The dashboard says churn went up. It cannot tell you why, or whether last quarter's price change caused it. We answer questions like that, with the working shown.
Four people are looking at the same bar chart and arguing about what caused it. One of them has a theory, another has a competing theory, and the meeting ends with whoever spoke last. The number on the dashboard is real; the explanation behind it is not in the dashboard. Getting to an answer somebody can defend takes the raw records, a stated method, and a willingness to say how confident the answer is.
Work usually arrives as a sentence: why did this move, what happens next, who are these people, did our change do anything. Those are four separate jobs, with different methods and different things they cannot tell you.
The question to ask of any finding is what would have to be true for it to be wrong. That means the assumptions are written down rather than buried in a filter somebody set weeks ago, and the rows that were dropped are named along with why. Most disagreements about an analysis turn out to be disagreements about which rows counted.
So the work ships with its working: the query or notebook that produced the number, the version of the data it ran against, and a result you can re-run yourself next quarter and get the same figure. A finding that cannot be reproduced has a short life, because the first person who dislikes it will ask for it again.
Four things, and the last two are the ones that get skipped elsewhere. A number without its method is a rumour, and a number without its limits gets quoted in a board pack.
One or two pages at the front: what we found, what it means for the decision you asked about, and how strongly the data supports it.
Which records were used, which were excluded and on what rule, what was assumed, and which alternative approaches we tried and set aside. Written for someone who wants to disagree.
The notebook, SQL and any scripts, in your repository, running against your warehouse. Your analysts can re-run the whole thing on next quarter's data without asking us.
The caveats, plainly: the population this does not cover, the period it holds for, and the questions people will try to use it to answer that it cannot.
The first step is usually the one that gets cut, and it is the one that decides whether the rest is worth anything. Most of the risk in this work sits in the data rather than in the modelling.
Most briefs arrive as a topic rather than a question. We push until there is a sentence with a decision attached to it, and a definition of every term in it, including what counts as churn. You sign off that sentence before anything else happens.
Before any analysis, we check what the records can support: coverage, gaps, duplicate identities, fields that changed meaning when a system was replaced. You get a short memo on what is trustworthy and what is not, and sometimes it changes the question.
The method follows from the question rather than from whatever is fashionable. We keep a simple baseline alongside anything more elaborate, because a regression that beats no benchmark tells you nothing. Interim findings come to you while they are still cheap to redirect.
A result gets tested against the obvious objections before you see it: does it survive if we change the window, drop the largest account, or define the segment differently. The review session walks you through the checks that failed as well as those that held.
The write-up, the code and the caveats land in your repository. Then we sit with your analysts while they re-run the whole thing themselves and answer questions. If the answer needs refreshing monthly, you leave with something your team runs, not a standing invoice.
We scope the first step, not the whole answer. The data audit is a fixed piece of work that ends with a memo saying what is answerable and what it will take. Plenty of engagements stop there, because the memo says the question needs better tracking before it can be answered at all. That is a cheaper outcome than finding out at the end.
Describe the problem in a paragraph and we will tell you which service applies.
