Dominic Chiappe · Henley-on-Thames

Thinking about how organisations perform in an AI-enabled world

Assurance

Agent Competency Assurance

Research proposition

Ask what an agent can reliably do in this configuration, and what it is authorised to do.

Define the work, prove capability and grant bounded authority, then observe and revalidate.
  1. DefineWork and limits
  2. EvidenceTest the task
  3. AuthorityBounded permission

Observe · learn · revalidate

Existing authority-cycle diagram: define work, gather evidence, grant bounded authority and revalidate.

The problem

A model score does not establish competence in a particular workplace. Tools, prompts, permissions, memory, supervision and changes to the environment can alter the deployed agent's behaviour.

Illustrative example

A small, practical starting point

Before an agent reconciles a set of records, the proposed assurance process would define the task and limits, examine relevant test evidence and allow only bounded work under supervision. A change of tools or configuration would trigger a fresh review.

Keep exploring