Skip to content

Reporting and decision support

Numbers everyone agrees on, delivered where people work.

Leaders wait days for reports and argue about whose number is right. We define each metric once, deliver reports where people already work and test AI analysts against verified answers before anyone relies on them.

Who it's for

Leaders who wait days for reports, argue about whose number is right or want to ask questions of their data in plain English.

What we deliver

A metric dictionary

Each metric gets a definition, an owner and its edge cases. Take "average order value": it has at least three valid answers, depending on whether it includes freight and canceled orders and whether it's per order or per item.

A semantic layer

So every tool uses the same definitions: dbt MetricFlow, Snowflake semantic views or Databricks metric views.

Automated reports and dashboards

Delivered where people already work: email, Slack and Sheets.

AI analysts that are tested before anyone trusts them

Genie, Cortex Analyst or Claude connected over MCP, set up on the semantic layer and benchmarked with a golden set of business questions and verified answers. Two scores are reported separately: "the SQL ran" and "the answer was right". Remaining errors are sorted into six categories (schema understanding, joins and aggregation, metric ambiguity, filters, time logic and missing business context), so it's clear which ones better data modeling can fix.

Why this matters: on the Spider 2.0 enterprise text-to-SQL benchmark, models that score about 90% on simple schemas drop to about 10 to 20% on enterprise-scale databases (Spider 2.0 paper). It shows why data modeling and definitions matter. It doesn't mean any specific fix gives a specific gain.

AI on text

Turning reviews, tickets or survey comments into reportable categories, checked against a hand-labeled sample with a confusion matrix and per-category precision and recall. Personal information is masked before text reaches a model.

Activation

Sending lists back into business tools (reverse ETL), each with an owner, a refresh schedule and a plan for what happens if the data is wrong.

How we measure it

  • Time from question to answer.
  • Share of golden questions the AI analyst gets right.
  • Classification precision and recall.
  • How many people actually use the reports.

Background

These methods come from And's course plan for ETM 538, Decision Support Systems: Data Warehousing, at Portland State University (Winter 2027).

Questions

Common questions

How do you know the AI is working?
Each project gets a test set from your real work and agreed metrics (for example, answer correctness for an assistant or hours saved for an automation), measured before and after every change.
We tried AI before and it didn't work. Why would this be different?
We start with a baseline and a test set from your real work, so you can see whether it works before you rely on it. If it doesn't clear the bar you set, you'll know early.
Can you work with our existing data warehouse?
Yes: Databricks, Snowflake, BigQuery, Redshift, Postgres and others. The modeling and testing methods don't depend on the vendor.

Have a process that eats hours every week? Let's look at it together.