Preventive egress decision
An agent-specific rule rejects the first confidential package bound for OpenAI before provider access, incurring no provider cost for that denied call.
Use case · Procurement and legal operations
Use AI for vendor reviews without sending confidential documents everywhere.
Procurement and legal teams can use AI to draft advisory reviews while company policy restricts where confidential inputs may be sent. Talon's n8n demonstration proves a disallowed provider path was rejected before access and an approved, redacted path still produced useful output.
synthetic vendor documents
profile + proposed DPA + review criteria
↓ confidential classification
OpenAI egress → DENIED (zero provider call)
↓ allowed route only
email + IBAN redaction
Anthropic → advisory review draft
↓ signed session evidenceConfidential documentsDestination restrictionPII transformationApproved model routeSigned session
Business scenario
A confidential vendor package should not be routable to any model.
The pinned n8n demonstration assembles a fictional vendor profile, a proposed data-processing addendum and the company's internal review criteria. The workflow attempts one forbidden destination probe and then one permitted review, using the same authenticated vendor-review use case and session.
Keep useful business work
An allowed Anthropic route receives redacted contact email and IBAN values and creates a synthetic advisory review packet.
Verify the reason
The resulting session evidence links the blocked and permitted model decisions, data-handling facts and signatures without asserting control of out-of-band document sharing.
One effective policy, two outcomes
Deny the wrong destination. Continue on the authorized one.
The provider list alone is not enough. Talon checks the configured AI use case's effective policy, including the document's confidentiality and destination constraints. The failed OpenAI probe is a real negative-policy test, not a synthetic request with missing credentials.
Content review remains advisory. The output does not substitute for a human legal, data-protection, security or procurement decision, and the demo makes no claim of automatic regulatory compliance.
vendor-contract-review identity
↓ policy + confidentiality
OpenAI egress DENY
provider calls zero
↓
Anthropic route ALLOW
synthetic PII redacted
advisory draft produced
↓
signed decision evidenceEvaluate the actual perimeter
Questions procurement and platform teams should ask
What data is classified?
Which application input is marked confidential? Can the use case weaken a company-level provider or sovereignty constraint?
What can the LLM reach?
Does every relevant model request traverse Talon? Direct attachments, emails and API uploads outside the path are not covered.
What proves compliance with a rule?
Which denied and allowed requests have independently verifiable records? A valid signature supports integrity, not complete legal compliance.
Reproduce confidential egress control
Test a forbidden provider before relying on the positive path.
Compare the blocked first request with the approved review. The same evidence set should explain both outcomes.