Your AI team is building a data platform nobody decided to fund
I have seen an important decision made on a week-old picture because the data behind it was refreshed by a notebook. The notebook ran outside the data platform's orchestration and monitoring, so when it stopped, nothing alerted, and for a week the use case kept serving data that was no longer true.
That shortcut is not new. I have seen data science teams build their own data pipelines and copies outside data engineering, and I have seen organisations learn that shared standards were cheaper than every team rebuilding the same work. The ones that matured put the frameworks in place, and data science got the common benefits instead of duplicating them.
AI teams are now repeating that lesson, and the urgency to ship AI is where I see the standard frameworks shortcut today. Many growth-stage startups never learned the data science version, because nobody on the team lived through it. The result is a second data platform nobody decided to fund or govern.
Two teams, no interface
The AI team is set up to face the product and is measured on shipping AI features. The data team is measured on platform stability and its backlog. Both incentives are reasonable; what is missing is the interface between them. There is no agreed data contract for the data AI features consume, and no access path faster than a ticket.
So the AI team does what any capable team under a deadline would do. It extracts what it needs into notebooks, object storage, and vector stores, and re-implements the cleaning, joins, and definitions the data team already maintains. Each copy starts accurate and drifts from the governed data sources from the day it is made.
The shortcut fails silently
The shortcut runs outside the platform's orchestration and monitoring, so nothing alerts when it stops. In an AI feature, the cost of a stale refresh is quiet: for example, an account summary prepared for a customer renewal from data that predates the latest support escalation. The answer reads fluently and looks complete. It is simply out of date.
Within months, the second platform collects the problems the first one already solved. Key definitions exist twice, once in the data warehouse and once in the AI team's feature code. Access is granted through a service account with broad read rights, because requesting the right access would have taken weeks. When something breaks, the incident surfaces in the AI feature, and the data team is asked to debug pipelines it has never seen.
The pattern does not stop at team level either. A Howdy.com survey reported by CIO.com found that 16% of knowledge workers already build their own AI agents. The question of what is shared and what is local is arriving everywhere in the company at once.
Nobody owns the boundary
The data lead and the AI lead often report to different executives, and each is doing their job well. The decision that sits between them, what the AI team shares and what it owns, is never made explicitly, so it is made by default, one copy at a time.
The leadership mandate
Speed was the reason for setting up a separate AI team, and it is a good reason. Merging the teams would give that speed away, and forcing everything into the central data platform would too; no single architecture fits every component. The CTO's decision is narrower and more useful: which parts are shared and governed, and which belong to the AI team. Data sources, business definitions, access, orchestration, and monitoring belong in the first group. Prompts, evaluation sets, and model-specific features belong in the second. Once that line is drawn, both teams can move fast on their own side of it.
The path forward
Three agreements make the line hold:
- A data contract between the data team and the AI team for every source the AI features consume.
- A fast path for new data, measured in days rather than weeks, so the shortcut stops being the faster option.
- A joint review before any new copy of production data is made.
None of this is new knowledge. It is a lesson the industry already paid for with data science teams, and for a scaleup, the cheapest moment to import it is before the second platform has users.