---
title: "KYE™ Delegated Auditability™ — Runtime accountability for AI agents, without replacing IAM"
description: "KYE™ adds a delegated-authority and evidence layer on top of your existing IAM, OAuth, API gateway, workflow, SIEM, GRC and AI-agent stack"
url: https://kyeprotocol.com/ai-agent-auditability/
lang: en
source: "KYE Protocol"
---

KYE™ Delegated Auditability™

# Runtime accountability for AI agents — without replacing IAM.

KYE™ adds a delegated-authority and evidence layer on top of your existing IAM, OAuth, API gateway, workflow, SIEM, GRC and AI-agent stack. Start in shadow mode: observe AI-agent actions, capture who or what acted, on whose behalf, under what authority, inside what scope, and generate Evidence Packs™ before enforcing runtime controls.

[Apply for pilot](https://kyeprotocol.com/pilot-apply/) [View Evidence Pack™](https://kyeprotocol.com/evidence-pack-demo/) [See financial-services workflow](https://kyeprotocol.com/financial-services/)

The problem

## Why IAM is not enough.

An OAuth token says _this caller is identified and authorised to call this API_. It does not say:

- **On whose behalf** the agent claims to be acting — principal, delegation chain, time window.
- **For what purpose** the call is admissible — data classes, jurisdiction, restrictions.
- **Under what scope** the action is bounded — resource, amount, blast radius.
- **Whether a regulator can replay** the decision offline.

Delegated Auditability™ is the smallest possible KYE™ adoption that closes those four blind spots — without altering one line of your production code path.

Adopt KYE™ incrementally

## Adoption is staged — you pick how far you go.

Customers begin in observe-only mode (read-only stack binding, signed Observed Actions, production unchanged) and graduate through evidence, alert, guard, enforce and expand stages on their own schedule. The specific stage enumeration, the per-stage promotion criteria, and the rollback contract are proprietary and are not disclosed in this repository — they live in the private Adoption Stage State Machine and are governed per-customer by signed transitions.

### Observe

Read-only stack binding. Production unchanged.

### Guard

Smallest possible Authority Gate or Purpose Permission installed — in shadow first.

### Enforce

Promoted Guards block on production. Each promotion is a signed adoption-stage transition.

Shadow mode

## Shadow mode is a flag, not new code.

Every KYE™ Engine — Authority, Purpose, Decision — supports a `mode` parameter. Under `mode: shadow`, every check still runs, every Decision Map™ is still sealed, every Evidence Pack™ is still signed — but the Commit Boundary™ suppresses every side effect. `production_action_blocked` is invariantly `false`.

[Full shadow-mode contract](https://kyeprotocol.com/shadow-mode/)

Evidence Pack™ example

## A real bundle from a supplier-payment pilot.

One observed action — _supplier\_payment\_agent prepares £950 payment for invoice inv\_123_. The KYE™ Shadow Evaluation runs all six engines and returns `simulated_requires_approval`. The Authority Gap classifier opens a `missing_authority_grant` gap. The Guard Recommendation proposes an Authority Gate. The Evidence Pack™ bundles every artefact for offline replay.

[Open the Evidence Pack™ viewer](https://kyeprotocol.com/evidence-pack-demo/)

Built above your existing stack

## Eleven read-only Stack Bindings — no migration required.

### IAM / SSO

Okta, Entra ID, PingFederate, Auth0, Keycloak. Roles, groups, sessions.

### OAuth / OIDC

Scopes, tokens, claims, refresh chains.

### API gateway

Kong, Apigee, AWS API GW, Cloudflare. Per-call observation.

### MCP servers

Any MCP-conformant server — tool calls captured as Observed Actions.

### AI-agent frameworks

LangChain, LlamaIndex, AutoGen, OpenAI Agents SDK.

### Workflow engines

Temporal, Camunda, AWS Step Functions, n8n.

### SIEM

Splunk, Sentinel, Elastic, Chronicle — alert fan-out.

### GRC

OneTrust, Drata, Vanta, ServiceNow IRM.

### Policy engines

OPA, Cedar, Styra DAS.

### Audit logs

Read-only ingest from any append-only log.

### Data stores

Snowflake, BigQuery, Postgres, S3 — for data-class tags only.

Integration path

## From day 1 to first Evidence Pack™ — in three steps.

1. **Bind.** A KYE™ operator helps you install one or more read-only Stack Bindings. Default mode is `read_only`. No production change.
2. **Observe.** Each bound stack streams Observed Actions to the KYE™ Evidence Gateway™. A Shadow Evaluation fires for each one.
3. **Review.** KYE™ returns Authority Gaps and Guard Recommendations. Your CISO and AI risk officers triage them in KYE™ Cloud™.

Pilot CTA

## Apply for a pilot.

Pilots run 30–90 days, end with a signed Audit Pilot Report and a prioritised list of Guard Recommendations. Applications are manually qualified within 2 business days.

[Apply for pilot](https://kyeprotocol.com/pilot-apply/) Talk to the team

Canonical KYE™ surfaces referenced on this page: [Audit Pilot](https://kyeprotocol.com/auditors/) · [Delegated Auditability](https://kyeprotocol.com/glossary/) · [Evidence Pack™](https://kyeprotocol.com/evidence-pack/) · [KYE™ Cloud™](https://kyeprotocol.com/apps/) · [KYE™ MCP Server](https://kyeprotocol.com/mcp/) · [KYE Protocol™](https://kyeprotocol.com/) · [Purpose Permission](https://kyeprotocol.com/glossary/).