Autonomous data discovery

Find what your data is hiding, before anyone thinks to ask.

DecisionBox connects read-only to your data warehouse. Its AI agents decide what to investigate, write and run their own SQL, and check every number with a second query before they report it.

  • Read-only: nothing is written or copied
  • Every figure checked twice
  • Self-hosted, with your choice of AI model
DISCOVERY RUN · LIVEhypothesisrepeat bookings fall in one age groupquery 37 of 84SELECT … GROUP BY age_bandresultone band down, the others flatre-queryindependent SQL: figures matchHIGHvalidated insight · confidence 0.86Repeat bookings are falling in one age groupEvery figure checked by a second query. Source SQL attached.Recommended actionpriority high · 4 steps · effort low
Illustration, with an invented finding: the agent's log, then the insight it validated and the action it recommends.

Where it hurts today

The problems DecisionBox is built for

Nobody asked the right question

Important patterns stay in the warehouse because no one thought to look.

A backlog that never clears

Analysts hold a list of exploratory questions they never reach.

AI answers you cannot check

Numbers arrive with no way to see where they came from.

Findings that go nowhere

An insight without a segment, an impact and next steps does not become action.

What it does

Everything DecisionBox covers

Explorewithout being asked
Autonomous discovery runs
The agent forms its own hypotheses, writes the SQL, reads the results and chooses what to look at next.
Schema discovery
Enter credentials and the agent learns the tables by itself. No SDK, migration or pipeline change.
Playbooks
Industry context in plain language: the areas, metrics and segments that matter to you.
Self-healing SQL
A failed query is diagnosed and rewritten for the warehouse in use.
Learning across runs
Each run builds on earlier discoveries and on your feedback.
Verifyevery number
Insight validation
Each figure in a finding is queried again independently. What does not match is corrected or dropped.
Validated insights
A severity, a one-line headline, the rows affected, a confidence score and the exact SQL.
Full transparency
Every query, reasoning step and decision is logged and can be watched live.
Acton what it finds
Recommendations
Each insight comes with an action: priority, target segment, expected impact, effort and steps.
Ask
Plain-language questions about your data and its findings, with the steps shown. Enterprise edition.
Schedules and notifications
Daily, weekly or monthly runs, with results sent to Slack or by email. Enterprise edition.
Discovery ledger
What has been explored, what remains and the status of every finding. Enterprise edition.
Governwho sees what
Read-only access
Only SELECT queries run. The data stays in your warehouse.
Data governance
Sensitive columns are hidden from the AI, blocked in queries and redacted in results. Enterprise edition.
Sign-on and roles
Any OIDC identity provider, three roles, and isolation between organisations. Enterprise edition.
Audit log
Every request tied to a person, searchable and exportable. Enterprise edition.
Choice of AI model
Claude, OpenAI, Gemini, AWS Bedrock, Azure AI Foundry, or a model run locally.

How it works

From first connection to steady state

  1. Connect

    Choose your warehouse and enter read-only credentials.

  2. Describe

    Pick a playbook or write your own in plain language.

  3. Run

    The agent explores unattended, and every query can be watched.

  4. Validate

    Each number is queried again. What fails is corrected or discarded.

  5. Act

    Every insight arrives with a prioritised recommendation.

Each run builds on earlier discoveries and on your feedback.

Where it is used

What teams use DecisionBox for

Banking and wealth

Early warning of client attrition across deposit, advisory and transaction data.

Telecom

Revenue drift, plan moves and churn signals across billing, CRM and network data.

Retail and ecommerce

Channels, products, stores and loyalty read together.

Marketing

Which channels keep customers, and which cohorts leave.

Sales and revenue operations

Account traits that predict expansion, and usage signals that come before non-renewal.

Data teams

The backlog of exploratory questions, worked through overnight.

100%
of findings queried again before they are shown
50 to 100+
SQL queries in a typical run
6
AI model providers to choose from
Open source
core, under AGPL v3

Figures as published by DecisionBox.

Works with

What it connects to

  • Snowflake
  • Google BigQuery
  • Amazon Redshift
  • Databricks
  • PostgreSQL
  • Microsoft SQL Server
  • Oracle Database (Enterprise edition)
  • SAP HANA (Enterprise edition)
  • Slack (Enterprise edition)
  • Claude
  • OpenAI
  • Gemini
  • AWS Bedrock
  • Azure AI Foundry
  • Ollama

Deployment

Where it runs

  • Your own infrastructure
  • Docker Compose
  • Kubernetes with Helm
  • Terraform

Self-hosted. Credentials and data stay in your environment.

With Channel Nexus

Talk to someone who knows DecisionBox

One call, several technologies

On the booking page you choose DecisionBox and any other technology you want to hear about.

A consultant first, the vendor when it helps

A Channel Nexus consultant takes the call and brings in DecisionBox's own specialists for the detail.

A time that suits you

You pick the time. If a consultant cannot make it, we propose other times that suit you.