We made a video series for the people who act on DecisionBox findings rather than set it up. 7 short episodes are live, all on a public demo dataset. Start with episode 1: how far to trust a finding, and where to see the queries behind its numbers.
Each step of a DecisionBox run can now use its own model. In one comparison in our docs, mixing models cut a run's cost from $13.46 to about $10.10 and its time from 46 to about 26 minutes, with no loss in finding quality.
4
DecisionBox Enterprise can run discovery on a schedule: daily, weekly, monthly or your own cron. After each run it writes an executive summary where every figure links to the finding behind it, and sends it to the right people by email or Slack.
An AI finding earns trust when it can still explain itself later. In the Discovery Ledger, each finding links to the SQL and rows behind it, and every status change keeps its reason and the run that made it.
6
New series: DecisionBox for business, presented by Jale, an AI host. Episode 1 shows two agents checking a finding against the data. One finding held up. Another said the most expensive price band was 24% of sales. The real share was 15.2%, so it's marked Rejected.
1
11
Each step in DecisionBox now takes its output limit from its own model (v0.37.2). A step on a smaller model is no longer asked for more than it can return, and one on a larger model is no longer capped short. Reasoning effort is now set explicitly, not left to the provider.
9
Prompt caching is on by default in DecisionBox for Claude models on Anthropic, Bedrock, Vertex AI or Azure. Each exploration call reuses the cached start of the previous prompt. On a production-scale ERP warehouse, it cut the whole run's cost by about 75%.
16
DecisionBox asks your data the questions nobody thought to ask. Dashboards and AI assistants answer the questions someone already has. A database with 200 tables has 19,900 possible pairs of tables, too many for any team to search by hand.
With validation on, the default, each DecisionBox finding goes through two checks. We followed one real finding through both, on a public e-commerce dataset.
1
7
Validation also catches slips. On a public ticket-sales dataset, a DecisionBox headline said the top price band was 24% of sales. The rows said 15.2%. Both agents rejected it.
1
Watch DecisionBox catch its own mistake. After 4 queries, it sees that a join has inflated its sales numbers, so it rewrites the query to total sales separately. Later the data contradicts its hypothesis and it changes its mind. Both corrections stay on the record.
Every finding from an AI analyst deserves a check before anyone acts on it. A new public benchmark grades data agents by the consequences of their actions, across 210 tasks on a 235-table warehouse. The strongest of 14 models scored 95 or more on 34.8% of them.
5
Each step in DecisionBox can now use its own model (v0.37). For a project you already have, open its settings once and save the model setup, which comes pre-filled with what it uses today. Until that save, discovery and scheduled runs for it wait.
8
Tip for DecisionBox v0.37, in the open-source Community edition: you can give each step its own model. Exploration makes many calls, so a cheaper model fits there. Give the two validation agents your strongest model. Set it under Settings → Models.
Since April we've shipped 39 open-source releases of DecisionBox, v0.5 to v0.37. The biggest changes are in this thread, and the full write-up is linked in the last reply.
1
1
10
DecisionBox v0.36 added prompt caching. With the same model, settings and day, a full run on Claude cost 76% less. Exploration cost fell by 95%.
1
3








