Skip to main content
AI Discovery Agents

They don't wait
to be asked.

DecisionBox agents are not chatbots. They form hypotheses, write SQL, read the results, and decide what to investigate next — autonomously, end to end.

Exploreform hypothesesAnalyzerun SQL, read resultsDecidewhat to investigate next
Inside a real run

Watch the agent think

Discovery run
76 min · showing 7 highlights from 74 steps
Complete
Exploration
Investigating booking cancellation patterns across 2025
14:02
Query
Aggregated cancellation rate by month across 47,618 bookings
14:09
Finding
Overall cancellation rate is 21.15%, but it spikes to 28.8% in July — the highest month of the year
14:17
Exploration
Is the July spike concentrated in a specific booking pattern?
14:28
Query
Joined cancellations with booking lead time for July
14:41
Finding
Same-day bookings made up 26.4% of July volume, cancelling at 40.8% — nearly double the baseline
14:54
Validated insight
The July cancellation spike is driven entirely by a surge in same-day bookings, which cancel at 1.9× the platform baseline
15:16

Example trace from a vacation rental warehouse.

Built for discovery, not conversation

Loop, not one-shot

Runs form hypotheses, test them, and branch into follow-ups. Like watching a good analyst work.

Fully transparent

Every SQL query, every reasoning step, every decision — logged and visible in real time.

Read-only, always

Agents never write back to your warehouse. Read access is the only permission needed.

Self-healing SQL

Failed queries get diagnosed and rewritten against warehouse-specific error patterns — no human in the loop.

A typical run writes 50–100+ queries

The agent explores until it runs out of productive leads. Runs typically take anywhere from 30 minutes to a couple of hours, fully unattended, and produce a set of validated findings — not chat transcripts, not dashboards to interpret. You get to the point.

app.decisionbox.io/discoveries/142
AI agent run view
Keep reading
Validated Insights
What the agent produces at the end of a run.
Recommendations
From finding to action in a single handoff.
How It Works
Where the agent fits in the end-to-end flow.

Try it in two minutes

Clone the repo, run docker compose up, and point it at your warehouse.