Data-Team Economics After LLMs: The Collapse of the Analyst-Report Pipeline

For fifteen years, the modal hire on an enterprise data team has been a mid-level analyst who could write SQL, build a dashboard, and answer the recurring questions of a business partner. That role is now, in practical terms, being priced out of existence. Large language models paired with a well-instrumented semantic layer can answer most of the questions that used to fill an analyst's queue, and they can do it during the meeting rather than three days later.

The consequence is not that data teams shrink. It is that their shape changes. The analyst-report pipeline collapses, and the value moves upstream to the semantic layer, the metric definitions, the evaluation harness, and the governance around who is allowed to ask which questions of which data. Chief data officers we have spoken with in the last two quarters are hiring against a different profile: closer to a platform product manager, with deep opinions about schema design, metric ownership, and access control, and comfortable being on the hook for the quality of what the model returns.

The implication for how data organizations are structured is significant. The old dotted line, where analysts sat inside business units and were coordinated by a central data function, no longer justifies its overhead. The new pattern is a small central platform team that owns the model, the semantic layer, and the guardrails, plus a much smaller number of embedded specialists who own hard problems, principally causal inference, forecasting, and experimentation, that models still get wrong in expensive ways.

There is also a hard truth about data literacy. When any executive can pull a number by asking, the marginal cost of being wrong drops to zero and the volume of wrong numbers going into decisions goes up. The organizations handling this well are investing not in more analysts but in fewer, better metric definitions and a clear rule that any number that appears in a board pack must resolve to a governed metric. The organizations handling this badly are watching their forecasting accuracy quietly deteriorate.

The question for a CDO to bring to the next executive session is not how many analysts you need. It is which twenty metrics your company actually runs on, who owns each one, and whether the definition survives the model asking clarifying questions about it. If you cannot list them, your data function is about to be graded on outputs it does not control.