Golden Data

Issue #22 · September 20, 2026

Omni Let Anyone Build the App. The Analyst Still Owns the Definition.

Paul BrownWrites Golden Data from Ogden, Utah7 min read

“Whose number is in it?”

Someone in ops will ask the agent to build a tool on your model this quarter and show it to you finished. That question won’t have an obvious answer. It should.

On August 6, Omni shipped Apps. A user describes the tool they want, an agent builds it, and one-third of customers were building them within weeks. Five demo videos followed, every Friday through September 4, every title with the word Apps in it.

Under the videos sits the sentence the whole thing rests on. Every query in a generated app is a reference to a field in your semantic model, the one built by your experts. The app builder is anyone with a Create Apps permission. The model builder is “your experts.”

One vendor. Two permissions. One got cheaper this month.

That’s one product. The wider read: Looker’s verified queries went GA on August 28, 30 to 50 hand-picked question-and-query pairs per agent. Snowflake’s Semantic Studio entered preview on August 26 with Git and a debugger. dbt Labs said in June that an analytics engineer competing with AI on code production will find the role shrinking. Three vendors. One object. The dashboard got automated. The definition under it got a version number and an owner.

The counter-read is real, though. The layer under the app is being automated too. Omni’s Modeling Agent scaffolds the first pass, then pauses for approval. In every mode, the vendor kept one verb human. Approve. The job moved to the pause.

One thing to do this Friday. Take the ten metrics your Monday call cannot run without. Score each on four lines: owned by a named person, tested by a check that fails the build, documented in prose a stranger’s agent can read, tied to a named decision. Forty cells, one point each. A four can carry an app. A zero is the one the app gets built on first, because it was the easiest to find.

Omni has posted a demo video every Friday since August 7. Five in a row, through September 4. Every title has the word Apps in it. August 21 added writeback, skills workflows, and AI summaries. August 28 added user memory and a title that reads “Omni With Any LLM You Can Imagine.”

The blog post under the videos went up August 6 and was revised August 26. A user describes the tool they want. An agent builds it. “In the first weeks after release, over one-third of customers were already building and using them.” Then the sentence the whole thing rests on: “Every query in the generated App is a reference to a field in your semantic model.” Which model? “The same one built by your experts and trusted to run your dashboards.”

This publication spent the summer saying the dashboard was never the product. It is odd to have a $1.5 billion vendor agree, in a changelog, on a Friday.

Read the post as an org chart. The app builder is anyone with a permission: “A Create Apps permission lets admins choose who can build apps while everyone else can still use them.” The model builder is “your experts.” One got cheaper this month. The other did not. The story where AI replaces the analyst needs those to be the same person. The vendor permissioned them separately.

What’s actually moving in the market

Omni, August 6 through September 4. Every app is paired with a workbook holding up to 100 queries, “where anyone with access can open it, inspect the SQL and semantic model behind it, and tune it.” One hundred is the ceiling on what an app can ask. Zero is the number of definitions it can hold. Nothing in an app is a definition. Everything in it points at one.

Omni, April 9 and the week of July 20. The counter-read. The layer under the app is being automated too. The Modeling Agent, announced April 9 by Jade Khiev, lets teams “let AI scaffold the first pass and then review.” In Review mode “the agent writes changes inline, then pauses for approval before continuing.” The week of July 20, “AI-powered semantic model generation is now generally available.” In both modes the vendor kept one verb human. Approve. The job moved to the pause. Pricing is not published. Who pays for the pause is a sales call.

Looker, August 28. Verified queries for Conversational Analytics went generally available: “a predefined pair of natural language questions and their exact, corresponding Looker Explore queries.” Google recommends 30 to 50 per agent. The best-practices page draws the line itself: “Unlike LookML, which is governed, instructions are often free-form text and can become ‘stale’.”

Snowflake, August 26. Semantic Studio entered preview, “the authoring environment for semantic views in Workspaces, combining conversational authoring with CoCo, direct YAML editing, and deployment to the live Snowflake object in a single surface.” It has Git and a debugger for wrong answers. Five days earlier, version targeting for Cortex Analyst evaluations went generally available. A definition with a version number, an eval pinned to it, and a debugger is a file with an owner. Omni, Looker, Snowflake. Three vendors. One object. Each shipped a new way to build on it and a new way to check it. None shipped a way to skip it.

dbt Labs, June 16 and April 14. Daniel Poppy, dbt Labs blog, June 16: the deliverable is now “the contracts that protect those models, the tests that validate them, and the semantic definitions that make them machine-readable.” And: “An analytics engineer who spends 2026 competing with AI on code production will find the role shrinking.” The 2026 State of Analytics Engineering report, April 14, 363 respondents: 72% prioritize AI-assisted coding. 24% prioritize AI-assisted pipeline management, including testing and observability. 71% name hallucinated outputs reaching stakeholders as a top concern. The fast half got automated. The half that catches the wrong number stayed where it was.

What I’d do this week

The action. One, and it fits in a Friday afternoon.

Take the ten metrics from the August 26 Signal, the ten the Monday call cannot run without. If you skipped that one, take the ten you would least like to see inside an app someone else built. Score each on four lines.

Owned. A named person, and that person knows. Not a team. Not “data.” If you have to ask, the answer is no.

Tested. A check that fails the build when the number is wrong. A not_null test checks the column. A reconciliation to last quarter’s reported figure checks the definition. Only the second one counts here.

Documented. Grain, filter, exclusions, in prose a stranger’s agent can read. A column name is not documentation. Neither is the SQL.

Has a named decision. The meeting where this number changes what someone does. No decision, not owned. Computed.

Four columns, ten rows, forty cells. One point per cell.

The user moment: someone in ops asks the agent to build a tool on your model and shows it to you finished. The shape: a ten-by-four table with a score column, dated, committed next to the model. The time budget: two hours. Success: before that demo, you know which of your ten can carry an app and which will carry a wrong number into a room you are not in. A four is app-ready. A two is a dashboard with the owner missing. A zero is the one the app gets built on first, because it was the easiest to find. Credible failure: every row scores a three or better, and the row that lost its point lost it on Owned.

Omni let anyone build the app. It also wrote down who it still needs. That sentence is a job description. It is a better one than the JD.

Sources

Crafting runs weekly. The job is to be useful, not to sell. If a line in here is close to a move you're working on, the email reaches a person.

Hiring rather than hunting? Signal is the column for your side of the desk.

Email meRead more Crafting

← All Crafting issues · Drafted with Claude · Edited by Paul Brown