Every startup ends up in one of two failure modes on data. Either the founders are flying blind — no dashboards, decisions made on vibes, no idea what activation or retention actually looks like — or the company has spent $500K on a modern data stack, hired three analysts, and still cannot answer the CEO's question about last week's conversion rate.
One product analytics tool. Pick one: PostHog, Amplitude, or Mixpanel. Instrument the 10 events that matter: signup, activation, first key action, second key action, upgrade, churn, and a few core product interactions. Do not try to be comprehensive. You will replace the taxonomy in 12 months anyway.
One dashboard visible to the whole team: DAU/WAU/MAU, signups, activations, revenue. Update daily. That is the entire data stack.
No data warehouse, no dbt, no BI tool. If a founder can't answer a question from PostHog in 5 minutes, the question is probably premature.
The cracks show up. The CEO wants a self-serve funnel by acquisition channel. The head of sales wants ACV by segment. Finance wants MRR waterfall reconciled to Stripe. PostHog cannot do all of that alone.
A data warehouse. BigQuery, Snowflake, or Redshift. BigQuery is easiest at this stage.
A data pipeline. Fivetran or Airbyte to sync Stripe, HubSpot, PostHog, and your production database into the warehouse. dbt for transformations. Even a small dbt project prevents 90 percent of the "which number is right" fights.
A BI tool. Metabase (open source) or Hex or Mode for team dashboards.
One analyst or analytics engineer. Not a data scientist. Not an ML engineer. Someone who can write clean SQL, model data in dbt, and ship dashboards.
A small data team: 2 to 4 analytics engineers and analysts under a head of data. Add a governance layer:
Metric definitions stored in code (dbt semantic layer, LookML, or Cube)
A single source of truth for every north-star metric — every board deck, exec dashboard, and investor update pulls from the same place
Data quality monitoring (Monte Carlo, Metaplane, or open-source alternatives)
Experimentation platform (Statsig, Eppo, GrowthBook, or homegrown) — most consumer and product-led companies need this by now
A CDO or VP of Data. Separate teams for analytics, ML/AI, and data platform. Real ML infra if the product requires it (feature stores, model registries, online inference). Data as a customer-facing surface: reports, exports, embedded analytics.
1. What decision would we make differently if this data were 10x better? If the answer is "none," don't build the pipeline. 2. Who owns this number end to end? Every north-star metric should have a single named owner. Numbers without owners rot.
Instrumenting everything, using nothing. 400 events tracked, 3 dashboards actually opened weekly.
Warehouse before you need it. Every hour spent on data infra at seed is an hour not spent on product.
Data team siloed from GTM and product. Analysts embedded in a squad ship 10x more useful work than analysts sitting on a "data team" island.
Hiring a data scientist before an analytics engineer. The ML models will not help if the underlying data is broken. Fix the foundation first.
A good data function makes the whole company smarter. A bad one makes the whole company slower. The difference is discipline about what to build and when.
