Whether your data team is two people, twenty, or doesn’t exist yet, the kind of organization most companies actually need — clear ownership, real escalation, decisions on the record — is one nobody can afford to hire on day one. Crewdata is a lab exploring whether AI agents can fill those roles: not a copilot, not another observability layer, but teammates with a domain to own, an escalation path when things break, and a paper trail when they decide.
Most data problems aren't data problems. They're organization problems. Crewdata starts from three convictions about how to fix that.
A team with great tools but no organization produces mediocre results. A team with clear organization — even with modest tools — produces reliable ones. We don’t sell skills to agents. We give them an organization to operate within.
Most incidents don’t start with bad data. They start with a change nobody announced, a metric defined two different ways, or a decision made in a meeting that never reached the model. An organization with clear ownership, working escalation, and decisions on the record isn’t bureaucracy — it’s how the translations stop getting lost.
A tool waits to be used. A teammate takes ownership. Crewdata agents don’t just execute instructions — they own a domain, escalate what doesn’t fit, document their decisions, and learn from what went wrong.
Data is never the first priority in a startup. Not the second. Not the third. But it's exactly what investors ask about when they're deciding whether to keep investing.
By the time a team is big enough to tackle data seriously, the debt is already there. Every shortcut taken when data wasn't the priority. Every model built to answer one question, now answering ten. Every ownership decision deferred because there was always something more urgent.
And when the org gets larger? The team fragments. Domains, tribes, squads. Twenty people in data can still mean twenty different mental models of the same problem. A strong central governance team helps — but it doesn't fix the fragmentation. The sins just move around.
That's the observation we kept coming back to. And it's what led to Crewdata — a lab exploring whether a small team can have real data organization from day one, before the debt arrives, before the fragmentation happens.
Built by Pol Martí — data leader who has built data organizations from scratch, more than once. This is what he kept seeing break.
A coordinated row tells you everything is running. The displacement — the element that breaks the line — is where the organizational work lives. The incident nobody owns. The metric that moved before anyone noticed. The escalation that stalled because it wasn’t clear whose job it was.
In a small team, nobody is assigned to catch this. There’s no data quality manager. No governance lead. No escalation path. The aggregate looks fine until it doesn’t, and by then it’s someone else’s crisis.
The lab is building the organization that catches the displacement before it becomes a problem — and knows exactly what to do with it when it does.
Formula 1 teams don’t test new setups in a race. They test them in a simulator that replicates the car, the track, the temperature, the tyre wear, the driver’s reflexes — every variable that matters. By the time a setup reaches the track, it’s been run against thousands of scenarios.
Data organizations have no equivalent. New ownership models, new escalation rules, new governance structures — they all get tested live, on real data, with real teams, against real consequences. When something breaks, you usually find out months later: after the dashboard goes wrong, after the postmortem, after the people who made the call have moved on.
Crewdata is built as that simulator. A complete fictional company — with domains, dashboards, dependencies, and a team of agents operating as the data org — running close enough to reality that we can ask the questions you can’t ask in production:
We don’t just observe — we inject. A column that silently changes meaning. A metric defined two ways in two teams. A stakeholder asking the same question twice and getting different answers. Then we watch what the organization catches, what it misses, and where the escalation stalls.
Even if Crewdata never becomes a product, the simulator leaves something behind: a public record of how data organizations actually break, and a realistic environment to test any framework — ours or anyone else’s — against patterns nobody else has the patience to reproduce.
A CDO. A data director. Domain managers. Analysts — each trained on one slice of your business. Operating within a hierarchy with explicit escalation rules.
These are the roles that keep data reliable in a mature organization. In most teams, they don’t exist — there’s one person, or two, doing all of it. In some companies they were never even on the hiring plan. Crewdata builds that structure as a system of AI agents, so a team of two has the same operating model as a team of twenty — and a team of zero has somewhere to start.
What AI really changes isn’t intelligence — it’s consistency and speed. Every incident handled the same way. Every metric documented the same way. The reviews that used to wait for someone with time, the documentation that lived in someone’s head, the work everyone knew was needed but never reached the top of the queue — they actually happen. The marginal cost of doing things properly drops to nearly zero.
Human review is part of the design. Agents escalate what they can’t resolve. Decisions are documented. The humans on your team stay in control of what matters — they just stop doing the work the org should be handling automatically.
First incident flows in production. Data quality escalation end-to-end.
Agent memory, cross-domain ownership, semantic layer integration.
Whatever the lab learns gets written down. Plain language, dated, reproducible. We won't share code or implementation details — but we will share what works, what breaks, and what surprises us along the way.
We’ll send the next note when it’s ready. No newsletter loop. No growth hacking. Just what the lab learns — and what you can use.