Skip to main content
This page explains how Interact handles data in practice. It is an overview for administrators and users — not a legal contract. Binding terms are in Interact; see Legal documents.

​
How Interact uses your data

Interact connects to the data warehouse and platforms you configure, runs queries and analyses, and returns answers in chat, reports, jobs, and monitors. In that flow:
  • You decide which systems to connect, which semantic models agents may use, and who in your organisation can access which agents.
  • Turntwo Interact processes that data as a processor under the DPA when it is Subscriber Data (for example chats, query results shown in Interact, and reports).
  • AI model providers receive the prompts and context needed to answer a question for the models enabled on your agents — nothing more than that request requires.
Interact is built for marketing and business analytics. Do not bring special-category personal data into Interact, and keep connected data limited to what your analysis needs.

​
Data access & connectors

Data Analyst agents reach your warehouse only through a connector you assign per agent, and only through the semantic models you select for that agent. That combination is the main access-control model in Interact.
Service account credentials are encrypted with Google Cloud KMS before storage. They are decrypted only at query time and are never shown to AI models.
Each agent selects its own warehouse connection. You can create multiple connectors with different credentials and IAM scopes (for example a marketing-only service account vs a broader analytics account) and attach the right one to each agent.
For every agent you select which semantic models it may use. Each model maps to a warehouse source table. The agent queries through those models; if raw SQL is enabled, it may only read the source tables of the selected models — not arbitrary tables in the project.
Provision each warehouse service account with the minimum access required — typically read-only roles on the datasets that agent’s models need. Prefer separate accounts when agents serve different sensitivity levels.
Connectors run read-only SELECT queries. Interact does not write to, modify, or delete data in your warehouse.

​
Query & LLM limits

These are separate controls: one protects your warehouse bill (bytes scanned), the other caps AI spend in Interact.

​
Warehouse query limits

On each warehouse connector you can set a maximum billable data volume per query. For BigQuery this is configured in megabytes (default 1000 MB ≈ 1 GB) and passed through as maximumBytesBilled. If a query would scan more than that limit, the warehouse rejects it — so a single agent query cannot run an unbounded scan against your project.

​
LLM cost caps

AI usage is measured in Credits (based on model-provider cost). Administrators can cap spend at several levels: Which member limit applies: personal override → otherwise the strictest (lowest) group limit among the member’s groups → otherwise the organisation default → otherwise no member limit. Organisation credit limits and member LLM caps are independent of the warehouse per-query byte limit.

​
Chat & AI processing

​
Chat history

Interact stores chat history so users can continue conversations and review past answers.

​
AI models

Agents use Anthropic Claude by default. You can choose a different model per agent (for example Google Gemini or OpenAI GPT). Only the providers for the models you use receive that request’s prompts and data. AI model providers are often located outside the EEA (typically the United States). Transfers use an adequacy decision (such as the EU–US Data Privacy Framework where the provider is certified) or the European Commission’s Standard Contractual Clauses. Query results and chat content are used to answer your questions in Interact. They are not used to train Interact’s own models. If you select another model, that provider’s terms apply. Provider policies can change; material updates are reflected in the documents listed under Legal documents.

​
Guardrails

You can add guardrails on an agent — for example PII detection and redaction on user messages, or checks on agent output — to reduce the chance that sensitive values reach a model or a response.

​
Access control & governance

​
Hosting & data residency

Interact’s application and primary data stores run in the European Union (for example application hosting and database storage in Frankfurt). Some subprocessors — especially AI model providers and certain background services — process data in or from the United States under the safeguards described above. Services you connect yourself (your warehouse, Slack, Teams, sign-in providers, MCP servers, and similar) are under your control and are not Turntwo subprocessors.

​
Security practices

Turntwo maintains an information security management system aligned with ISO/IEC 27001. Measures include:
  • Encryption in transit (TLS) and at rest
  • Encrypted storage of connector credentials via a managed key management service
  • Role-based access and organisation isolation
  • Logging, error monitoring, and incident alerting
  • Regular backups with point-in-time recovery of the application database
  • Agent actions on external platforms only after explicit user approval
  • Staff confidentiality and least-privilege access to production systems
Full detail is in Annex 2 of the Data Processing Agreement.
  • Create separate connectors (different service accounts) when agents need different warehouse privileges, and assign them per agent.
  • On each agent, allow only the semantic models (and therefore source tables) that agent should see.
  • Prefer views or column masking in the warehouse when PII must stay out of those source tables. For example, hash customer_id values so agents can still join and count customers without seeing the original IDs.
  • Set a sensible max billable data volume per query on each warehouse connector.
  • Set organisation credit limits and member / group LLM caps where you need spend control.
  • Enable 2FA for password logins (and SSO where available) for your organisation.
  • Use guardrails on agents that handle broader or more sensitive questions.
  • Review member, group, and agent access periodically.
  • Keep data that must never reach an AI interface out of connected datasets entirely.
Questions about privacy or security: info@turntwo.com.