Continuous AI assurance

Know when your AI is no longer what you approved.

MetricsLM establishes an evidence-based Boundary for AI systems, detects material change, and triggers reassessment when existing approval is no longer valid.

Audit-Ready. Always-Ready.

Book a demo See how it works

Customer Refund Agent

Boundary · v1

Approved

Approved 16 Sep 2026

Purpose: Resolve eligible customer refund requests
Model: Claude Sonnet
Data: Orders · Customer profiles
Tools: CRM · Refund API
Authority: Refunds ≤ £500
Oversight: Human approval > £500

Material change detected

2 hours ago

Refund authority: £500→£5,000
Source: refund-service/config.ts

Reassessment required

Partnerships, standards & recognition

Silver Partner

Consortium

Project Female Founder

AI changes after approval. Your assurance shouldn't stay still.

A system can be reviewed, approved, and documented today, then become a different risk tomorrow. Point-in-time assurance is no longer enough once models, tools, and authority can change underneath the last sign-off.

“A point-in-time assurance is not an assurance of what you are running today. If the system can change after approval, the record has to change with it.”

Lena Smart, Strategic Advisor & Investor

A new model

The approved endpoint is no longer the one running in production.

A new tool

The system can now call capabilities that were never reviewed.

A different risk owner

Authority moved, but the last approval still looks current.

A broader permission

Limits, data access, or refund authority quietly expand.

From approval to reassessment·Operating model

The closed loop of continuous assurance.

Define the boundary, watch the system, and reassess when it no longer holds.

1Define Boundary
Guardrails, permissions, and scope
2Baseline v1
Approved state and expectations
3Detect Change
Meaningful runtime changes
4Reassess
Update status and approve

MetricsLMTrusted AI

1Define Boundary
Guardrails, permissions, and scope
2Baseline v1
Approved state and expectations
3Detect Change
Meaningful runtime changes
4Reassess
Update status and approve

Evidence first. Human approved.

MetricsLM does not ask anyone to trust a score. It discovers what a system actually is, proposes a Boundary from that evidence, and keeps every material decision with a human owner.

Book a demo

Findings

Customer Refund Agent · Review 2

3 require review

Finding	Section	Owner	Status
Refund API exceeds approved limit	Authority	Maya Chen Security	Blocked
New refund-policy tool	Tools	Tom Lewis Engineering	Restricted
Model endpoint changed	Model	Priya Shah Responsible AI	Reassessment
Purpose still matches	Purpose	Daniel Reed Product	Approved