Before Writing Another AI Policy, Record the Systems You Use

A business announces an AI policy while teams adopt assistants, product features and supplier services. Start by recording which systems and business uses the policy governs. Then connect each use to its owner, data, decisions and operating evidence.

ISO/IEC 42001 addresses an AI management system through policies, objectives and processes. NIST's AI Risk Management Framework is voluntary, and its Generative AI Profile provides a companion resource for generative AI risk work. These references support governance planning; the specific deployment still needs assessment. [1–3]

Record the use, not only the product name

One product often supports several activities with different data and consequences. Record the system and business use separately: what someone is trying to accomplish, what information enters the system and how the output influences work.

For a customer-support assistant, review access to customer records, where outputs appear, who reviews them and which supplier arrangements address retention. For an internal drafting tool, examine the material entered, the intended audience and the approval process before publication.

A shared technology label is insufficient to decide the safeguards for both uses. Preserve the distinction between the supplier product, its configuration and the activity the organization performs with it.

Include AI used by the security team

Ask the security owner which product features use AI and what those features do in the actual workflow. Include the uses identified within the agreed inventory boundary; do not limit discovery to customer-facing products and employee chat tools.

For an alert-triage feature, record the information it receives, whether it ranks or groups alerts, who reviews the result and whether a connected workflow can act on it. A feature that proposes an investigation priority and one permitted to disable an account need separately recorded authority and review arrangements. Ask the technical owner to confirm the configuration rather than inferring it from an “AI-powered” product label.

Link the inventory entry to the relevant operating procedure and evaluation evidence. Record unresolved questions about incorrect or missed alerts for the responsible specialist. This applies the inventory's existing ownership and evidence fields to security operations; it does not establish detection performance. NIST's voluntary framework supports inventory and human–AI responsibility decisions. [6]

Build the first inventory around decisions

Use fields the business will maintain:

Build the first inventory around decisions
Field Question to resolve
System and use-case ID Which product and activity are under review?
Owner and user group Who authorizes the use and who performs it?
Purpose and deployment status Why is it used, and is it planned, piloted or operating?
Data categories and sources Which information enters the system and where does it originate?
Supplier and configuration references Which agreements and settings govern the use?
Output destination and decision Where does the output go and what does it influence?
Human review and escalation Who checks the output and what happens on an exception?
Assessment and approval references Which review supports the current decision?
Next review and change triggers When must the team revisit the use?

Mark unknown information explicitly, assign an owner and set a follow-up action. A completed inventory records the uses identified; it does not establish whether every use has been found or every risk evaluated.

Keep sensitive configurations, prompts, records and contractual material in controlled locations. Reference them from the inventory instead of placing unnecessary personal or confidential data in the tracker.

Separate approval from operating evidence

An approval records who authorized a use and under which conditions. Operating evidence shows whether those conditions remain in place. Retain both and define the evidence needed for each material safeguard.

A condition requiring human review needs an identified reviewer and an operating review process. A restriction on data use needs a documented data flow and supporting configuration or agreement. A supplier assurance document needs a review of its scope and relevance to the service in use.

Record exceptions with an owner, decision and review date. A policy statement alone does not resolve an exception in the deployed service.

Reassess material changes

Set review triggers for new data categories, a customer-facing use, a supplier change, altered access, an expanded user group or a different decision relying on the output. Record the change and the subsequent assessment result.

For example, an internal summarization tool later becomes part of a customer-facing workflow. The new output destination and users require another review even if the product name stays the same. This is an illustrative scenario, not a Cyvelle customer case or a prescribed universal risk classification.

Keep impersonation risks in the right record

An attacker's use of AI is a threat scenario, not an approved system in your organization's inventory. Record suspected voice impersonation in the relevant risk or incident record, with the business process, affected roles and verification procedure it could challenge.

For an unexpected request involving money, access or confidential information, verify the person through a trusted contact route obtained independently of the incoming message or call. The FBI recommends an independent callback when checking a suspected impersonation. [7] Link the scenario to the organization's existing approval and reporting process. The inventory and incident records answer different questions; cross-reference them when an internal AI tool is also involved.

Connect the inventory to governance

The next step is to connect owners, risks, controls and decisions around the inventoried uses. Review ISO 42001 Readiness →for management-system preparation and EU AI Act Readiness → for an obligations-focused scope.

Book an AI Governance Consultation to discuss your business use and the governance question requiring attention.

Common questions

Is a list of AI product names enough?

Record each use as well as the product: its purpose, owner, users, data, supplier, decisions and review evidence. One product may support several uses with different risks and responsibilities.

Does an approved AI use mean its controls are operating?

Approval records a decision. Retain separate evidence of the controls, monitoring and review relevant to that use, and track any conditions or unresolved actions.

Does ISO 42001 certification settle every AI Act obligation?

A management system, a voluntary framework and applicable law serve different purposes. Establish the organization’s role and the requirements applying to the particular activities; use current sources and qualified legal advice where needed.

The inventory fields and change triggers are our practical editorial guidance. They are not a reproduction of ISO or NIST requirements.

Sources

  1. ISO — ISO/IEC 42001:2023 overview

  2. NIST — AI Risk Management Framework

  3. NIST AI 600-1 — Generative Artificial Intelligence Profile

  4. EUR-Lex — Regulation (EU) 2026/1744, Digital Omnibus on AI

  5. European Commission — AI Omnibus enters into force

  6. NIST — AI RMF Core

  7. FBI IC3 — Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud