
AI Agent Data Privacy: Protect Tool Calls and Documents

By Erin Harris, AI Strategy & Business Transformation Consultant
AI agent privacy requires controls at the points where an agent calls outside tools and receives their results. Protecting only the initial prompt does not prevent an assistant from passing personal details to a search API, SaaS application or document processor.
On November 19, 2026, AWS and Skyflow published a reference implementation for Amazon Bedrock AgentCore Gateway. It uses request and response interceptors to de-identify sensitive fields, restore authorized values and substitute de-identified documents. The interceptor is open source, Skyflow is separately available through AWS Marketplace, and signed-URL document de-identification relies on a preview capability. Source: https://aws.amazon.com/blogs/machine-learning/protect-agent-tool-calls-with-skyflow-and-amazon-bedrock-agentcore-gateway/
Start with a data-flow inventory
Before selecting software, list each agent tool, the business reason it is called, the exact fields it receives and whether the receiving provider is inside or outside your control. Include returned documents and logs, not just outgoing requests. A single unnecessary email address in a search query is a meaningful design defect even if the final answer is correct.
A field-level decision table
For each field, classify it as required, useful but optional, or unnecessary. A local-weather lookup may need a city, while a billing lookup may require a customer ID. If a third-party tool needs a stable reference but not the person's identity, evaluate a token or pseudonymous identifier. Do not assume tokenization is anonymous if a trusted system can reverse it.
Three control layers
First, authorization determines which tools and records an agent may access. Second, minimization removes information that a permitted tool does not need. Third, monitoring records whether controls worked, including refusals and unexpected data flows. These layers answer different questions and should be evaluated separately.
The AWS and Skyflow design places an interceptor at the gateway so agents and tools can remain unchanged. Its modes address outgoing tool calls, model context and linked documents. String de-identification is described as generally available. Document processing through Skyflow Detect v2 is preview, and Region support, latency and document size limits must be validated. Source: https://aws.amazon.com/blogs/machine-learning/protect-agent-tool-calls-with-skyflow-and-amazon-bedrock-agentcore-gateway/
Hypothetical service-business test
A plumbing company's AI assistant reads a service request containing a customer's name, phone number, street address and a description of a leak. It calls a parts-search service to find a replacement valve. The parts service needs the valve specifications, not the customer's identity. In a test, send the request through a controlled layer and verify that only the part details leave. Then deliberately place an email address in the free-text description and confirm the detector handles it. Finally, simulate a protection-service outage and observe whether the call is blocked. This example is illustrative, not a vendor benchmark.
How to pilot without creating a new blind spot
Use synthetic records for initial tests. Include false positives, missed identifiers, oversized documents and third-party tool failures. Measure detection precision, legitimate task completion, processing time and whether logs accidentally retain raw values. Establish who can authorize re-identification and how long tokens and source records are retained. Do not turn on production access merely because a demo hides a name.
When a simpler solution wins
If an agent only needs a few fields, reducing the tool schema may be easier to maintain than adding a full tokenization service. If external calls are unnecessary, eliminate them. The objective is minimum necessary data exposure, not the maximum number of security products.
Key Takeaways
- Map tool requests, results, documents and logs.
- Enforce authorization separately from de-identification.
- Pilot preview document features with explicit limits.
- Measure both missed sensitive fields and blocked legitimate work.
- Remove unnecessary tool calls before adding controls.
Where to begin
An AI Opportunity & Transformation Assessment with EH Synergy can identify the first workflow worth mapping and the controls proportionate to its risk. Inquire through EH Synergy's established contact channel.

