An assessment begins with an agreement about the question it will answer. Before asking for evidence, establish which organization or business unit is being examined, which period the work covers, and who will use the conclusion. A list of interviews is not an assessment scope.

The Assurance Line examines organizational governance through the eight domains, supported by testing of selected AI systems and the controls around them. Its ratings describe what the evidence supports within the stated scope. The assessment does not award certification, guarantee safety, establish legal compliance in every jurisdiction, or issue an audit opinion. A separately commissioned audit may use the framework within its own professional requirements.

Choose the assessment type

A self-assessment records management’s supported judgment about its own controls. An internal review can challenge that judgment without being independent of the organization. An independent assessment requires independence from the activities assessed, appropriate competence, and the authority to obtain evidence and report concerns. State which applies. An external consultant who helped design the controls should disclose that involvement rather than imply independence through an external job title.

Name the sponsor, the person accountable for the engagement, the assessors, and the intended recipients. Identify who can approve scope changes and who will receive urgent findings. The sponsor should enable access and resolve obstacles; sponsorship does not entitle someone to suppress a supported conclusion.

Set the organizational and system boundaries

Start with the organization’s documented definition of an AI system under VIS-3. Check that it captures AI embedded in products already in use, including features activated through a vendor setting. Obtain the inventory and compare its definition with the activities people describe. Record disputed inclusions before selecting the sample.

For each selected material system, establish the boundary under VIS-4. Record the purpose, model, system instructions, retrieval sources, memory, connected data stores, tools and permissions, vendors, human decision points, and affected population. Record the version or configuration examined. A product name alone cannot establish this boundary.

Distinguish two layers of work:

• Organizational controls: inventory governance, delegation, risk appetite, assurance planning, reporting, and other arrangements that apply across the entity.
• Selected system testing: evidence that those arrangements operate for the named systems and processes examined.

Describe both layers in the report. Testing one well-managed system does not establish effectiveness across every system. If a shared process is claimed to cover all systems, examine its population and exceptions as well as the selected examples.

Select systems for a reason

Record the population from which the sample is selected, its known limitations, and the reason for selecting each system. Consider potential harm, decision authority, sensitive data, affected people, financial significance, regulatory exposure, vendor concentration, and material change. Include a less visible use where it can test the completeness of management’s account.

A system without a material classification can still deserve examination when the classification itself is doubtful. Ask what would make a seemingly minor use consequential. Do not let management’s label replace the assessor’s scope judgment.

Agree an assessment period and an evidence cutoff. Separate the condition at the cutoff from improvements made afterward. State whether the review concerns control design, implementation, operation over a period, or all three. For a newly deployed system, report the limited operating history rather than imply that a short period establishes sustained effectiveness.

Make the engagement workable

Agree evidence request dates, interview availability, review meetings, and the report date. Identify technical, privacy, legal, and other specialists needed to examine the selected uses. Record what the team can assess competently and where a specialist conclusion will be required.

Agree how evidence will be transferred, who may access it, how sensitive information will be minimized, and how working papers will be retained and disposed of. Request controlled access or redacted records where full copies are unnecessary. Do not put live customer data, credentials, or confidential documents into an unapproved AI service to summarize the assessment.

Set an escalation route for immediate concerns, such as unauthorized access, uncontrolled decisions affecting people, or a failure that cannot be suspended. Do not wait for the final report to tell the accountable owner about an active material exposure.

Complete the scope before fieldwork

The agreed scope should record:

1. Entity, business units, locations, and intended recipients.
2. Assessment type, purpose, framework version, period, and cutoff.
3. Organizational processes and selected systems, including selection rationale.
4. System boundaries, materiality considerations, and known inventory limitations.
5. Inclusions, exclusions, dependencies, and restrictions on access.
6. Sponsor, accountable owners, assessors, specialists, and conflicts.
7. Evidence handling, timetable, escalation route, and scope-change approval.
8. The conclusions the work is intended to support, and the claims it cannot support.

Chapter output: an agreed scope and engagement plan. Revisit it when evidence reveals a different population, boundary, or risk from the one originally described.