CausoAI drafts a causal diagram from your data automatically, but it does not ship that diagram straight into the analysis. You review it first. This is the one step in the process that genuinely needs a person, and it is worth understanding why before you look at what to check.
Why software cannot finish this on its own
Data tells you two things move together. It cannot always tell you which one drives the other. Ice cream sales and drownings rise together every summer; nothing in the numbers reveals that the sun is behind both, or which direction any arrow should point.
Algorithms narrow this down considerably — timing, statistical structure and known patterns rule out a great many possibilities. But they routinely reach a point where two arrangements fit the data equally well and only knowledge of the business can separate them. That knowledge is yours, not the model’s.
You are not being asked to check the statistics. You are being asked whether the picture matches how your business actually works.
Step 1: Check the arrows that point the wrong way
Start here, because it is the most common error and the easiest to spot. Read each arrow out loud as a sentence: "A causes B." Most mistakes become obvious the moment you say them.
Watch for arrows that run backwards in time. If the diagram says revenue causes ad spend, ask whether that is genuinely wrong — sometimes it is exactly right, because last quarter’s revenue set this quarter’s budget. Timing is the test, not intuition about what sounds sensible.
Step 2: Look for the thing that is missing
This is the highest-value check you can make, and the one the software is least able to do for you. A diagram can only include what is in your data. If something important drives two variables at once and is not in your file, the analysis will attribute its effect to whatever is.
Ask yourself: what was going on during this period that is not in this dataset?
- —Seasonality, holidays, or the shape of your fiscal quarter
- —A brand campaign, PR moment, or competitor event running at the same time
- —A pricing change, packaging change, or website redesign
- —A reorganisation, a new sales leader, or a change in how a team was compensated
- —Anything that changed about how the data itself was collected
If one of these was happening and is not represented, that is the most valuable thing you can add. Even a crude column — a flag for the weeks a campaign was running — is far better than leaving it out.
Step 3: Watch for the middle-man problem
Some variables sit between the thing you changed and the outcome you care about. Email campaigns drive site visits, and site visits drive signups. Site visits are in the middle.
Middle variables need care. If you hold one constant while measuring the effect of the thing upstream, you erase most of the effect you were trying to measure — because that effect travelled through the middle. The diagram needs to show these as a chain rather than as three unrelated arrows into the outcome.
Rule of thumb: if B only happens because A happened, B belongs between A and the outcome, not beside it.
Step 4: Question anything that surprises you
If the diagram shows a connection you do not recognise, do not delete it reflexively and do not accept it because the software drew it. Both reactions lose information.
A surprising arrow means one of three things: the data has found something real that you did not know about, two variables are being driven by something absent from the file, or a column means something different from what its name suggests. The third is more common than people expect, particularly with fields that changed definition partway through the period.
What good enough looks like
The diagram does not have to be a complete theory of your business. It has to be right about the paths connecting the thing you are testing to the outcome you care about, and it has to include the major factors that influence both.
For a first analysis this is usually a ten-minute review, and the readiness score afterwards will tell you if something important is still missing. Every version you save is kept, so you can change your mind, re-run, and compare.