Practical guide
High-impact use restrictions policy for AI-enabled work
Set a policy that classifies uses by context and impact, names prohibited and approval-gated routes, assigns authority, controls exceptions, and requires evidence-based review and suspension.
Set a policy that classifies uses by context and impact, names prohibited and approval-gated routes, assigns authority, controls exceptions, and requires evidence-based review and suspension.
Write the objective and scope first
Use an observable objective: “Prevent prohibited AI uses and ensure uses with potentially material effects on health, safety, rights, access, or consequential decisions receive the required classification, authority, safeguards, and evidence before operation.”
Define covered systems, models, integrations, vendors, workflows, data, users, geographies, and lifecycle stages. Include pilots, configuration changes, copied workflows, and material changes to intended purpose. Name the policy owner and the people authorized to classify, approve, suspend, and retire uses.
“High impact” is not a universal legal category. The EU AI Act separately identifies prohibited practices and high-risk classifications with detailed scope, conditions, exceptions, operator roles, and application dates. A policy using that source must retain the applicable facts and qualified legal review rather than copying a label. Regulation (EU) 2024/1689
Sources: [1]
Use four policy routes
### Route 1 — Prohibited
List practices the organization will not design, procure, deploy, or use. Each entry needs a source or approved risk rationale, scope, examples, owner, and review trigger. Do not use vague wording such as “harmful AI.” Describe the intended purpose, affected people, decision, data, and behavior that cause the prohibition.
Where law controls, the policy should cite the current jurisdiction-specific provision and retain counsel's applicability decision. Internal prohibitions may be broader than legal minimums, but they should be labelled as organizational choices.
### Route 2 — Restricted and approval-gated
Define uses that may proceed only after enhanced review. Entry conditions might include material influence over essential access, health or safety, rights, financial conditions, education, worker evaluation, or other consequential decisions defined by the organization.
Require a decision record covering:
NIST's AI RMF Core supports documenting context, targeted application scope, impacts, human oversight, internal controls, and responses. It is voluntary risk-management guidance and does not establish legal compliance or make the approval decision. NIST AI RMF Core
### Route 3 — Standard controlled use
Permit lower-impact uses only within approved tools, data rules, human-accountability requirements, source-verification practices, and logging. A standard route is not uncontrolled. Specify what changes would send the use back for classification, such as a new affected population, decision purpose, data class, integration, autonomy level, geography, or vendor.
### Route 4 — Unresolved
When facts or authority are incomplete, pause the proposed use or confine it to an explicitly bounded non-operational test. Assign an owner and due date for classification. “Unresolved” must not quietly become approval through delay.
- intended purpose and foreseeable misuse;
- affected people and possible impacts;
- data categories and provenance;
- system and human roles;
- alternatives, including not using AI;
- evaluation and monitoring plan;
- human oversight and appeal routes;
- security, privacy, accessibility, and vendor review;
- residual risks, authority, and expiry.
Sources: [2]
Use a SCOPE classification record
Require one record for every new or materially changed use:
| Field | Required decision evidence | | --- | --- | | S — System and setting | Model, provider, integration, workflow, users, geography, lifecycle stage | | C — Consequence | Decision influenced, affected people, magnitude, reversibility, appeal | | O — Objective and alternatives | Intended purpose, necessity, non-AI or lower-impact options | | P — Protection and proof | Data, evaluation, oversight, security, monitoring, records | | E — Executive authority and expiry | Classifier, approver, conditions, review date, suspension trigger |
The record must preserve the facts that drove the route, not only the outcome. A copied classification is invalid when the use context or affected decision differs.
Control exceptions and emergency action
An exception must never authorize a prohibited use or bypass applicable law. For an internal restriction that permits exceptions, require scope, rationale, accountable owner, compensating measures, monitoring, expiry, revocation trigger, affected records, and closure evidence.
Allow immediate suspension when a prohibited route is detected, the classification facts were materially wrong, required safeguards fail, an incident threshold is crossed, or governing requirements change. Restoration requires a new recorded decision, not only technical recovery.
Worked policy schedule: synthetic use cases
| Proposed use | Policy route | Required action | Release condition | | --- | --- | --- | --- | | Infer worker emotion from webcam images to score engagement | Prohibited | Reject request; retain rationale and notification | None under this policy; legal review cannot convert an internal prohibition automatically | | AI recommendation materially influences access to an essential customer service | Restricted | Complete SCOPE record, impact and alternatives review, evaluation, oversight, appeal, monitoring, and approval | Named authority approves bounded use and evidence plan | | Draft internal meeting summaries from non-sensitive approved inputs | Standard controlled | Approved tool, participant notice, source check, access controls, retention rule | Owner confirms controls and logging | | New vendor proposes autonomous case closure but cannot explain data lineage | Unresolved | Pause operational use; request lineage, evaluation, oversight, and incident information | Reclassify only after evidence is complete |
The schedule is synthetic and not a legal classification. A real organization would map the applicable jurisdictions, sector rules, contractual duties, and its own risk tolerances with qualified reviewers.
Test and review the policy
Policy approval is design evidence, not proof of operation. At least quarterly, or after material change, reconcile use inventories to classification records; sample approvals and denials; test technical and workflow restrictions; review exceptions and expiry; inspect incidents and bypass attempts; and verify remediation closure.
Report the source population, sample, results, failures, open exceptions, and review date. Do not conclude “effective” from training completion, zero reported incidents, or a policy signature alone.
AI can help compare inventories, surface missing fields, draft scenarios, and summarize records. Accountable people own classification, legal applicability, approval, suspension, and residual risk. Two Prune may assess governance reasoning in a synthetic work simulation; it does not provide legal advice, operate customer controls, certify compliance, or make employment decisions.
Sources
European Union: Regulation (EU) 2024/1689 (AI Act)
- European Union: AI Act official PDF
- NIST: AI RMF Core
- NIST: AI RMF Playbook
- NIST: AI RMF Playbook FAQ
Sources: [1] [4] [2] [3] [5]
Sources
- 1.Regulation (EU) 2024/1689 · eur-lex.europa.eu
- 2.NIST AI RMF Core · National Institute of Standards and Technology
- 3.NIST AI RMF Playbook · National Institute of Standards and Technology
- 4.European Union: AI Act official PDF · eur-lex.europa.eu
- 5.NIST: AI RMF Playbook FAQ · National Institute of Standards and Technology