Multi-criteria decision matrix

Use the published, signed weights for this event. Rank the shortlist on those weights, attach the evidence for every cell, and stop. The committee still awards. A model that invents weights, or silently reuses last year's category template, is ranking its own taste.

That distinction is the quality bar. Challengers, internal audit, and the board will ask which criteria were published, who signed them, and whether anyone changed them after bids were opened. If the answer is that the model "optimized" the weights to fit the scores, you do not have a defensible award. You have a preference with arithmetic on top.

The matrix does not replace the work that produces each column. Quality and delivery still come from proposal scoring against rfp criteria. Cost should be total cost of ownership modeling, not the cover-page price. Concentration and plant risk come from single-source risk flagging. ESG belongs only where this event actually scores it, fed from esg supplier screening. The matrix applies the signed weights to those owned inputs and shows how close the result is, not only who sits first.

Sourcing and evaluation tools such as SAP Ariba, Coupa, Jaggaer, and Inventive AI sit in this class. This page does not rank those products, and it does not treat a software rank as the award.

Agree the criteria, the weights, the scoring scale, and the mandatory pass/fail gates in writing with the business stakeholder and the award committee before the RFP goes out. Put that scheme in the event documents the bidders see. Keep one version. A private second weighting, held "in case we need it," is how the preferred supplier gets found after the fact.

Reweighting after you have seen the scores is the failure that ruins the rest. A one-point shift from quality toward price that happens to put the favored bidder first is visible to anyone who can subtract. If the first ranking is uncomfortable, gather more evidence, treat a near-tie as a tie, or record a reasoned override. Do not edit the weights.

Keep the locked scheme small enough that a committee member can defend it. Five to seven scored dimensions plus a short list of mandatory gates is enough for most awards. Scoring forty suppliers across twenty criteria is not a decision. It is a table nobody in the room can stand behind line by line. Shortlist against published rules first. Run the matrix on that shortlist only.

Store the signed scheme as a versioned file on the event: who signed, on which date, which pack went to bidders. The model reads that file. It does not propose a "better" weighting because a different mix would separate the scores more cleanly. Separation is not the goal. Fidelity to the published scheme is.

Build each supplier row from four kinds of input, then apply the signed weights. Do not hide the parts inside one unexplained index.

• Scored criteria. Claims extracted against the RFP, with the source passage attached. Gaps stay visible. A polished proposal is still a claim until someone tests it.
• TCO. Bid price plus logistics, transition, expected quality failure, and exit, using this event's TCO model. If a cost line has no source, omit it. Do not invent it into the total so the ranking looks complete.
• Risk. Concentration, single-plant, and region exposure. A cheap bid that puts the category on one plant is not a bargain until the committee accepts that concentration.
• ESG. Only the dimensions this event scores. Pulling a full screening file into the matrix and then ignoring it is worse than leaving ESG out.

Mandatory gates sit outside the weighted average. Insurance, required certifications, living-wage or labor clauses, data residency, and any published legal pass/fail are binary. A supplier that fails a mandatory item is out. Do not let a high TCO or quality score pull a fail back into the ranking. Averaging a fail is how a non-compliant bid looks close enough in a dashboard and then becomes an award the legal team cannot defend.

Show the fail on the face of the pack, with the evidence, and drop the row. A still-ranked supplier with a fail in a footnote is not a ranking. It is a buried exception.

The walkthrough is illustrative, not a measured result.

A campus facilities award publishes its scheme before bids open: TCO 40, quality and mobilization 30, operational risk 15, ESG 10, strategic fit 5. Mandatory gates include specified insurance and a named living-wage clause. Three suppliers reach the shortlist the event defined.

One fails the insurance gate. The matrix removes that row. It does not keep the supplier in the table with a muted score. The remaining two are close on the weighted total. One is cheaper on TCO and thinner on mobilization evidence. The other is stronger on quality claims and carries more single-site concentration. A small, unpublished nudge of the TCO weight would flip the order. That sensitivity is something to print, not a reason to open the weight file.

What the committee needs on the page is not "supplier B wins." It is the locked weights, the two remaining rows with evidence links, the mandatory fail already removed, and a plain statement that the pair is a near-tie. The committee then decides on grounds the matrix already showed, such as mobilization risk or concentration, and records that decision in the award minute. If they want a commercial move from the cheaper bidder first, negotiation position generation is the next artifact, built from the same TCO and alternatives, not from a second secret weighting.

Do not widen the matrix to forty names so a favored supplier can sit in a respectable band. If they did not make the shortlist the event published, they are not in the ranking. A long ranked list is how a committee avoids saying who it will award.

Treat the output as a recommendation with a rationale, stamped with the weight file it used. The award minute still names the committee, the scores they accepted or overrode, and why.

Pasting the model rank into the award letter as if it were the decision is the failure that follows a clean matrix. Challengers will ask who judged, against which criteria, and whether anyone looked at the near-tie. If the answer is that the system ranked them, you do not have an award. You have a tool output.

Record overrides in the same pack: which criterion, which evidence, who signed. An override that cannot be explained is a reweight by another name. An override that is explained is how a committee uses judgment without pretending the arithmetic was the judge.

When the ranking holds across a small sensitivity band you agreed before bids opened, say so. When it does not, say that too. Do not manufacture a winner by hiding the band, by averaging away a mandatory fail, or by ranking a field so large that first place stops meaning a decision.