Decision Matrix

Make a comparison traceable by showing which criteria matter, how evidence maps to a common scale and whether reasonable changes alter the ranking.

In one minute

A Decision Matrix places comparable alternatives in rows and decision criteria in columns. A weighted version usually:

1. filters alternatives through mandatory requirements;
2. defines non-overlapping criteria and observable scales;
3. assigns criterion weights before performance scores are known;
4. rates each alternative with cited evidence and confidence;
5. calculates weighted totals;
6. tests sensitivity to uncertain scores, weights and option changes;
7. records an accountable decision, including risks and dissent.

The total is a model of stated preferences and evidence—not proof of an objectively best answer. A close or unstable ranking is a reason to improve evidence or make the trade-off explicit.

Best for: defined alternatives that answer the same decision question across several differentiating criteria.
Avoid when: alternatives are not comparable, one requirement already decides eligibility or uncertainty cannot be represented by point scores.

The problem it addresses

Decision discussions often mix criteria, facts and preferences. One participant emphasises cost, another risk, and a preferred option receives favourable interpretations after the fact.

A matrix exposes the structure. It can also create false confidence if teams double-count criteria, use undefined scales, score missing evidence at the midpoint or adjust weights until a favourite wins. Sensitivity analysis is therefore part of the method, not an optional decoration.

When to use it

Use a Decision Matrix when:

• two to eight viable alternatives answer one decision statement;
• the choice depends on several distinct outcomes;
• criteria can be operationally defined;
• trade-offs need to be visible to reviewers;
• procurement, architecture or product concepts need a documented comparison;
• qualitative and quantitative evidence must be combined carefully;
• the team can test uncertain weights and scores.

When not to use it

Do not use it:

• to discover the alternatives;
• to compensate for a failed mandatory requirement;
• when criteria are causally dependent and double-count the same benefit;
• to hide political judgement behind arithmetic;
• with incomparable units and no explicit normalisation;
• as the sole basis for a regulated or irreversible decision;
• when one scenario-specific downside needs separate risk analysis.

Use Pairwise Comparison for one-criterion ranking. Use Kepner-Tregoe Decision Analysis when Musts, Wants and adverse consequences require a stricter sequence.

Inputs required

Prepare:

• one decision statement, owner, horizon and review trigger;
• a comparable option set, including status quo where legitimate;
• mandatory pass/fail requirements;
• four to eight differentiating criteria;
• operational scales and direction of preference;
• weights agreed before alternative scoring;
• evidence, uncertainty and source notes for each rating;
• an external risk or consequence review.

Step-by-step process

1. Frame the decision

State what is being chosen, for whom, over which period and by whom. Keep each alternative at the same level.

2. Test mandatory requirements

Remove or explicitly condition alternatives that fail a genuine non-negotiable requirement. Do not give a Must a large weight and allow compensation.

3. Define independent criteria

Each criterion should represent a distinct decision-relevant outcome. Check for overlap such as “ease of use,” “adoption” and “user satisfaction” all capturing the same effect.

4. Build operational scales

Define what every score means before rating. Prefer actual units where comparable. If normalising different units, document the rule and its limitations.

5. Assign weights

Allocate importance to criteria before revealing alternative performance. Record who set the weights and why.

6. Rate evidence and confidence

For every cell, record the score, source, assumption and confidence. Keep “unknown” visible; do not automatically turn it into average performance.

7. Calculate and inspect

Multiply each score by its criterion weight and sum by alternative. Then inspect the cell-level reasons rather than reading only the final rank.

8. Test sensitivity

Vary pivotal weights and uncertain scores across credible ranges. Recalculate after excluding a redundant criterion or adding a legitimate option. Identify which assumptions can reverse the result.

9. Review consequences

Examine downside, reversibility, distribution of harm and implementation dependencies outside the compensatory total.

10. Decide and document

The accountable owner records the choice, rationale, sensitivity, dissent, conditions and review trigger. A matrix recommends; it does not own authority.

AI automation lens

AI can validate arithmetic, find missing cells, identify duplicate criteria, link ratings to source notes and run sensitivity scenarios.

It must not:

• invent scores or evidence;
• infer stakeholder weights from seniority or past behaviour;
• choose the final option;
• hide a failed mandatory requirement inside a total;
• normalise incompatible measures without explanation;
• optimise the matrix until a preferred answer wins.

For consequential choices, preserve the complete version history and human decision record.

Visual model

Text alternative: comparable alternatives first pass mandatory gates. Eligible options are evaluated against defined criteria and fixed weights using evidence and confidence. The weighted result is stress-tested; unstable results return for evidence or explicit trade-off, while stable results proceed to accountable decision and separate risk review.

Interactive example

Scenario

A logistics company compares three customer-service platforms. All pass data-location and audit-export requirements. Criteria are twelve-month cost, integration time, accessibility, case-routing quality and exit effort. The current favourite receives 9/10 for routing based only on a vendor demonstration.

Your move

Define an operational scale for routing quality, decide how to handle the demonstration score and name two sensitivity tests.

Worked answer

Routing quality is measured on a representative blinded test set: 9 means at least 95% correct team routing and at least 98% urgent-case recall; 7 means at least 90% and 95%. Until the test runs, the favourite has “unknown,” not 9.

Sensitivity tests vary routing performance across the credible range and double the weight of exit effort under a contract-risk scenario. If either change reverses the ranking, the recommendation is conditional on the test and negotiation evidence.

Facilitation notes

• Define criteria and scales before naming the favourite.
• Ask whether two criteria reward the same underlying outcome.
• Separate measurements from preference weights.
• Use ranges for uncertain cost, schedule and performance.
• Show cell evidence during review.
• Preserve dissent and the decision owner's authority.

Expected output

• a precise decision statement and comparable options;
• a mandatory-gate record;
• operational criteria and non-overlap check;
• weights with rationale and ownership;
• scores with source, assumption and confidence;
• a reproducible calculation;
• sensitivity and consequence analysis;
• an accountable decision and review trigger.

Common mistakes

1. Undefined scales. “8/10” is meaningless without an operational definition.
2. Double-counting. Correlated criteria make one preference dominate invisibly.
3. Unknown equals average. Missing evidence must remain visible.
4. Weight tuning. Changing weights after seeing the winner is retrospective justification.
5. False precision. Decimal totals do not make qualitative evidence precise.
6. Skipping sensitivity. A ranking that flips easily should not be presented as robust.

Quality checklist

• Alternatives answer the same decision question.
• Mandatory requirements are pass/fail.
• Criteria are differentiating and not duplicated.
• Every scale has an operational definition.
• Weights were fixed before performance scores.
• Each material rating has evidence and confidence.
• Unknowns are visible.
• Sensitivity, risks, dissent and review trigger are documented.

Template

Add:

• mandatory gate results;
• total and normalisation method;
• sensitivity ranges and rank changes;
• consequences outside the score;
• decision, dissent, conditions and review trigger.

Use the structured Decision Matrix workspace template for the complete record.

Knowledge check

Question: Option A leads by 0.7 points, but a small credible change in one uncertain criterion reverses A and B. What is the strongest conclusion?

A. A is objectively best.
B. Add decimal places.
C. The ranking is sensitive; improve the pivotal evidence or make the trade-off conditional.
D. Remove option B.

Answer: C. Sensitivity reveals that the apparent lead is not robust to credible uncertainty.

Related tools

• Kepner-Tregoe Decision Analysis separates Musts, Wants and consequences.
• Pairwise Comparison ranks a short list against one criterion.
• SCAMPER generates alternatives before evaluation.
• FMEA analyses failure modes outside the weighted total.
• PDCA/PDSA Cycle tests a pivotal assumption.

References

1. National Aeronautics and Space Administration. NASA Systems Engineering Handbook, NASA/SP-2016-6105 Rev 2, section 6.8. Official handbook (opens in a new tab). Accessed 26 August 2026.
2. National Aeronautics and Space Administration. “Systems Engineering Handbook Appendix: Trade Study.” Official definition and reporting requirements (opens in a new tab). Accessed 26 August 2026.
3. Frey, D. D. et al. “The Pugh Controlled Convergence method: model-based evaluation and implications for design theory.” Research in Engineering Design, 20, 2009. DOI (opens in a new tab). Independent analysis of a related matrix method.
4. Cinelli, M. et al. “How to support the application of multiple criteria decision analysis? Let us start with a comprehensive taxonomy.” Omega, 96, 2020. DOI (opens in a new tab). Independent review covering sensitivity, robustness and rank reversal.