Aggregate insights
Offer fixed, parameterized headcount, attrition and tenure tools. The analytics backend remains responsible for valid SQL, cohort size and preventing sensitive rows from appearing in results.
Use case · People Analytics
Give People Analytics AI access to trends, not unrestricted employee records.
A workforce analytics copilot should answer aggregate questions about headcount, attrition and tenure, not inherit every permission of a privileged HR service. Talon's existing model and MCP governance controls can enforce an AI-use-case-specific tool boundary on intercepted traffic—with signed decision evidence.
workforce-analytics AI use case
→ Talon-authenticated model / MCP path
→ policy: aggregate tools only
headcount_aggregate ALLOW
attrition_aggregate ALLOW
tenure_aggregate ALLOW
employee_export DENY
direct database access:
must be blocked by HR IAM + deployment
signed Talon evidence: intercepted callsAI-use-case identityLimited HR toolsPre-call MCP denialBackend access controlsSigned decisions
The business problem
Existing warehouse governance is necessary, but AI introduces a new action surface.
An HR warehouse may already have department permissions, aggregation rules and sensitive employee fields protected by established access controls. Adding an AI assistant should not mean passing privileged raw-data credentials into the new runtime. The useful boundary is an authenticated analytics use case that can invoke only the permitted reporting operations.
Prevent employee exports
When an agent attempts a forbidden employee_export MCP call through Talon, policy can deny it before the upstream tool executes and record the denial. Hiding a tool alone is not direct-call authorization.
Preserve auditability
Attribute routed requests and allowed/denied calls to the authenticated AI use case. Signed Talon evidence supports verification of observed decisions; it is not a complete HR activity log.
Two enforcement layers
Talon governs routed AI actions; HR infrastructure prevents bypass.
Talon applies company policy at the model gateway and MCP proxy to traffic it receives. But a compromised AI application could bypass a proxy if it also has credentials, filesystem access or network reachability to the private HR warehouse. That protection must come from established HR authorization, separate service identities, network restrictions and data-storage isolation.
The synthetic workforce example uses DuckDB and separate services to illustrate both layers. Its isolated browser/AI container does not mount the database or have a direct network route to the private HR service. The separate aggregate dashboard backend serves fixed metrics without exposing arbitrary SQL. These are deployment decisions, not hidden Talon features.
AI application
├─ model traffic → Talon gateway
├─ HR tool call → Talon MCP proxy
│ ├─ aggregate metrics: allowed
│ └─ employee export: denied
└─ raw database access: no route
↑ external IAM + network isolation
Talon evidence: observed policy actions
HR service logs: backend activity
neither replaces the otherHow to evaluate
Prove a useful answer and zero forbidden exports.
Use fictional employees only. Compare a deliberately privileged direct HR-backend call against an isolated AI deployment; do not claim Talon blocked that privileged direct call. Then show a permitted aggregate request routed through Talon and deliberately invoke a forbidden export through its MCP proxy. Inspect the backend dispatch counter and signed evidence independently.
1. Ask an aggregate question
Which department has the highest attrition over six months? Require parameterized aggregate tool results and avoid invented causes for differences.
2. Attempt an individual export
Submit a forbidden employee-record export through the governed MCP path. A denial should prevent the upstream export handler from running.
3. Verify both boundaries
Confirm the AI runtime cannot directly reach raw HR data, then verify that allowed/denied Talon records match the actual intercepted calls.
The example is synthetic, not a Workday or other production HRIS integration. This use case relies on capabilities already implemented in Talon; the end-to-end demonstration and optional real-provider smoke are being integrated separately in talon-full-demo PR #39. Do not represent an unexecuted provider smoke as passed, or claim production GDPR compliance, universal data isolation, or complete agent activity visibility.
Bring one reporting operation
Keep analytics useful. Stop unrestricted employee access.
Start with a single aggregate HR tool and one forbidden export. Identify the actual policy enforcement point and the separate infrastructure controls that prevent bypass.