Every data leader inherits an org model, usually one chosen by accident. The wrong shape doesn’t just slow delivery — it dictates which problems are tractable and which never get solved. There are four models worth taking seriously, and each one fails in a predictable way.

1. Centralized

A single data team reports to one leader (often a VP of Data, sometimes the CFO or COO). All analysts, engineers, and scientists sit together. Business units file requests through a queue or a single point of intake.

When it works. Early-stage companies under ~150 employees, or any business where reporting needs are still converging on a shared definition of the metrics that matter. Centralization forces a single definition of “revenue,” a single pipeline, a single dashboard layer. It’s the cheapest way to get a coherent first version of the data estate. It also concentrates rare senior talent — one strong head of data can lift the floor for the whole company.

Failure mode. The queue. Once the business has more than three or four functions making requests, the queue becomes the bottleneck and the team becomes a ticket-handling shop. Analysts stop developing domain expertise because they rotate across requests. Stakeholders stop trusting timelines. The senior people leave first, because they did not take the job to be a Jira-driven internal agency. Centralized teams that survive past 200 employees almost always do so by quietly evolving into one of the next three models.

2. Embedded

Analysts (and sometimes engineers) report into the business unit they serve — marketing analysts report to the CMO, finance analysts to the CFO, product analysts to product. There may or may not be a dotted line to a head of data.

When it works. When domain context is the binding constraint on quality. Pricing analytics, growth experimentation, sales operations — these reward a person who knows the business cold and sits in the same standups as the people running it. Embedded analysts ship faster on their domain than any centralized team will, because they skip the intake-and-translation tax.

Failure mode. Definitional drift. Three embedded analysts will produce three different revenue numbers within six months. Tooling fragments. Everyone builds on a different warehouse layer, or worse, on extracts. Career paths break — there’s no senior analyst above the embedded person, so growth means leaving. And because each analyst is a team of one, there’s no peer review; quality is whatever that one person’s standards are. Embedded-only orgs typically discover, around year two, that they have no idea which of their numbers are right.

3. Hybrid (hub-and-spoke)

A central data platform/engineering team owns infrastructure, the warehouse, and shared models. Embedded analysts in business units consume from that shared layer and build last-mile analyses. Definitions live in the central team; speed lives at the edge.

When it works. Most companies between 200 and 2000 employees, and a fair number above. It captures the best of the first two: shared definitions, shared tooling, shared career ladder, but with domain-embedded analysts who don’t have to file a ticket to ask a question. The central team becomes a force multiplier rather than a bottleneck.

Failure mode. Two of them, both organizational. First, the central team gets pulled into being the queue anyway, because embedded analysts ask them to build “just one more” model. The hub forgets its job is platform, not service. Second, the dotted line breaks. Embedded analysts, under pressure from their business unit, route around central definitions. Within a year, you have three definitions of churn again, only now you also have a platform team that thinks it owns the canonical one. Hybrid orgs require explicit governance — RFC processes, contract tests on shared models, a metrics layer with teeth — or they decay back into embedded chaos.

4. Platform-team-of-experts

The central group is small and elite. They build a self-serve platform — strong semantic layer, well-documented models, opinionated tooling — and then teach the rest of the company to use it. Embedded analysts exist, but the goal is that any reasonably technical PM, marketer, or operator can answer their own questions against the platform. The data team’s product is the platform itself.

When it works. Mature, technically sophisticated companies — the Airbnbs, Stripes, Shopifys of the world. It also works in smaller orgs where the population genuinely is technical: dev-tools companies, infra companies, quant firms. The model demands a culture that rewards self-service and a leadership team that stops asking the data team to “just pull the number.”

Failure mode. It’s a category error in most companies. Self-serve assumes your stakeholders want to learn SQL or its equivalent, and most don’t — not because they’re incapable, but because their job is something else. The platform team builds elegant abstractions that nobody uses; meanwhile real questions go unanswered. The other failure mode is that the platform team becomes monastic — beautiful internals, no business impact — and gets cut in the next cost cycle.

Maturity trajectory

A useful default for most B2B companies, by stage:

  • Pre-50 employees. One analyst, embedded with the founder/CEO. No model needed yet.
  • 50–200. Centralized team of two to five, reporting to the COO or CFO. Get to one definition of revenue, one warehouse, one BI layer.
  • 200–800. Hybrid. Spin out a platform/engineering function (two to four people), embed analysts in marketing, finance, and product. Invest hard in a metrics layer before drift sets in.
  • 800–3000. Hybrid matures. The platform team grows into a real product team — owning ingestion, transformation, semantic layer, governance, observability. Embedded analytics scales out by function. A head-of-data role with real authority becomes non-optional.
  • 3000+. Optionally evolve toward platform-team-of-experts, but only if your stakeholder population is genuinely technical and your platform is genuinely good. Most large companies stay in mature hybrid forever, and that’s fine.

The mistake is not picking the wrong model — it’s keeping the right model for too long. Centralized at 300 employees is a queue. Embedded at 800 is three definitions of revenue. Re-org before the failure mode locks in, not after.