The answer that arrived before the question
August 22, 2026
“Can you pull the numbers that show the Toronto expansion is working?” It sounds like a data request. It is a verdict with a research task attached.
A reasonable request, on a Wednesday
The request is not sinister and the person making it is not cynical. A director believes the Toronto expansion is working. They have been there, they have talked to the team, they have a strong and possibly correct instinct. What they need is something for the board pack, and they need it Friday.
So the analyst does what a helpful colleague does. There are twelve ways to measure whether an expansion is working, and three of them look good. Those three are defensible. Each is a real metric, honestly computed, correctly labelled. The chart goes in the deck. Nobody lied, nobody fabricated anything, and the board approves phase two.
This is the elephant we call Bias: cognitive blindspots, and gut feeling dressed up as data. Its distinguishing feature is that every individual step is defensible. It never requires anyone to be dishonest. Only to be helpful, in order, starting from the answer.
The evidence
This is not an accusation from outside. It is a description from inside.
78%
of business leaders say people often make the decision first — then go looking for the data to justify it
Oracle and Seth Stephens-Davidowitz, The Decision Dilemma, 2023 (n=14,250 across 17 countries).
If everyone knows it is happening, why doesn't it stop?
That figure is the interesting part of the study, and it is easy to skim past. It is not researchers catching executives out. It is executives, 14,250 of them across 17 countries, describing what they watch happen in their own organizations. The behaviour is not hidden. It is known, named, and stable.
It persists because the incentives are pointed at it. A decision that has been voiced in a meeting has already cost its owner something to voice. Unvoicing it costs more. The request arrives phrased as a task rather than a question: “pull the numbers that show”, not “find out whether”. A task has no failure mode except being late. And the analyst who answers “actually, they do not show that” is asking a senior person to be wrong in front of colleagues, which is a large favour to ask of someone three levels down.
None of that requires bad faith. It only requires that nobody in the chain has a job description that includes disagreeing.
What does the Bias actually cost?
The obvious cost is the bad decision that got approved. That one is real, but it is also rare and survivable. Most organizations can absorb a wrong call about Toronto.
The expensive cost is what happens to the evidence function. If analysis is a service that reliably produces support for whatever was already decided, then the organization gradually and correctly learns to discount all of it — including the analysis that would have been right. You end up paying for a capability whose output nobody weighs, because everyone knows how it was commissioned.
And the deepest cost is the loss of the ability to be surprised. An organization that only ever confirms itself has no mechanism for discovering that its model of the world has drifted. It will keep executing a strategy long after the conditions that justified it have gone, and it will have a folder of charts showing that everything is fine. That is the elephant charging: not one bad decision, but the removal of the thing that would have caught it.
How do you know the Bias is in your room?
It is hard to see in yourself and easy to see in the paperwork. Look at how questions are worded and how often answers come back inconvenient.
- Requests that name the finding rather than the question: “show that” instead of “find out whether”.
- An analytics function whose results have not contradicted a senior view in living memory.
- A metric that entered the deck when it agreed and quietly left when it stopped.
- The phrase “we just need something to show”, said without embarrassment.
- Post-mortems that examine the execution and never revisit whether the original number meant what it was taken to mean.
Isn't this just what having a hypothesis looks like?
It is a fair objection, and the distinction matters more than the disapproval. Starting with a belief is not the problem. It is how anyone competent works, and a leader with no priors is not open-minded, they are uninformed. Expertise is largely a stock of well-earned hunches.
The difference is whether the answer could have come back the other way. A hypothesis is a claim you have arranged to be able to lose. A justification is a conclusion you have arranged to be able to defend. From outside they can look identical on the day. The tell is what was said beforehand about what would count as disconfirming, and whether anyone wrote it down.
Which is why the cheapest intervention here is also the least technological: before the analysis, get the decision-maker to say out loud what result would change their mind. If the honest answer is “nothing would”, that is useful too — it means the decision is already made, and the analyst can be spared a week of producing a costume for it.
How do you get the Bias out of the room?
The same three moves we bring to any elephant, pointed at this one.
Map
Follow the requests, not the dashboards. Take a quarter of analysis tickets and sort them by how they were worded and how they landed: how many asked a question, how many named their answer, and how many ever came back inconvenient. That ratio is the diagnosis, and it takes days rather than quarters to produce.
Prove
Pick one recurring, consequential decision and run it with the disconfirming condition written down first, stated by the decision-maker, before the data is pulled. One decision, one cycle. The point is not the outcome. It is that the organization sees a senior person survive being contradicted, which is the actual barrier.
Scale
Make the wording a habit rather than a policy. “What would change your mind?” asked routinely, by people who are expected to ask it, does more than a governance framework nobody reads. Unlike a framework, it costs nothing and cannot be delegated to us.
Find out which decisions your data is dressing up
The Elephant Safari names your herd in ten questions and ranks them by what they cost you. Or start from the business problem instead: having more data than ever and still deciding on gut feeling. Get the Bias out of the room. Ask the question you could lose.
Every figure in this article traces to a primary source. See it in Knowledge