Your CEO is going to ask AI about your numbers either way, and if nothing defines them, it will guess. I help small data teams build a semantic layer (the agreed definitions behind your numbers) and an AI agent on top of it, so every answer comes from a definition your team agreed on.
First answer by week 2, in a moderated Slack channel with just your data people. By week 8 the agent handles the questions your leadership actually asks, and your team knows how to keep it right.
Built for teams that already run dbt on a warehouse. No warehouse yet? You're a step early.
models: - name: orders semantic_model: enabled: true agg_time_dimension: ordered_at metrics: - name: net_revenue label: Net revenue description: > Gross bookings minus refunds and credits. Excludes tax. type: simple agg: sum expr: gross - refunds - credits filter: | {{ Dimension('order__is_test') }} = false config: meta: owner: finance # agreed in week 1
What was revenue in September?
$1.42M in net revenue. "Revenue" has two agreed definitions here, so I used net_revenue (owner: Finance). Gross bookings were $1.61M. The breakdown by channel is on Revenue overview.
And churn by region?
Nobody has defined churn yet, so I won't guess. I've flagged it for the data team.
Example data. The agent always says which definition it used, and when there isn't one, it says "I don't know."
Want to know where your own metrics stand first? Start with a free review.
If three of these sound like a normal Tuesday, let's talk.
The models got good and cheap, so insight is cheap now too, as long as your definitions are ready for it. Most teams know they should get there, but nobody has the time to lead it, so it slips behind whatever is on fire this week.
Hiring me is the forcing function. For eight weeks you get a champion who has done this before, and your team picks up the skills along the way. Once it's live, people get a quick answer in Slack instead of hunting for the right dashboard.
Your data team stops being the search engine and gets back to the analysis that changes decisions.
Your team writes the questions and the answers they expect, and we run them as an eval: a weekly test against those answers to measure accuracy and drift. By week 8 the agent gets at least 90% right, and the rest are "I don't know," never a guess. If it falls short, I keep working at no cost until it does. The full terms come with the quote.
$24,000 for the standard build, fixed price and fixed scope. That's about two months of one analytics engineer's salary. I build it, and your team's part is mostly the first two weeks.
The first two builds run at a founding rate, in exchange for a written case study. Ask about it in the review.
After handover, a 30-day check-in is included. If you want a hand after that, there's an optional monthly plan with me as an advisor: we review the eval results together, and I help your team work out what slipped, how to fix it, and what to add next.
You'll need dbt, Core or Cloud, on Snowflake, BigQuery, Databricks, Redshift, Postgres or Trino, and someone who keeps it running. No warehouse yet? You're a step early: get your data into one first, then come back.
Your BI tools stay. The agent runs on your own AI provider or inside your cloud, with read-only access, so the model sees answers, not your raw tables.
I'm Kevin Styers. Seven years in data, from environmental sensors to telecom networks to leading an analytics team (I started in field ecology, so I came at this sideways).
Every one of those jobs had the same conflict, two people with two numbers for the same thing, and a new dashboard never once settled it. So I built a Slack agent for it, and the agent was the easy part. The real work was everything under it: agreeing on the definitions in dbt, grading its answers against real questions until they held up, and setting the rules for what it's allowed to answer at all. It's the only kind of agent I'd put in front of a CEO.
For teams on dbt. We spend 90 minutes on your project. You leave with a written list of the metrics most likely to disagree, and the first one I'd define. It's yours whether we work together or not.
Reviews are open now. The two founding builds start in Q1 2027.
Book a free review or email me with your data problems