12 August 2026 ยท 7 minute read
Why a single number is not an answer
The most dangerous output any model produces is one number with two decimal places on it.
Somebody builds the model. Somebody presents the model. The slide says the price rise is worth 1.4 million dollars. Everyone in the room now has a number to argue with, which feels like progress, and for the next forty minutes the discussion is about whether it is 1.4 or 1.1 or 1.8.
Nobody asks the only question that matters, which is what would have to be true for that number to be the wrong sign.
This is not a failure of the people in the room. It is a failure of the artefact. A single number offers exactly one handle to grab, so that is the handle everyone grabs. Give the same room a range and they argue about the width of the range, which is a better argument, because the width is where the information is.
Where the number came from
Trace almost any decisive looking figure back through the spreadsheet that produced it and you find the same structure. Three or four cells hold real data. Another twenty hold arithmetic. And somewhere in the middle sit two or three cells holding a number that somebody chose, in a meeting, eighteen months ago, based on a feeling.
The volume response to a price change is almost always one of those cells. So is the ramp on a new salesperson. So is the share of a cost increase you can pass on. These are not observations. They are positions, and they were taken by people who are no longer in the building.
What the spreadsheet then does is remarkable: it takes three estimates with enormous ranges, multiplies them together, and reports the product to four significant figures. The precision is not a lie exactly. It is arithmetic performed faithfully on numbers that do not deserve it.
The same question, six hundred and forty times
The useful move is not to make the middle number more accurate. It is to look at the shape of the whole set.
Two answers with the same median can be completely different decisions. One where every run lands within ten percent of the middle is a decision you can make on the median. One where the middle is the same but a fifth of the runs end with the company short of cash in month eight is a decision about survival that happens to have a pleasant average.
A spreadsheet cannot tell those two apart, because it only ever computed the one path. It is not that the spreadsheet got the range wrong. It never had a range to get wrong.
What to ask instead
Four questions that make a number earn its place
None of these requires software. They are just the questions that a single number is designed, structurally, to prevent you from asking.
- What is the bad case, and how bad is it, and how often does it happen
- Which input is this standing on, and how do I know that input is right
- When would I find out I was wrong, and what would the first sign be
- If the answer changed sign, which assumption would have had to move, and by how much
The objection, which is fair
Boards want a number. Investors want a number. You cannot stand up and say the answer is somewhere between minus 400,000 and plus 2.2 million, and expect anyone to thank you for the honesty.
That objection is correct and it is not an argument for the single number. It is an argument for leading with the median and never presenting it alone. The sentence that works is: the middle case is 1.4 million, one run in ten is below zero, and the whole thing turns on whether our largest two accounts renew before the change reaches them. That is one sentence. It is still decisive. It just happens to also be true.
What it buys you is the meeting you have three months later. When the number turns out to be 600,000, the person who presented a range and a mechanism has a conversation about which assumption moved. The person who presented 1.4 has a conversation about their credibility.
A number is a summary of an argument, not a substitute for one
The reason to run something hundreds of times rather than once is not statistical sophistication. It is that running it once forces you to pick values for the things you do not know, and picking values for the things you do not know is exactly the step that makes the output feel more certain than the input was.
Run it many times with the uncertainty left in and the output carries the uncertainty forward. The middle is still there. You can still quote it. But now it is sitting next to the information about how much it is worth, instead of standing alone pretending that information does not exist.
The precision was never the problem. The problem is that a number with two decimal places on it does not look like an argument, so nobody argues with it, and the assumptions underneath it survive by being invisible.
Run the question instead of arguing about it
Build a twin from whatever record you have, and see the range rather than the one number somebody liked.