Data Access and Security
When employees upload files individually, it’s hard for your organization to know what data reaches the AI app. With Decision Layer, you can define what data the app can receive and review the query conditions and results in the run record.
Decide who can query which data.
Configure the connection to share only authorized query and analysis results with internal AI apps, ChatGPT, or Claude.
See the query conditions and the results they produced.
The connected Semantic Layer defines the data views, metrics, and fields available for analysis. You can exclude sensitive fields from views or configure access policies to return masked values.
To enforce individual access, configure the connection so the user’s permissions carry through to the data query. If a shared service account runs the query, its permissions apply. If you need a separate scope for each user, verify how those limits are enforced when the query runs.
When the AI app calls an MCP tool to answer a question, it sends query inputs such as a date range or customer account. Depending on the tool, it may also send part or all of your question.
Query permissions and tool settings determine what the AI app receives, such as totals, individual records, or analysis results and supporting evidence. The app’s policies and organization settings govern how it uses that information to answer and how it retains it.
The user asks a question about company data.
ChatGPT or Claude retains conversations according to the app’s policies and organization settings.
When connecting
authentik verifies the user’s identity when they set up the MCP connection.
When asking a question
Analyzes the data using the registered procedure.
Queries only the rows and fields the account is allowed to access.
Review query conditions and analysis results in Decision Layer.
Enterprise workspace admins can approve access to the connection. Review connection access and data permissions separately. For analysis, expose only the query tools you need and grant the account read-only access.
The Decision Layer run record shows query conditions, the versions of the methods and procedures used, step results, and validation details. It can also show generated SQL when the connection captures it. A record can contain the question or result data, so set permissions for reading it separately.
The open-source Decision Layer’s local example keeps run records in SQLite in the environment where it runs. For a deployment, check where run records are stored, who can access them, how long they are kept, and how to delete them. Before sharing a remote MCP connection, verify user authentication and access isolation for run records.
Disconnecting MCP stops future access through that connection. Existing Decision Layer run records and ChatGPT or Claude conversations are managed separately under the policies of the systems that store them.
Check whether data is used for training and how long conversations and tool results are retained. The policies below apply to business services. Personal accounts follow their own service policies and settings.
OpenAI does not use business data or API inputs and outputs for training by default. Separate terms apply if you explicitly opt in to data sharing. Conversation retention depends on the plan and workspace settings. API retention depends on the features and data controls in use.
Anthropic does not use commercial product inputs or outputs for training by default. Separate terms apply when you submit feedback or agree to share data. Conversation and API data retention depend on the service, settings, and contract terms.