Role and authority
How should an agent’s role, delegated authority and prohibited actions be made operationally clear?
Flagship research programme
Assuring the Agentic Enterprise
As AI agents gain the capacity to interpret, decide and act within organisations, assurance must address not only what a model can do, but how agency operates inside real work.
Programme status
This is a developing research programme. Its questions and themes are being defined openly; no findings, study participants or completed publications are implied.
The research question
Who assures agency inside the enterprise?
Task completion alone does not establish that activity was appropriate, authorised or aligned with organisational intent.
The programme investigates the operational conditions needed for AI-agent activity to remain within defined roles, permissions, policies, responsibilities and escalation paths as work unfolds.
The question is not only whether an agent is capable. It is whether its activity can be understood, bounded and assured in context.
Research themes
Eight questions shaping the programme
Intent and context
How can current policy, process and situational context reach an agent at the moment of action?
Human oversight
Which decisions require human judgement, approval or the ability to interrupt activity?
Assurance Points
Where can consequential activity be evaluated without creating indiscriminate checkpoints?
Exceptions and escalation
How should uncertainty, ambiguity and out-of-bound conditions be recognised and escalated?
Evidence
What records are needed to understand decisions, actions, approvals and outcomes?
Learning
How should operational evidence improve instructions, controls and future assurance?
Shared accountability
How do responsibilities remain clear across leaders, process owners, technology teams and users?
Two assurance questions
Capability is not the same as assured participation
Model assurance and operational assurance address related but different concerns. Organisations may need both.
Model assurance
Operational assurance
Primary signal
Model behaviour and technical performance
Activity within a specific organisational workflow
Relevant boundary
Defined evaluation conditions
Policies, permissions, responsibilities and operational context
Question
Can the system perform the task?
Did the activity remain aligned with organisational intent?
Evidence
Test results and performance measures
Decisions, actions, exceptions, approvals and outcomes
Primary signal
- Model assurance
- Model behaviour and technical performance
- Operational assurance
- Activity within a specific organisational workflow
Relevant boundary
- Model assurance
- Defined evaluation conditions
- Operational assurance
- Policies, permissions, responsibilities and operational context
Question
- Model assurance
- Can the system perform the task?
- Operational assurance
- Did the activity remain aligned with organisational intent?
Evidence
- Model assurance
- Test results and performance measures
- Operational assurance
- Decisions, actions, exceptions, approvals and outcomes
Agency in workflow
Assure the moments where agency becomes consequential
This illustrative workflow shows where roles, boundaries, human judgement and evidence may need to meet AI-agent activity. It is not a universal control design.
Illustrative workflow
Assurance is placed at material moments—not around every action.
- 01
Request
Human activity
- 02
Interpret
Agent activity
AP 01 - 03
Decide
Shared activity
AP 02 - 04
Act
Agent activity
AP 03 - 05
Evidence
Human activity
Define authority and boundaries
Evaluate material moments
Retain usable evidence
Questions for practitioners
Begin with the work, not the technology alone
- 01
What work has the agent been authorised to perform?
- 02
Which source expresses current organisational intent?
- 03
Where could an incorrect action materially affect the intended outcome?
- 04
When must a person review, approve or interrupt activity?
- 05
How are exceptions recognised and escalated?
- 06
What evidence would allow the organisation to reconstruct what happened?
Developing research agenda
An inquiry being shaped in the open
The Institute intends to develop the programme through research papers, executive interviews, practitioner conversations and roundtables. Methods, limitations and evidence will be stated alongside any future findings.
Participation or contribution does not imply endorsement, partnership or a published finding.
Explore the communityRelated concepts
Continue through the field
Participate
Contribute evidence, experience or a question
The Institute welcomes relevant perspectives from people working across enterprise AI, operations, governance, risk, audit and behavioural systems.