Case Study — Private AI Gateway & Compliance

AXA: Private AI Gateway with In-House PII Protection at Global Scale

A private, high-throughput AI gateway providing centralised routing, in-house PII/PHI detection and auditable governance across a global operation.

Industry

Insurance

99.99%Gateway Availability
AI GatewayPII/PHI DetectionPolicy EnforcementAuditable Governance

The challenge

Insurance means regulated, sensitive data moving through AI systems at global scale — personal data in prompts, across many markets and teams, with different obligations in each. The central problem is not model capability. It is that sensitive data must not reach a model, and the difference between a control that exists and a control that operates is the entire risk.

Application-level enforcement does not hold at that scale. If every team implements its own redaction, controls diverge, there is no way to verify compliance after the fact, and the first evidence of a problem is an incident. Detection quality is also harder than it looks: formatted identifiers are easy, and the risk lives in unstructured free text — claims narratives, medical references, internal identifiers in ordinary prose.

Beyond prevention, the gateway had to produce evidence. Regulated operations get asked to demonstrate what happened, when, under which policy — which means a complete audit trail as a first-class requirement, not a nice-to-have.

The approach

01

One enforced path for all model traffic

Every model call was routed through a single gateway so that policy is enforced by position rather than by each application remembering to enforce it. Routing, access control and budgets were unified in the same place, which turned the gateway from a security control into a financial one.

02

In-house PII/PHI detection before dispatch

Detection runs before the prompt reaches a model, because once sensitive data is in a prompt it has already crossed the trust boundary — and most providers will have logged it. Rather than regex, detection uses purpose-built models covering the data classes actually processed, which is what makes unstructured free-text detection viable at all.

03

Policy as a routed decision

Policy is evaluated per request against data class, model and environment, producing an allow, deny or redact decision. Routing is then made by policy rather than by whichever client library a developer happened to import, which is what makes the answer to 'why did this request go there' auditable.

04

Auditable governance by construction

Every request is logged with its policy decision, the redactions applied, the model selected and the outcome, retained to the compliance window. The audit trail is generated by the enforcement path rather than reconstructed afterwards, which is the difference between evidence you have and evidence you can produce.

The result

  • Gateway availability held at 99.99% across a global deployment.
  • All model traffic flows through one enforced path with consistent policy decisions.
  • PII and PHI are detected and handled before prompts leave the perimeter.
  • Every request carries an audit record sufficient to reconstruct the decision that was made.

Facing a similar problem?

Private AI Gateway & Compliance is one of the four service lines we operate. We can walk through your traffic profile, data constraints and cost targets directly.