: it requires a falsifiable question, a comparison and an explicit link between evidence and action.
- Decision statement
- Name the decision-maker, deadline, options, constraints and the evidence that could credibly change the choice.
- Hypothesis ledger
- List competing explanations, their observable implications and the data that would weaken or strengthen each one.
- Comparison design
- Select an appropriate baseline, cohort, segment, time window or scenario. Avoid comparing unlike populations or periods.
- Driver analysis
- Decompose the outcome into volume, mix, price, conversion, retention, capacity or other mechanisms relevant to the business model.
- Option assessment
- Compare credible actions on expected effect, cost, reversibility, implementation time, risk and information value.
- Feedback rule
- Define leading signals, review dates and the conditions for continuing, adapting or stopping the chosen action.
Evidence required
The minimum data depends on the decision, not the tool.
Decision intelligence can work with imperfect data if the limitations are visible. A smaller coherent dataset with stable definitions is often more useful than a larger extract whose population and transformations cannot be explained.
| Input | Role in the analysis | Failure to avoid |
|---|---|---|
| Outcome measure and denominator | Defines what improved or deteriorated and for which population. | Reporting a rate without the changing base behind it. |
| Time and event history | Separates trend, seasonality, lag and intervention timing. | Comparing partial periods with completed periods. |
| Cohort and segment dimensions | Reveals mix effects and heterogeneous behaviour. | Assuming an aggregate change occurred within every segment. |
| Exposure or treatment record | Shows who experienced a campaign, product change or operating intervention. | Attributing all subsequent movement to the intervention. |
| Cost, capacity and policy constraints | Keeps options feasible and comparable. | Recommending the largest theoretical effect without implementation context. |
| External market evidence | Tests whether the movement is company-specific or broader. | Using market data with incompatible definitions or populations. |
Outputs
A compact evidence package for a named choice.
The deliverable records the reasoning so that the decision can be challenged, implemented and reviewed without repeating the entire analysis.
- Decision frameDecision owner, deadline, options, constraints, criteria and the cost of waiting or reversing.
- Hypothesis and evidence matrixCompeting explanations with supporting, conflicting and missing evidence.
- Driver decompositionThe contribution of mix, volume, price, conversion, retention or other mechanisms to the observed result.
- Option comparisonExpected effect, cost, uncertainty, reversibility and operational dependency assessed on common criteria.
- Recommendation with confidenceA clear choice, the reasoning behind it and what remains uncertain.
- Monitoring and review ruleSignals, thresholds, owner and review date that determine whether the decision remains valid.
Worked example
An aggregate decline caused by mix, not segment performance.
A sales team sees lead conversion fall from approximately 10.0% to 8.6% and proposes retraining account executives. Before changing the sales process, the analysis tests whether conversion deteriorated within segments or whether the lead mix changed.
| Segment | Earlier lead mix | Current lead mix | Conversion in both periods |
|---|---|---|---|
| Enterprise | 30% | 20% | 20.0% |
| Small business | 70% | 80% | 5.7% |
Current conversion = (20% x 20.0%) + (80% x 5.7%) = 8.56%
Neither segment became less effective. The aggregate rate fell because the lower-converting segment represents a larger share of leads. Retraining the sales team may still have value, but the observed decline does not support that diagnosis.
The immediate decision shifts from changing sales execution to understanding why the acquisition mix changed and whether that mix is economically desirable after revenue, cost and capacity are considered.
Limits
What the analysis can and cannot establish.
Decision intelligence improves the structure and traceability of a choice. It does not turn observational business data into certainty.
- Association is not automatically causationA credible causal claim may require randomisation, a natural experiment or a stronger quasi-experimental design.
- Unobserved factors remain possibleThe analysis can only test explanations represented by available data or defensible external evidence.
- Models depend on assumptionsScenario outputs should be read with sensitivity ranges, not as exact forecasts.
- Judgement remains accountableData informs the choice; leaders remain responsible for values, risk appetite and implementation.