Platform · Data agents
A chatbot answers. An analyst shows their working. DataFact agents are built to do the second thing: plan the investigation, cross-check what they find, and hand back something you could defend in a review.
Anatomy of a run
Every run is inspectable: the sub-questions it formed, which channel each step went to, how long it took, and the point where it reconciled two sources that disagreed.
When someone challenges the conclusion, the argument is about the evidence rather than about whether the machine can be trusted.
Capabilities
These behaviours are the difference between something people try once and something they route real work through.
Planning
A causal question cannot be answered by one retrieval. The agent breaks it into sub-questions, and what it finds in the first step changes where it looks in the second — rather than following a chain fixed before it knew anything.
Corroboration
When the report says one thing and the table says another, averaging them is the wrong answer. The run surfaces the disagreement, states which source it trusted, and says why.
Clarification
Half the questions people type are underspecified — which region, which fiscal calendar, gross or net. Guessing produces a confident wrong answer. The agent asks, with the options it can see, then continues from your reply.
Output
What comes out has data, charts, a written conclusion and a citation trail — durable enough to paste into a decision memo, and re-runnable next month against fresh data.
Delivery
An agent is published to the surfaces your people already use: the console, an embedded widget in your own application, an API for your systems, or a bot in the chat tool where the question would have been asked anyway.
Agent types
Different questions want different machinery. An agent is set up with the resources it needs and nothing more, which keeps both its answers and its permissions tight.
Knowledge
Bound to selected document collections. Returns text, tables and figures with page-level citations.
Analytical
Bound to a schema and an analysis model. Produces governed SQL, results and charts.
Investigative
Bound to both channels. Plans, splits and corroborates before writing a conclusion.
Assistive
Prompt and skills only, for drafting and reasoning tasks that should not touch company data.
The recurring investigation that takes an analyst two days. We will build the agent for it during the demo and show you the run.
SaaS subscription or private deployment · No data migration either way