PYX data product agentic engineering
Field notes

When AI does BI

Short recipes from someone who spent fifteen years arguing about grain in meeting rooms, and has finally found something a language model is genuinely good at. Not writing the SQL. Reading the ambiguity, at a volume no human ever could, and handing a person a decision worth making.

The premise
Most of an analytics team is a translation layer.

Count the roles whose real output is somebody else’s meaning restated in a different notation. A spec into SQL. SQL into DAX. DAX into a slide. It’s most of them, and every translation is a chance to lose the meaning, with nothing left behind that says what was decided, or who decided it.

The wager
The translation work goes. The judgement work grows.

A model can list every ambiguity in a requirement and tell you what each one implies. It can’t decide that Finance counts at contract account because that’s how the invoice runs. So give it the enumeration. Keep the judgement. That trade has a name now, and it makes you an agentic data analyst: someone who directs the work and rules on it, rather than typing it.

Field notes

Things to do differently on Monday

Each one is a habit, not a feature: what the old loop was, what changed, and how to run the new one without pretending the AI has judgement it doesn’t have.

03Discovery

Your semantics are already written. They are in a spreadsheet.

The most complete model of your business isn’t in your warehouse. It’s in a workbook your finance lead has kept for nine years. Stop calling it shadow IT.

6 min·Read →
04Design plane

Show the numbers before you have the tables

Six weeks of platform work to find out the requirement was wrong. There’s a way to find that out on day one.

6 min·Read →
05Control

Bind the control before the first row lands

Data quality arrives last because it’s scoped last. Move it to the moment the meaning gets decided and it stops being a project.

6 min·Read →
06Migration

The migration that passes every test and changes every number

Green pipeline, different answer. Both true. Nothing in a normal stack is built to notice.

6 min·Read →
07Discipline

Vibe analytics is what your Genie pilot actually was

Ask a question, get a number, accept it because it looks about right. Software named this two years ago and built a discipline to escape it. We haven’t.

6 min·Read →
08Evaluation

Your claims are already an eval suite. Stop writing them twice.

The best way to tell an agent what “correct” means isn’t a better prompt. It’s the set of checks that have to hold, which you already wrote, in week one.

6 min·Read →
09Trajectory

Check the path, not just the number

A figure can be right this month and arrive by a route that’s wrong next month. Nothing in your stack is looking at the route.

7 min·Read →
10Harness

You don’t have a model problem. You have a harness problem.

The model is about a tenth of a working agent. The other nine tenths is the thing your team has to build, and in data, almost nobody is building it.

6 min·Read →
From reading to doing

Every one of these is testable on one of your own requirements.

Tell us what you're trying to build or improve. We can talk through whether PYX is a fit.