Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Why 90% of Data Teams Are Failing at Data Modeling - Freestyle Friday (May 15, 2026)
TL;DR
The real data-modeling crisis is organizational, not technical — In Joe Reis’s April survey of 334 respondents, 90% reported pain points, with time pressure and unclear ownership showing up as the two recurring root causes across this survey and his larger January state-of-data-engineering survey of 1,100 people.
People want skills, time, and ownership far more than new tools — When asked what would improve things, respondents chose training and education (28%), clearer business requirements (24%), more time for modeling (21%), and dedicated ownership (21%), while only 4.8% asked for better tooling.
Most teams are doing physical modeling by default and 'vibing it as they go' — Joe says the de facto pattern is whoever builds the pipelines or dbt/SQL transformations also owns the physical model, which leads to ad hoc standards and constant firefighting.
His most practical fix is almost boring: assign a single accountable owner — Joe argues teams can improve fast in “one meeting” by naming a custodian who understands how business reality maps to the model and is empowered to maintain it.
AI isn’t removing the need for modeling fundamentals — it’s exposing them — Joe’s January survey found 82% of respondents use AI tools at least daily, but his point is that AI amplifies what teams already know and don’t know, so weak fundamentals get speedrun into bigger problems.
Data modeling is not just schema design; it’s shared meaning for humans and machines — Referencing semantic-layer debates and a conversation with Looker cofounder Lloyd Tabb, Joe frames the model as a central representation of meaning that agents and humans both need, even if companies disagree on implementation.
The Breakdown
Walking Salt Lake City and reading the survey tea leaves
Joe opens literally out on a walk in Salt Lake City, using the fresh-air Freestyle Friday format to revisit his April data-modeling survey. He situates it alongside his January state-of-data-engineering survey with 1,100 responses and March’s AI-usage survey with 193 respondents, then says the same two problems keep surfacing: people are crushed for time and nobody clearly owns the work.
The numbers say the bottleneck isn’t tooling
The most revealing stat dump comes fast: training and education led the "what would help" list at 28%, followed by clearer business requirements at 24%, more time for modeling at 21%, and dedicated ownership at 21%. Better tooling got just 4.8%, which Joe says should be a wake-up call for vendors: yes, tools reduce friction, but they don’t solve a skills gap or a broken org.
Sponsored detour, with a point about undifferentiated heavy lifting
Joe shouts out Fivetran and Revify, but ties both reads back to the theme. His Fivetran pitch is basically a reminder from Fundamentals of Data Engineering: don’t waste precious time on "undifferentiated heavy lifting" if pipelines aren’t your core competency; and Revify gets a nod for a case where it cut a Snowflake bill by 50% across hundreds of warehouses in 48 hours.
Why 90% of teams hurt: ad hoc modeling creates a fire-fighting loop
Back to the data: 90% of respondents said they have some kind of data-modeling pain point. Joe’s diagnosis is blunt — disciplined teams with standards fight fewer fires, while teams "vibing it as they go" with ad hoc SQL and dbt scripts stay trapped in a feedback loop where lack of time causes bad modeling, which creates more fires, which destroys even more time.
The simplest fix: pick an owner in one meeting
His advice is almost aggressively practical: choose an owner or custodian for the data model and make that person directly accountable. That owner should understand the business domain, translate reality into the model, and keep that shared map coherent for everyone else.
AI makes the fundamentals matter more, not less
Joe pushes back on the idea that AI is an easy button. Citing a recent Stockholm talk, he repeats the line that AI amplifies what you know and what you don’t know; speed helps if you can maneuver, and hurts if you can’t. In his framing, data modeling isn’t just schema design — it’s the work of creating shared understanding among stakeholders, now including machines.
Semantic layers, Lloyd Tabb, and the hard work nobody wants to hear about
He brings up semantic layers as the current flashpoint, noting some say they’re essential and others say they’re not, but he’s firmly in the camp that says you need some central representation of meaning. Recounting a conversation with Looker cofounder Lloyd Tabb about Malloy, Joe says the frustrating truth is that rolling this stuff out in companies just takes forever — there is no easy button.
Leadership pressure, org charts, and the architecture question AI is forcing
The back half turns more human and systemic. Joe says the free-form comments felt like a therapy session, then references a conversation with Eric Weber, formerly head of data at Yelp, Grammarly, and Superhuman, about how leaders — especially in the Bay Area — are under intense AI-fueled pressure to do more with less. He closes on a more speculative note with Conway’s Law and his joking "Reis’s law" that the data model mirrors the architecture, wondering what happens to systems, ownership, and time pressure as org charts get flattened or even agent-run.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
