Validated InsightsNot a chatbot answer.
Not a chatbot answer.
A finding you can ship.
DecisionBox insights are structured, severity-ranked, and verified against your actual warehouse before they reach you. Each one is something an analyst could take to a stakeholder as-is.
Before vs after
The difference shows up immediately
A chatbot answer
"Looking at your data, it seems like guest ratings might affect whether guests come back. You may want to look into why some guests don't rebook. Would you like me to explore this further?"
A DecisionBox insight
Guests with low-rated stays repeat-book at 7.35% — a third of the rate for guests with good experiences.
7.35% repeat bookings in low-rated cohort — vs 25.6% for 3–4 star guests
Verified against warehouse
Anatomy of an insight
What's inside every finding
01SeverityRanked critical / high / medium / low so you know what to tackle first.
02HeadlineOne sentence, no hedging. The finding, stated plainly.
03Affected countExactly how many users, listings, or rows this finding touches.
01 Critical severity
02Guests with low-rated stays repeat-book at 7.35% — a third of the rate for guests with good experiences.
5,969 guests whose average review score is 2 or below are coming back at just 7.35%. Guests with 3–4 star experiences repeat at 25.6% — a 3.5× gap driven almost entirely by first-stay satisfaction.
05SELECT avg_rating_bucket, COUNT(DISTINCT guest_id), AVG(repeat_rate)
FROM guest_cohorts WHERE avg_rating_bucket IN ('1-2 star', '3-4 star');
FROM guest_cohorts WHERE avg_rating_bucket IN ('1-2 star', '3-4 star');
04ConfidenceCalibrated confidence score based on sample size and effect magnitude.
05Source SQLThe exact query that produced the finding — copyable, auditable, runnable.
06EvidenceThe supporting numbers, already reconciled against an independent verification query.
Every number is verified before it reaches you.
After the agent forms a finding, DecisionBox re-queries the warehouse with an independent verification query. Numbers that reconcile are kept, numbers that don't are adjusted or discarded. What you see is what your data actually says — no hallucinated statistics, no numbers pulled from context.
app.decisionbox.io/insights
Keep reading
Try it in two minutes
Clone the repo, run docker compose up, and point it at your warehouse.