The AI calculates
You paste raw data into a language model and ask for metrics. The model guesses the arithmetic and often gets it wrong.
/0.4 · Ask
Work with your data in Claude, ChatGPT or your own agents. Ask questions, run analysis and build dashboards in Dataloader or as Claude artifacts, all from the same model, through the Dataloader MCP server and API.
What it is
People use dashboards. Claude, ChatGPT and your own agents use the MCP server. Both read the same model, so they give the same numbers.
Answers come with their source, ingest run and time.
KPIs without enough connected data are rated low confidence or not shown, instead of guessed.
The difference
You paste raw data into a language model and ask for metrics. The model guesses the arithmetic and often gets it wrong.
The language model writes SQL or Python to query your database. Different prompts produce different logic, risking wrong joins or double counts.
Defined KPIs are calculated on linked data by the Dataloader server. The AI reads the result with its source, so you get the same answer every time.
Language models are good at language, not arithmetic.
Capabilities
Ask for a dashboard in plain language. Publish it in Dataloader for your team, or keep it as a Claude artifact you can share. Both read the same model.
Claude, ChatGPT and your own agents read the Observatory model through MCP.
Values come back with their source, ingest run and time, so an answer can be checked.
Language models run with providers in the EU or EU-hosted regions, receive only redacted, pseudonymised data, and are contractually barred from training on your data. Generated values are tagged so they stay distinguishable from observed ones.
In practice
Illustrative example
ROAS × margin
Illustrative example
Illustrative example
Unit A: 31%
Illustrative example
Customer result
+0.5 pp
Owner-reported customer result
How it works
Point Claude, ChatGPT or your own agent to the Dataloader MCP endpoint.
Ask in plain language. The assistant reads the Observatory model.
The Observatory returns values with their source, ingest run and time.
Technical notes
Questions
Language models cannot reliably perform arithmetic on large datasets. With Dataloader, metrics are calculated on linked data by the server. The AI simply reads the result with its source, so you get the same answer every time, with its source.
Yes. Ask Claude, ChatGPT or your own agent for a dashboard in plain language. Publish it in Dataloader for your team, or keep it as a Claude artifact. Every number comes from the same model, with its source.
Claude, ChatGPT and your own agents, through the MCP server and the API.
No. Language-model providers used by Dataloader are contractually barred from training on your data.
If you want them. Dashboards and AI assistants read the same model, so they give the same numbers.
Every value in the model carries its provenance: source, ingest run, time, and who produced it. Answers return it with the value.
Language models run with providers in the EU or EU-hosted regions, receive only redacted, pseudonymised data, and are contractually barred from training on your data. Generated values are tagged so they stay distinguishable from observed ones.
Related solutions