Golden Data

Issue #8 · June 12, 2026

The Analytics Interview Changed This Month. Most Prep Is Still for Last Month.

Paul Brown Writes Golden Data from Ogden, Utah 5 min read

Every Friday I check what moved in the analytics craft. Most Fridays the movement is incremental — a JD adjustment, a comp tick. This Friday the data said something.

Three things happened since May 19 that change what a Staff Analytics Engineer needs to walk into an interview knowing. dbt shipped an AI coding agent that reads project lineage and writes dbt models from natural language. Meta, Shopify, Canva, and Rippling announced they now allow AI tools in live technical screens. And the Ravio 2026 Compensation Trends report confirmed a 12% IC salary premium for AI/ML practitioners — alongside entry-level analytics hiring down 73% year over year.

These aren’t separate events. The interview is moving from “can you write a dbt model” to “can you govern the model the agent wrote.” The 12% is what the second skill pays. The 73% is what happened to the pipeline that only knew the first.

The prep guides are still describing the first thing.

What’s actually moving in the market

dbt Labs shipped Fusion 2.0 and the Developer Agent in May 2026. The Staff AE interview question is shifting from “write it” to “govern it.” The Developer Agent reads project lineage, validates changes against upstream dependencies, and authors or refactors dbt models from natural language — available on Snowflake without reconfiguring your project. JDs at companies running Fusion on Snowflake have started adding “AI-generated SQL validation” and “transformation governance” to Staff AE required qualifications. The Snowflake Summit (June 2–5) treated governance tooling as a headline feature. The JDs got there first.

Ravio’s 2026 Compensation Trends report (400,000+ verified employee records) put a number on the bifurcation: 12% IC premium for AI/ML roles, entry-level analytics hiring down 73%. The premium sits at the practitioner level, not management (management premium: 3%). The entry-level collapse means some of the candidates now competing for Staff roles have four years of experience and are still looking. The 12% and the 73% are the same story from opposite ends of the distribution.

PayPal (4,760 layoffs, 20% of workforce, May 5–9) and Dune Analytics (25% reduction, May 14) both named AI automation as the mechanism and analytics roles as the target. PayPal: AI automation and flatter org design, data science and quantitative analytics confirmed in scope. Dune: explicit AI fluency pivot, following 300 interviews conducted on that standard. The fintech pattern is consistent — analytics IC layers below Staff are thinning; Staff roles with governance ownership are staying open.

Meta, Shopify, Canva, and Rippling now allow candidates to use AI tools during live technical interviews — and some require prompt disclosure. Per Karat’s 2026 Engineering Interview Trends report, the screen is no longer “can you write code without assistance.” It’s “can you own an outcome and explain what the AI produced.” The prompt-disclosure norm is emerging as a proxy for senior-level intellectual transparency: can you name the tradeoffs in the code the AI wrote for you?

Power BI’s May 2026 update shipped Git-friendly Copilot metadata storage — the semantic governance story Fabric was missing. Metric definitions are now version-controlled and auditable outside the Analysis Services database. For BI Directors at companies running Fabric: “who owns the definition of revenue?” now has an answer that auditors can find in Git history.

What I’d do this week

Governance drill: write the AI-generated SQL evaluation the Staff AE interview is starting to ask for. The technical screen at companies running dbt Fusion and Cortex Analyst is shifting toward “evaluate and govern AI-generated SQL.” Most prep materials still describe the prior version of this question. The interview has moved.

  • The user moment: Before your next Staff or Principal AE technical screen. Or the next time a PM says "can we just ask the AI layer that question?" — which is this month's version of "can we just query the database directly?"
  • The shape: Portfolio addition — one written evaluation of AI-generated SQL from a real dbt project.
  • The time budget: 90 minutes. Use the dbt Developer Agent (Preview), Snowflake Cortex Analyst, or Cursor/Copilot against a dbt project you know. Generate a transformation or query. Find the mistake it makes. Write three sentences: the question, the wrong output, the lineage assumption that caused it.
  • The artifact: One paragraph. A specific failure mode with a specific cause. Not a framework — an incident. The governance question the interview is asking is: what failure mode have you caught? A framework is not an answer. An incident is.
  • What success looks like: A month from now you have three of these. That's a governance record. The Staff AE who brings a governance record to a technical screen is answering a question most candidates haven't realized is being asked yet.

Interview prep update: practice once with AI tools explicitly on the table. The shift has already happened at Meta, Shopify, Canva, and Rippling. It will spread. If you haven’t practiced a technical screen where AI tools are available, the first attempt will be in an interview.

  • The user moment: This Friday afternoon, or Saturday morning before a week with a loop scheduled.
  • The shape: Skill drill — one mock technical screen run with AI tools in play.
  • The time budget: 60 minutes. Take a published dbt or analytics engineering interview problem. Solve it using Cursor, Copilot, or the dbt Developer Agent. Write one paragraph about a tradeoff in the solution the AI produced. Practice stating that tradeoff out loud.
  • The artifact: One technical problem, solved, with one AI-generated tradeoff articulated in your own words. The candidate who can narrate their reasoning about AI-generated code is demonstrating governance instinct. The interview is beginning to test for that.
  • What success looks like: When the interviewer says "you can use any tools you'd like," you don't pause. You use the right one and explain the output. The pause is obvious in everyone else's answer.

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 me Read more Crafting

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