AI Policy

Playfair is analytics for people who don’t write SQL, built to be trusted by people who do. This policy explains how we use artificial intelligence to do that, and the limits we set for ourselves.

1. What the AI does

AI features help at specific steps of answering a question:

• Interpretation — turning “monthly revenue this year” into a structured plan: the measure, filters, time grain and assumptions.
• SQL generation — writing a read-only query for your database dialect, using your metric definitions, joins and semantic rules.
• Chart selection — choosing the form that fits the question (line for time, horizontal bars for categories) following documented visualisation rules.
• Narrative — a short headline and caveats in plain language.
• Descriptions — suggesting descriptions for tables and columns, which you can edit.
• Driver analysis — comparing periods across dimensions to rank what changed and why.

When no AI provider is configured, Playfair uses a deterministic built-in engine for the same steps.

2. Show the work: a human stays in charge

Every answer shows the interpretation, the exact SQL that ran, the rows behind the number, a confidence level with its reasons, and caveats such as excluded rows. When a question is ambiguous, Playfair asks rather than guessing silently. Analysts can edit the SQL, and anyone can mark an answer as wrong; accepted corrections can become rules that improve future answers in your workspace only.

AI output can be wrong. Review answers before relying on them, especially for significant decisions, and never use the Service to make decisions with legal or similarly significant effects on people without meaningful human review.

3. What data the AI sees

We send the minimum needed for each step:

• your question and the recent context of the thread;
• the relevant part of your schema: table and column names, types, descriptions, synonyms, metric definitions and joins;
• for descriptions, a handful of sample values — values from columns identified as personal data are redacted;
• for narratives and driver analysis, aggregated results rather than raw rows wherever possible.

Queries always run against your database from our servers, never from the AI provider, and always read-only.

4. No training on your data

We do not use your data, your questions or your answers to train AI models — ours or anyone else’s. Our AI providers process requests on our behalf under contracts that prohibit training on customer data and restrict retention to short-term abuse monitoring. Improvements learned from your corrections stay inside your workspace as semantic rules you can see and delete.

5. Providers

Our current AI provider is listed on the Subprocessors page. We may route different steps to different models (for example, a smaller model for interpretation) to balance quality, speed and cost; the same commitments apply to every model.

6. Your controls

• Admins can switch off AI-generated descriptions for a source and write their own.
• Analysts can edit generated SQL and save governed metrics that constrain generation.
• Workspace Owners can ask us to disable AI features entirely for their workspace; the built-in engine is then used.

7. Transparency and regulation

AI-generated text in Playfair is labelled as such, and every chart and number can be traced back to its query. We monitor developments under the EU Artificial Intelligence Act and will update this policy as obligations apply. Playfair is not intended for any use classified as high-risk under that Act.

8. Contact

Questions or concerns about our use of AI: legal@asrar.example. Security issues: security@asrar.example.