Contact

How Aegis4A works · From AI intent to governed execution

How Aegis4A governs AI agent requests

Every request takes the same path: check the rules, hold for a person if needed, record it, then forward it. There is no other route to your models or systems.

Book a demo

The request path

Four steps, every request

  1. 01

    Attach

    The agent uses an Aegis4A URL and key. Aegis4A holds the real credentials.

  2. 02

    Decide

    A rule matches the request and returns execute, approve, escalate or block. No rule, no action.

  3. 03

    Approve

    If approval is needed, a ticket opens. A reviewer approves it one time or always.

  4. 04

    Record

    The intent is logged before forwarding and the outcome after, hash-chained and copied off the server.

Works with

The model providers your agents already use

One OpenAI-compatible endpoint. Agents keep their SDKs; only the base URL and key change.

Walkthrough

A prior authorization request, step by step

  1. Request. An intake agent submits a prior authorization to the payer system through Aegis4A.
  2. Hold. The rule requires approval, so nothing is sent. The agent gets ticket A4A-000123.
  3. Approve. A reviewer checks the exact request in the console and approves it one time.
  4. Re-run. The agent sends the same request again. Aegis4A matches the approval and forwards it with the payer credential it holds.
  5. Record. The audit trail shows the request, decision, approval, approver and response, in order.

Change any field, such as the service code, and the approval no longer matches. A new ticket opens showing what changed.

The four boundaries

What each boundary checks

AnswerAnswers cite approved sources, or are flagged or withheld.
ActionRules per route and tool, with field conditions such as amount over 1,000.
AuthorityAdmin, manager and user roles. Always-approvals and escalations need a manager.
AuditA hash-chained log with an off-server copy. Any edit or deletion is detectable.

Fail-closed

If a check can’t run, nothing runs

If the rules can’t be read or the audit log can’t be written, Aegis4A refuses the request.

Near-matches

AI can advise. People decide

For near-matches, an optional AI reviewer adds a recorded verdict. By default it never approves anything itself.

See it live

See this walkthrough run on a live agent

Book a demo Use cases