Retrieval ignores the person
A search finds relevant records, but relevance alone does not establish whether this user and purpose are allowed to use them.
An authorized question is not permission to send an entire customer record to any available model. Frozion governs retrieval, context, tool responses and model routing so the data path follows the purpose of the work.
Allowed for this person. Needed for this task. Approved for this model.
Identity, tenant and resource scope determine the eligible records.
Declared context and tool-response policies limit what enters the model request.
Classification, region and handling rules determine the eligible model.
The final answer can look harmless even when the request sent unnecessary or unauthorized information to a provider. Control the inputs and the destination, not just the words returned.
A search finds relevant records, but relevance alone does not establish whether this user and purpose are allowed to use them.
An agent needs a refund status and receives the customer’s wider account history. A tool response adds even more fields than the task requires.
An approved model is busy. A fallback that considers only availability can send the same context to a provider or region the data was never cleared for.
The controls work together: permission determines what can be retrieved, context determines what is needed, tool policies govern returned data and routing determines where the resulting request may go.
Scope retrieval to the initiating identity, tenant, resource permissions and business purpose. A useful match still has to be an authorized match.
Supply the fields and facts the task expects. Keep their source, scope and freshness attached so the next step can distinguish a usable input from missing or stale information.
Apply configured response policies at the governed tool boundary, including redaction of designated sensitive fields. A tool’s ability to return a field does not mean the model needs it.
Filter model routes by data classification, provider, region and handling requirements before applying availability or cost preferences. Keep those constraints on fallback routes too.
A support representative asks why a refund is pending. Follow the sample request from retrieval to the approved model boundary.
Follow each checkpoint before a model receives the request.
Of 40 candidate records, six match the user’s permitted customer and case scope. The model receives no permission to inspect the other records simply because the search found them.
The task requires refund status, the recorded reason and the applicable policy. The declared input leaves unrelated account details out of this model request. Source and freshness information stay connected to the facts.
The refund lookup returns a designated sensitive field. The configured tool-response policy redacts it before the response is passed into model context. Policies must identify the fields and transformations needed for this integration.
This request is restricted to an approved private model in the EU. That model is temporarily unavailable; the managed route does not meet the request’s rules. The case waits. An answer can be produced after an eligible route becomes available.
Security, privacy and data teams can review where information enters the process and which rules govern its use.
Inspect the path from an authorized identity to retrieval, tools and model destinations. Keep fallback behavior inside the same policy boundary.
Control the path to the model.Connect the business purpose to the fields used and the approved handling requirements. Review the available evidence with appropriate access.
Explain why this data was used.Define useful task inputs with sources, scope and freshness. Stop treating the largest available record as the best possible context.
Share the right facts for the task.Data minimization, response policies and eligible model routes are application decisions that can be inspected and maintained as the work changes.
Running in your cloud gives you placement control. Separately approve model endpoints, connected systems and evidence destinations. Hosting location alone does not determine every data flow.
Removing fields can change the answer a model can produce. Define the required inputs and evaluate the task with its permitted context, including the path for missing information.
Bring one agent or application that handles sensitive records. We’ll define what it may retrieve, what the task needs and which model destinations are allowed.