Most decisions get made by whoever argues best in the meeting. A decision matrix makes the reasons visible.
What it is
Two common forms:
Pugh matrix. Developed by Stuart Pugh, a British design engineer, in the 1980s. You pick one option as the datum (the baseline, often the current way). Every other option is scored against it on each criterion: better (+), same (S), or worse (−). Count the pluses and minuses. It's quick, and it's good early on, when you have rough concepts.
Weighted decision matrix. Each criterion gets a weight. Each option gets a score on each criterion, say 1 to 5. Multiply and add. Highest total wins, on paper.
The value isn't the total. It's that everyone can see why.
Why people skipped it
Someone has to gather the facts for every cell. Six options, eight criteria, forty-eight cells, each needing research. So the matrix gets filled in by gut feel in the meeting, and the numbers mean little.
How a bot does it
- People choose the criteria and weights. This is the part that reflects what you value. Do it before the bot sees the options, so the weights aren't bent to favor someone's favorite.
- Give the bot the options and the evidence. Quotes, spec sheets, trial results.
- Have it score every cell with a reason and a source.
- Have it run the math, and a sensitivity check: does the winner change if a weight moves a little?
- People review the close calls and decide.
The bot brief
Score these options in a weighted decision matrix.
Options: [list, with source files]. Criteria and weights: [given — do not change them].
Score each option on each criterion 1-5. For every score, give a one-line reason
and cite the source. If no evidence, score it "?" and say what's missing; don't guess.
Compute weighted totals.
Sensitivity: for each criterion, would the top option change if its weight moved
by ±1? Report which weights the decision depends on.
Do not recommend a winner beyond reporting the totals and sensitivity.Starting point only. Test it on your own data and model, and adjust as needed before you rely on it.
An example
Choosing a model for a quoting bot. Weights set by the owner first.
| Criterion (weight) | Model A | Model B | Model C |
|---|---|---|---|
| Accuracy on 30 past quotes (5) | 5 | 4 | 3 |
| Cost per quote (3) | 2 | 4 | 5 |
| Speed (2) | 3 | 4 | 5 |
| Handles long price lists (2) | 5 | 5 | 3 |
| Weighted total | 47 | 50 | 46 |
Sensitivity: if the cost weight drops from 3 to 2, B leads by only one point (46 to 45). At 1, A wins (43 to 42). So the decision really depends on how much cost matters versus accuracy. That's a conversation for people, and now it's a short one.
Where bots differ
A bot can make a matrix look objective when the weights are the whole decision. That's why people set the weights before the scoring, and why the bot reports which weights the answer depends on.
Watch for scores without sources. A "4" with no reason is a guess wearing a number.
What this means for you
- Set criteria and weights first, as people.
- Let the bot gather evidence and score every cell.
- Always ask for the sensitivity check.
- Spend the meeting on the close calls.