The first data hire in most companies is a data scientist, because that is the title everyone has heard of, and it is usually the wrong one. A data scientist joining a company with no reliable data infrastructure spends their first year cleaning exports, writing queries for other people, and rebuilding the same dashboard — work they are overqualified for and did not join to do. They leave within eighteen months and the company concludes that data was oversold.
The first hire should almost always be an analytics engineer or a strong analyst who is comfortable building pipelines: someone who can get data reliably from the source systems into one place, model it so the definitions are consistent, and build reporting people trust. That is unglamorous and it is the foundation everything else needs. Modelling comes later, and comes much faster once the foundation exists.
The problem they are actually solving is definitional rather than technical. In most companies, three teams report three different numbers for revenue, active users or churn, all of them defensible, because each built their own query with their own assumptions. Until there is one agreed definition of each core metric, expressed once in a modelled layer that everyone reads from, every meeting includes an argument about whose number is right. Resolving that is worth more than any model.
Expect the first serious finding to be uncomfortable rather than insightful. Almost every company that builds proper reporting discovers that a number in circulation was wrong — a metric double-counting, a filter excluding a segment, a dashboard nobody had checked since the person who built it left. Handling that well, as a fixed problem rather than as a blame exercise, determines whether the data function is trusted or resented.
Be realistic about where machine learning fits. Most companies' highest-value data work for the first two years is describing what happened accurately and making it available. Prediction requires labelled history, stable definitions and someone to act on the output, and companies frequently have none of the three. A model built before the reporting layer exists is trained on data nobody trusts, which is the same problem one level further along.
Finally, embed the function rather than isolating it. A central data team receiving requests through a queue becomes a bottleneck and produces work that misses the point. Analysts who sit with the teams they serve, understand the operational reality, and are present when decisions are made, produce different and better work — and they surface the questions worth answering, which nobody in a request queue ever does.
