CONTENT POLICYFINANCIAL SERVICES
Dynamic routing between internal and external models based on content sensitivity.
A financial services team needed to use frontier models for complex analysis—but couldn't send PII or proprietary data to external providers. Manual classification was slow and error-prone.
!
The challenge
Analysts needed GPT-4 for research, but prompts often contained customer data, account numbers, or internal financial models.
→
The solution
Airlock intercepts every request, runs content classification, and routes sensitive prompts to an air-gapped internal model while public queries go to external APIs.
✓
The outcome
100% policy compliance with zero workflow friction. Analysts use AI freely—Airlock enforces data boundaries automatically.
HOW IT WORKS
INCOMING REQUEST
"Summarize Q3 revenue for account #4521..."
AIRLOCK CONTENT DETECTION
PII detected: account #
Financial data: revenue
Sensitivity: HIGH
Policy: INTERNAL_ONLY
SENSITIVE
ROUTE TO
Internal Model
Air-gapped, on-prem
PUBLIC
ROUTE TO
External API
GPT-4, Claude, etc.
100%
Policy compliance
0
Data leaks
<50ms
Classification time
COST OPTIMIZATIONINSURANCE / DOCUMENT PROCESSING
80% cost reduction by migrating document extraction to internal OSS models.
An insurance team was spending $40K/month on GPT-4 for document extraction. They suspected smaller models could handle it, but didn't have cycles for a lengthy evaluation.
!
The challenge
Document extraction was simple but high-volume: policy numbers, dates, coverage amounts. GPT-4 was overkill, but evaluating alternatives meant months of engineering.
→
The solution
Airlock's shadow mode let them route 10% of traffic to an internal Llama model while comparing outputs. No code changes. No pipeline disruption.
✓
The outcome
After 2 weeks of shadow evaluation, they migrated 100% of document extraction to internal models. Fixed cost, faster latency, no quality loss.
MIGRATION JOURNEY
1
Before: External API
GPT-4 for all document extraction
$40K/mo~800ms latency
2
Shadow evaluation
10% traffic to Llama 3.1 70B, compare outputs
2 weeks99.2% accuracy match
3
After: Internal model
100% migrated to on-prem Llama
$8K/mo (fixed)~200ms latency
80%
Cost reduction
4x
Faster latency
2 wks
Evaluation time
SECURITY INCIDENTENTERPRISE SOFTWARE
Caught an agent exfiltrating data to an unauthorized endpoint.
A development team deployed a third-party coding agent to help with refactoring. Within hours, Airlock flagged suspicious behavior: the agent was sending code snippets to an external API without permission.
!
The incident
A coding agent was silently POSTing source code to an analytics endpoint embedded in its runtime. No disclosure in the agent's documentation.
→
How Airlock caught it
Every agent action flows through Airlock. The egress policy flagged outbound requests to an unapproved domain. Full audit trail showed exactly what data was sent.
✓
The response
Agent was terminated immediately. Security team had forensic evidence within minutes. The vendor was notified and the agent was banned org-wide.
INCIDENT AUDIT TRAIL
14:23:01.442ALLOW
Agent "refactor-bot" → Read file: src/auth/login.ts
14:23:02.118ALLOW
Agent "refactor-bot" → LLM request: Anthropic API
14:23:03.891BLOCKED
Agent "refactor-bot" → POST to analytics.agent-vendor.io
⚠ Egress policy violation: unapproved domain
14:23:03.892ALERT
Data exfiltration attempt detected
Payload: 2.3KB source code from src/auth/*
14:23:04.001TERMINATED
Agent "refactor-bot" session killed
Reason: Security policy AUTO_TERMINATE_ON_EXFIL
1
Exfil attempt blocked
<1s
Detection to termination
Full
Forensic audit trail
Ready to govern AI at your enterprise?
See how Airlock can help you route intelligently, optimize costs, and catch security incidents before they become breaches.