Most businesses are not short of data. They are short of questions specific enough to answer with it. A folder of exports and a dashboard nobody opens are symptoms of the same thing: the analysis was never attached to a decision.
Start from the decision, backwards Ask which recurring decision is currently made on instinct. Which customers to call first. How much stock to hold. Which service line to staff up. Then work backwards to the smallest piece of evidence that would change that decision. This produces narrow, answerable questions instead of a request to "get insights from our data".
Expect the first phase to be unglamorous Reconciling two systems that spell customer names differently, agreeing what counts as an active account, deciding whether a cancelled order stays in the total — this is where the time goes, and skipping it is how organisations end up with two confident reports that disagree. Write the definitions down; they matter more than the modelling.
Description before prediction A clear picture of what happened last quarter, broken down along one or two dimensions that matter, resolves more arguments than a forecast does. Prediction is worth adding once the description is trusted and someone can articulate what they would do differently at a given predicted value.
When machine learning is the right tool It fits when the pattern is real but too complicated to express as rules, when enough labelled history exists, and when being right most of the time is genuinely useful. Price estimation, demand patterns and text classification often qualify. A one-off question with forty rows behind it does not; a well-built spreadsheet is the correct answer there.
Reading text at scale A great deal of business knowledge sits in language — reviews, support tickets, survey comments, call notes. Classifying that text into themes turns anecdote into a countable dataset, and it frequently changes which problems get prioritised, because volume finally becomes visible.
Build the pipeline, not the one-off An analysis that has to be reproduced by hand each month will be reproduced twice and then abandoned. A scheduled pipeline with checks on freshness and record counts keeps the answer current and, importantly, tells you when it has stopped being current.
Close the loop Every piece of analysis should have an owner and a stated action. If nobody can say what they would do differently, the work is interesting rather than valuable, and it is fair to stop.
Our Python, data and AI work follows this sequence. Contact us with the decision you would most like to make on evidence rather than instinct.
AI Studio · Published 2 Mar 2026
← All insights