Numbers come from computation, language comes from models

Put a general-purpose model in front of a dashboard and it will hand you a number that sounds right. Our division of labor: statistics compute, the agent reasons, and every figure must pass an evidence-grounding check.

The most common failure of "AI analytics" isn't a wrong answer — it's a fluent one. The model returns a well-structured, confident conclusion with a number precise to one decimal place, and the number is invented. For growth decisions, that is far more dangerous than no answer at all.

A strict division of labor

• Statistical detectors own every number: sweeps, aggregates, significance, threshold calls — deterministic computation, re-checkable, each snapshot fingerprinted.
• The language model owns explanation and orchestration: form hypotheses, assemble the evidence chain, write why it matters, decide what to investigate next.
• Between them sits structured output: every figure the agent cites carries an evidence ID pointing at the source row.

The grounding check

After a report is generated, one more machine pass compares every figure in the text against the evidence ID it claims. Citations that don't reconcile are stripped — not greyed out, removed. The cost is that reports are sometimes shorter. The benefit is that you can click any number and see where it came from.

When it isn't decidable, say so

Evidence chains run three ways: internal data, external market, and the two combined. When none of the three settles it, the report says "not decidable" and lists what data is missing, instead of picking whichever explanation sounds best. A system that can say "I don't know" is the only kind worth believing when it says "I do."