M↓ View as Markdown ↗

How AI Analyst handles your data¶

This page follows your data from the moment someone asks a question to the answer on screen. It is written for IT, security and governance teams evaluating the product.

The page describes Actian SaaS, the service Actian hosts. If your organization runs a self-managed deployment, also read Self-managed deployments below: where data lives and which AI models see it are different there.

Info

In short: AI Analyst queries your database with the read-only warehouse user you provide. It does not copy your database or tables. The query results it needs to answer a question are kept with that conversation and with any tables, charts or reports it produces. In Actian SaaS, Actian's AI providers process data under data processing agreements and do not use it to train models.

Core guarantees¶

Encrypted connections and credentials: Connections to your data source support TLS. You choose the SSL mode on each connection; we recommend Verify full where your database supports it, because some connectors (for example MySQL and ClickHouse) default to no TLS. Data source credentials and AI endpoint API keys are encrypted at rest with AES-256-GCM. In Actian SaaS, stored data is also encrypted at rest by the hosting platform. In a self-managed deployment, your operator is responsible for at-rest encryption of the database and storage.

Read-only access: AI Analyst only issues read queries. On PostgreSQL, MySQL, MariaDB, Amazon Redshift, MotherDuck and ClickHouse, read-only mode is also enforced by the database session. Other warehouses (for example Snowflake, BigQuery, Databricks, Microsoft SQL Server, Microsoft Fabric and Azure Synapse) cannot enforce it from the connection, so connect with a warehouse user that has read-only privileges. Studio warns you when this applies. Every query is also limited in rows and run time.

No bulk copies: AI Analyst never replicates your database or tables. It retrieves only the rows a question needs, up to a row limit per query. Those results are kept with the conversation and its artifacts, so the conversation can continue and artifacts can be reopened. CSV files you upload, and sample values saved in your semantic layer (for example, a dimension's listed values), are also stored by AI Analyst.

Not used for AI training: For Actian Inference, Actian's AI providers process data under data processing agreements and do not use it to train models. If your organization uses its own LLM or Embedding endpoint, or runs a self-managed deployment, your agreement with that provider governs how it handles data. See AI providers.

How your questions are answered¶

AI Analyst does not hand your database to an AI model. Each question follows a traceable process grounded in your semantic layer.

sequenceDiagram
    participant U as User
    participant AA as AI Analyst
    participant SL as Semantic Layer
    participant DB as Your Database
    participant LLM as AI provider (Actian Inference or your own endpoint)

    U->>AA: Asks a question in natural language
    AA->>SL: Looks up relevant metrics, models and business definitions
    AA->>LLM: Sends question + semantic layer definitions
    LLM-->>AA: Writes a SemQL query
    AA->>DB: Runs the compiled SQL (read-only)
    DB-->>AA: Returns the query result (specific rows only)
    AA->>LLM: Sends result rows for summarisation
    LLM-->>AA: Produces a natural-language answer
    AA-->>U: Displays the answer, with the query behind it

In practice:

The AI provider never has direct access to your database.
Your database credentials are stored encrypted (AES-256-GCM, with keys held outside the database). They are never shown again after saving and never sent to the AI provider.
The AI provider receives the user's question, the conversation so far, the relevant semantic layer definitions (names, descriptions, relationships and any listed dimension values), and the rows returned by the queries it runs. When the data team generates a model with AI in Studio, a small sample of rows (up to 20) and distinct column values from the source table are also sent.
Analyst queries are written in SemQL, a semantic query language, and compiled to SQL against your semantic layer. If your warehouse rejects the compiled SQL for a dialect reason, AI Analyst may ask the model to repair it once. The repaired query runs only if it reads nothing beyond the original query and keeps all of its filters.
Query results are written into the conversation's history, so follow-up questions can use them. They are also kept in any artifact (table, chart or report) created from them.

How Steward works with your data¶

Steward is the AI agent that helps data teams build and maintain the semantic layer. It interacts with your data more broadly than the AI Analyst agents, because building a trustworthy semantic layer requires understanding the real data, not just its structure.

sequenceDiagram
    participant DT as Data Team
    participant ST as Steward
    participant SL as Semantic Layer
    participant DB as Your Database
    participant LLM as AI provider (Actian Inference or your own endpoint)

    DT->>ST: Asks Steward to build or improve a model
    ST->>SL: Reads existing model and metric definitions
    ST->>DB: Reads table structure (column names and types)
    ST->>DT: Asks for approval before running a SQL query (by default)
    ST->>DB: Runs read-only SQL to inspect or validate real data
    ST->>LLM: Sends schema metadata, query results and semantic layer context
    LLM-->>ST: Proposes changes to the semantic layer
    ST-->>DT: Shows the proposed edits for review
    DT->>ST: Approves (if approval is required)
    ST->>SL: Applies changes to the semantic layer

What Steward accesses:

The structure of your tables (column names, data types and relationships) to understand what can be modelled.
Your existing semantic layer definitions, so new suggestions stay consistent.
Actual row-level data, when it runs a read-only SQL query to explore a table or validate a proposed model or measure. This can include real values from your tables, including personal data if present, and those values are sent to the AI provider as part of that step.
Files you attach to a Steward message. They are stored by AI Analyst, and what Steward reads from them is sent to the AI provider.
The web, if you let it use Search the web. The search query is sent to a third-party web search service. This tool is not available on air-gapped deployments.

What Steward does not do:

Write, update or delete data in your database. It only issues read queries; see Read-only access for how this is enforced per warehouse.
Keep exploration results outside the conversation. Results of Steward's SQL queries (up to 1,000 rows and 50 columns each) are kept in that Steward conversation so it can refer back to them. They are not added to your semantic layer unless you approve a change.

Who can use Steward, and with which database permissions: Steward is available to members with the Edit capability in a Workspace. It queries your data source with that data connection's credentials, not each user's own database permissions. By default Steward asks for approval before each SQL query it runs: Run SQL query is set to Ask first under Steward's Tool permissions. Members with Edit can change this setting, including to Always deny or Always allow, and individual sessions can override it, so treat it as a workspace preference rather than a security boundary.

To guarantee that sensitive columns are never seen, exclude or mask them upstream (for example with a scrubbed database view) and connect with a warehouse user that cannot read them.

Info

Unlike the AI Analyst agents, which work from queries compiled from your semantic layer, Steward can run its own exploratory SQL against your data source. Treat Steward access the same way you would treat direct, read-only database access for a data team member.

Workspaces keep data apart¶

Each organization can have several Workspaces. A Workspace holds its own semantic layer, AI Analysts, conversations, artifacts, and Steward memory and tool permissions. It is a hard boundary in both directions: nothing in one Workspace is readable from another. Data connections are attached to individual Workspaces, and a Workspace can own a private connection that other Workspaces never see. See Workspaces.

What is stored and where¶

In Actian SaaS, the PostgreSQL database that stores AI Analyst data is hosted on Google Cloud Platform in the European Union. In a self-managed deployment, data lives in your own PostgreSQL database and storage; see Self-managed deployments.

What we store:

Conversation history: the questions people ask and the answers AI Analyst provides.
Query results returned during a conversation, kept in the conversation's history so follow-up questions can use them.
Artifacts: the tables, charts and reports created from results, including a snapshot of their data.
Files you upload: CSV data files (stored and converted so they can be queried) and files attached to Steward messages.
Your semantic layer: model definitions, metrics, dimensions, filters, business logic and glossary terms, including sample and listed values saved with it (for example, a dimension's values).
A search index of your semantic layer, used to find relevant definitions.
Steward memory and AI Analyst configuration: instructions, access settings and saved prompts.
Ratings and comments people leave on answers.
Audit logs: a record of who did what and when, for your compliance needs.

What we do not store:

Copies of your database or tables. AI Analyst stores only the query results, values and files described above.

How artifacts use stored data: when you open a chart or table artifact that was built from a single re-runnable query, AI Analyst runs that query again against your current semantic layer and data. If the query cannot run, the artifact shows its stored snapshot and marks it as out of date. Artifacts without a re-runnable query, such as reports, always show the stored snapshot.

Diagnostic traces¶

Where diagnostic tracing is enabled, AI Analyst records traces of agent activity (prompts, model responses and tool calls) so engineering can debug issues. In Actian SaaS, traces are kept in Langfuse.

Before a trace leaves AI Analyst:

Warehouse result rows and sample values returned by the agent's tools are replaced by a placeholder; fields such as row counts and column names can remain.
Email addresses, phone numbers, US Social Security numbers, credit card numbers and IBANs are masked in the remaining text.

This masking is pattern-based. Values that appear elsewhere (for example in the user's question, in figures quoted in an answer, in values written into generated SQL text, or in listed dimension values sent to the model) are not removed, and other personal data such as names and addresses is not masked.

When tracing to Langfuse is enabled, ratings and their optional comments are also sent to Langfuse. Comments are sent as written, without masking.

Traces are used only by engineering and are never used to train models. Traces kept in Langfuse are deleted when your organization is deleted. Tracing is off in the configurations shipped for self-managed deployments; on an air-gapped deployment, vendor-hosted tracing services are not allowed. Your operator chooses where traces go and is responsible for keeping them on your network.

AI providers¶

The Inference page (Settings → Administration → Inference) shows which language model endpoint (Chat) and embedding endpoint (Search) your organization uses, and where each comes from. Studio members and members with the Manage inference capability can open it.

Source shown	What it means
Actian Inference	Actian-brokered access to hosted models from third-party AI providers. This is the default in Actian SaaS.
Provided by your administrator	An endpoint set for the whole deployment, for example by the operator of a self-managed deployment.
Your organization's own endpoint	An OpenAI-compatible endpoint your organization configured.
None	No endpoint. Without a Chat endpoint, conversations are blocked; without a Search endpoint, search matches keywords only.

Actian Inference¶

By default in Actian SaaS, AI Analyst uses Actian Inference: Actian-brokered access to hosted models from third-party AI providers. The product does not show which vendor or model served a given answer.

Actian Inference is never available on air-gapped deployments.

Your organization's own endpoints¶

Availability

Who can do this: members with the Manage inference capability for the organization.
Plan: Included only in plans that include customer-hosted inference.
Deployment: Actian SaaS and self-managed deployments.

Your organization can send chat and search traffic to its own OpenAI-compatible LLM endpoint and Embedding endpoint instead of the deployment's default.

Chat: once your organization uses its own LLM endpoint, every AI call made for your organization (questions, semantic layer context, query results, Steward's schema exploration, conversation titles, suggestions and reading PDF attachments) goes to that endpoint. Requests never fall back to Actian Inference, even if your endpoint fails.
Search: an Embedding endpoint receives only semantic-layer metadata and search queries. For models and metrics this is the name, description and agent guidance; for dimensions, measures and filters, the name, description and parent model name; for glossary terms, the term, definition and synonyms. If your Embedding endpoint fails, search falls back to keyword matching, not to the deployment's endpoint.
API keys are write-only: after saving, the key is never shown again. Keys are encrypted at rest in the same way as data source credentials, and a saved key is only ever sent to the base URL it was saved for.
Network: AI Analyst refuses to call addresses the deployment is not allowed to reach (the Test shows "This deployment isn't allowed to call that address.").
Model quality: the endpoint must support tool calling and must pass a Test before it can be saved. Answer quality depends on the model you choose.

With your own endpoint, prompts and query results go to the provider you chose. Your agreement with that provider governs its data handling, including whether it uses data for training.

Data shared with Actian for support¶

Bug reports that people submit (Report a Bug in Studio, Report a problem under an answer or in Steward) and issues the AI agents flag for Actian engineering are stored in your organization's database and, on Actian SaaS and connected deployments that send them, posted to Actian. On air-gapped deployments, nothing is sent: an administrator hands them to Actian in a diagnostic export. See What is shared with Actian for exactly what each contains.

Other services AI Analyst may contact¶

Depending on how it is configured, Actian SaaS (and a connected self-managed deployment) may contact third-party services for:

Messaging apps you connect, such as Slack and Microsoft Teams.
Sign-in, email delivery, billing entitlements and support chat.
Web search for Steward (Search the web).
Diagnostic tracing.
AI models: Actian Inference in Actian SaaS, or the endpoints your administrators configure.

An air-gapped deployment refuses to start if any vendor-hosted service is configured. Its sign-in, AI models, embeddings, storage, email and tracing all run on your network, and integrations that cannot be self-hosted (for example Slack and Microsoft Teams) are unavailable.

Self-managed deployments¶

In a self-managed deployment, your organization runs AI Analyst on its own infrastructure. The rest of this page applies, with these differences:

Topic	Self-managed deployment
Product database	A PostgreSQL server your organization runs: bundled with the Docker Compose installation, or your own with Kubernetes (Helm).
Uploaded files and attachments	A local Docker volume (Docker Compose) or your own S3-compatible storage (Kubernetes).
Sign-in data	Kept in the same PostgreSQL server when you use the Keycloak that ships with the product; otherwise with the identity provider your installation uses.
AI models	No AI model ships with the product. AI Analyst uses only the endpoints your administrators configure: a default set by your operator, or your organization's own under Settings → Administration → Inference. If no chat endpoint is configured, conversations are blocked; if no embedding endpoint is configured, search matches keywords only.
Credentials and API keys	Encrypted with AES-256-GCM using an encryption key that your operator generates and keeps in the installation's configuration.
Diagnostic traces	Off in the shipped configurations. Your operator can turn tracing on and decides where traces go. On an air-gapped deployment, vendor-hosted tracing services are not allowed, and your operator is responsible for keeping traces on your network.
What leaves your network	Air-gapped: the AI Analyst server makes no calls to Actian or to other vendors' services. Connected: only the services and endpoints your operator configures, plus bug reports and engineering flags if your deployment sends them to Actian.

Note

"Nothing leaves" describes the AI Analyst server. People's browsers load the product from your installation, but the sign-in page may still try to load some public media files from the internet.

When an organization is deleted¶

When an organization is scheduled for deletion, it becomes pending deletion: product work is paused, and the organization can be restored during a grace period of 14 days (the grace period can be skipped).

At the end of the grace period, AI Analyst permanently deletes:

The organization's product data, including its semantic layer, conversations, query results, artifacts and audit log.
Stored files: the organization logo, chart fonts, attachments, uploaded data files and audit exports.
Pending background jobs.
Where applicable, its billing record, its identity-provider organization and SSO connection, and its traces in Langfuse.

People's own sign-in accounts and shared chat-app installations are kept, because they can belong to other organizations. On self-managed deployments that use Keycloak, the Keycloak organization itself is not removed. After deletion, the organization cannot be restored.

Summary¶

	AI Analyst agents	Steward
Accesses your database	Read-only queries, compiled from your semantic layer	Read-only, including table structure and exploratory SQL (asks for approval first by default)
Sends data to the AI provider	Question, semantic layer definitions and query results	Schema metadata, semantic layer context and row-level data from exploratory queries
Stores your data	Query results are kept with the conversation and its artifacts; your database is not copied¹	Query results are kept with the Steward conversation; your database is not copied¹
Used for AI training	Not by Actian Inference providers; your own endpoint follows your agreement with that provider	Not by Actian Inference providers; your own endpoint follows your agreement with that provider

¹ Where diagnostic tracing is enabled, values stored in result-row fields are replaced by a placeholder before a trace leaves AI Analyst, and common personal-data patterns are masked. Result values can still appear where they are quoted in text, such as in a question or an answer. See Diagnostic traces.

For information about our broader security practices and contractual commitments, see How Actian AI Analyst protects your data.

For a full record of actions taken in your organization, see Audit Logs.