Notes from conversations with heads of growth and analytics at eight enterprises rethinking how their data teams spend the week.
Over the last two quarters we sat down with heads of growth, analytics and data science at eight enterprises rethinking how their teams spend the week. The composition of the teams has barely changed. What they do with their days has changed completely. These are notes from those conversations, with the specifics anonymized and the patterns kept.
In 2022 the archetypal week at a top B2C data science team looked roughly the same everywhere. Monday and Tuesday went to pulling and reconciling data for the standing dashboards. Wednesday was spent in reviews with the business, defending or explaining last week’s numbers. Thursday was the one day of actual modeling work. Friday was the model going into a slide for the following Monday.
That week produced roughly one shippable insight per person per month, generously counted. Nobody was happy with it, including the people doing the work. It was, however, the equilibrium the tooling allowed.
At the eight teams we studied, the week has reorganized around a different unit of work. It is not the report. It is not the model. It is the shipped, causally read experiment. Every activity on the calendar traces back to raising the rate at which those get produced.
In every one of the eight teams, the weekly BI report is no longer produced by hand. It is generated, checked and published automatically. The analyst’s job is not to produce the report. It is to interrogate it when a number looks off. That reclaimed roughly one day per person per week.
The pattern that used to consume Monday, a business stakeholder asks for a slice, the analyst pulls it, the stakeholder asks a follow-up, was replaced by a self-service surface over the same underlying model. The stakeholder pulls their own slices. The analyst is called in for the modeling questions, not the fetching ones.
A well-authored hypothesis is now the highest-leverage artifact in the team. It scopes the segment, states the expected direction, names the guardrail metrics and cites the priors. Bad hypotheses waste the entire downstream pipeline. The teams that treat authorship as a craft materially outperform the teams that treat it as an afterthought.
The conversation the head of data has with the head of growth used to be about dashboards. Now it is about experiments. Nobody argues about a chart in isolation; every argument grounds out in the incrementality readout of a specific shipped variant. This is the single largest cultural shift, and the one the interviewees mentioned most often unprompted.
The teams that made this shift did not grow. Two of the eight actually got smaller by attrition and did not backfill, and their throughput went up anyway. What changed is the internal shape.
Every team we spoke to had settled on a version of the same internal metric: shipped, causally read experiments per analyst per quarter. In 2022 the median for this group was two. In 2026 it is between twenty and forty, depending on the surface. That is the productivity delta, and it is the one number that predicts the ARPU trajectory of the business over the following two quarters.
Three attempts came up repeatedly as false starts. Worth naming, because they are the natural first moves and they do not pay back.
The consistent throughline from these conversations was the shift from producing artifacts to producing rate. Rate of hypotheses, rate of shipped experiments, rate of validated learnings. Every architectural choice downstream of that framing pointed in the same direction.
For the operating model these teams migrated into, see From campaign calendars to continuous decisioning.
Frequently asked
See it in the product
The same loops this note describes run 24/7 against your customer base. Watch the workspace decide, experiment and execute 1:1.
