Practical guide
Acceptable AI use policy template
This template gives an organization a usable starting point for everyday AI use. It is designed to be completed with legal, privacy, security, records, technology, accessibility, and operational owners. It is not a universal policy, legal advice, a compliance certificate, or a substitute for controls in a specific workflow. The governing idea is simple: permission follows the use case, data,.
This template gives an organization a usable starting point for everyday AI use. It is designed to be completed with legal, privacy, security, records, technology, accessibility, and operational owners. It is not a universal policy, legal advice, a compliance certificate, or a substitute for controls in a specific workflow. The governing idea is simple: permission follows the use case, data, action, and consequence—not a brand name alone. A tool that is acceptable for drafting public event copy may be unacceptable for restricted customer records or an irreversible action.
Completion instructions
Before issuing the policy, create an inventory of approved AI services and representative use cases. For each use case, record purpose, users, data classes, system access, output recipient, action authority, consequence, required review, retained evidence, and escalation route. Have the relevant control owners approve the result.
Replace every bracketed field. Delete examples that do not apply. Link terms such as “confidential,” “personal,” “high consequence,” and “approved tool” to existing organizational definitions. Publish a version, owner, effective date, review date, and change log. Test the policy on realistic cases before asking people to follow it.
Copyable policy
### 1. Purpose and scope
[Organization] permits responsible use of approved artificial intelligence services when the use supports an authorized work purpose and follows this policy and all applicable security, privacy, legal, records, accessibility, intellectual-property, professional, and operational requirements.
This policy applies to [people, contractors, systems, locations, and business units] using AI to generate, transform, classify, summarize, recommend, retrieve, predict, or take actions. It applies whether the AI is a standalone service, embedded in another product, accessed through an API, or configured as an agent.
The policy owner is [role]. Questions and exceptions go to [channel or role]. The policy does not transfer accountability from the person or function authorized to perform the work.
### 2. Approved tools and configurations
Users may use only the tools, accounts, plugins, integrations, models, and configurations listed in [approved-service register] for the stated use cases and data classes. Approval of one configuration does not approve every feature, connector, model, or future version.
Before enabling a new tool or material change, [technology owner], [security owner], [privacy or legal owner], and the relevant [workflow owner] must review the purpose, data flows, access, retention, supplier terms, system actions, testing, monitoring, exit, and incident route in proportion to risk.
The UK Government's current AI Playbook recommends understanding limitations, using AI lawfully and securely, maintaining meaningful human control, managing the lifecycle, choosing the right tool, obtaining relevant expertise, and applying organization-specific policy and assurance. Its rules bind its stated government context; they are useful prompts rather than private-sector compliance conclusions. Artificial Intelligence Playbook for the UK Government
### 3. Data that may and may not be used
Users must classify information before entering, uploading, connecting, or exposing it to an AI service.
Use the minimum information necessary. Do not assume that redaction, pseudonymization, an enterprise account, or a supplier statement makes a use lawful or safe. Follow [data-handling standard] and ask [privacy/security contact] when classification is uncertain.
### 4. Permitted and prohibited uses
Permitted uses are listed in [use-case register], including their required reviewer and evidence. Typical bounded examples may include brainstorming from public material, formatting user-authored text, summarizing approved internal documents for an authorized audience, generating test data, or drafting code that receives normal review and testing.
Users must not use AI to:
### 5. Human accountability and verification
The person submitting or acting on AI-assisted work remains responsible for meeting the same quality, evidence, and control requirements as other work. Review must be meaningful: the reviewer needs adequate skill, time, source access, authority to change or reject the output, and a defined check.
For material claims, calculations, citations, instructions, code, classifications, or recommendations, reviewers must inspect authoritative sources, test calculations or behavior, check omissions and conflicts, consider affected people, and record uncertainty. High-consequence use cases require the named human decision and intervention points in [control register].
NIST's AI RMF Core calls for documented context, roles, third-party risks, operator proficiency, human oversight, measurement, monitoring, and response. It is voluntary guidance and does not certify a policy. NIST AI RMF Core
### 6. Transparency, records, and attribution
Disclose AI assistance when required by law, contract, professional practice, customer commitment, publication standard, or [organization rule]. Do not claim AI output as independently verified work.
Retain [prompt, input, output, model or service version, source set, review, decision, and approval] for [use cases and retention period] where necessary for traceability. Do not retain sensitive content merely because a log is technically available. Follow the records schedule and access controls.
### 7. Incidents, uncertainty, and escalation
Stop and report through [incident route] when AI use may have exposed restricted data, produced harmful or deceptive content, taken an unauthorized action, bypassed a control, created a material error, or behaved outside the approved scope. Preserve relevant evidence without spreading the content further.
When instructions conflict or the user cannot verify a material output, mark it unresolved and escalate to [owner]. No one is required to accept AI output because a tool appears confident or because a deadline is close.
### 8. Monitoring, review, and exceptions
[Policy owner] reviews this policy at least [cadence] and after material changes to law, tool, model, supplier terms, data access, workflow, incident pattern, or consequence. Monitoring covers use cases, exceptions, incidents, control performance, accessibility, user feedback, and unexpected effects—not only usage volume.
Exceptions require [request record], a defined scope and duration, named approvers, compensating controls, retained rationale, and expiry. Emergency approval does not become permanent approval by default.
- Allowed without additional approval: [public or approved low-sensitivity examples] in an approved service.
- Allowed only under a documented control: [internal, personal, confidential, regulated, or client examples], using [approved configuration, minimization, access, retention, location, and agreement].
- Prohibited: [secrets, credentials, classified information, unapproved personal or special-category data, embargoed transactions, privileged material, or other local categories].
- bypass an approval, control, professional duty, contractual restriction, or access boundary;
- fabricate or obscure sources, actions, authorship, or test results;
- treat model output as verified fact, legal authority, policy, or source evidence;
- make an irreversible or externally consequential action without the authorization specified for that use case;
- impersonate a person or create deceptive material;
- conduct surveillance, profiling, or high-consequence decision-making outside an approved process; or
- connect a tool, plugin, or agent to systems or data beyond its approved scope.
Sources: [1] [2]
Worked example: customer-support summary
A support team wants an approved assistant to draft summaries of customer conversations for internal handoff. The record contains names, account details, and free-text complaints.
| Policy question | Completed answer | | --- | --- | | Purpose | Draft an internal summary; do not classify safety issues or send a customer response | | Approved service | Enterprise assistant in the configuration listed in register AI-03 | | Data | Minimized conversation text; payment data and identity documents excluded; access limited to the case team | | Human control | Case owner compares the draft with the original conversation and retains escalation authority | | Verification | Check requested action, dates, product, unresolved issue, quoted commitments, and safety cues | | Records | Retain final edited summary and review marker under the case schedule; do not copy raw prompts into a separate analytics store | | Stop condition | Any sign of restricted data leakage, invented commitment, or missed safety cue pauses the use case and triggers the incident route | | Boundary | The tool does not update the customer record or send messages automatically |
The use is acceptable only under those conditions. “Customer-support summarization” is not approved as an unrestricted category, and approval does not show that the tool is accurate across other languages, products, or configurations.
Practitioner quality check
Before publication, test the draft policy with at least these cases:
For each case, two readers should reach the same permission, control, owner, and escalation outcome. If they cannot, repair the definition or procedure. Verify that the technical configuration matches the written policy; prose cannot prevent a system from accessing data or taking actions.
- public brainstorming in an approved tool;
- an employee pasting personal or confidential data;
- an embedded assistant with broad document access;
- AI-drafted analysis containing a false citation;
- an agent proposing to send or modify a record;
- a plugin or model version not listed in the register;
- an urgent request where the normal reviewer is unavailable; and
- a user needing an accessible alternative or support.
Version history template
| Version | Effective date | Owner | Material change | Evidence or trigger | Approvers | Next review | | --- | --- | --- | --- | --- | --- | --- | | 1.0 | [date] | [role] | Initial policy | Inventory and scenario test | [roles] | [date] |
How this fits a Two Prune assessment
A Two Prune assessment can ask a participant to apply a supplied acceptable-use policy to synthetic workplace materials and produce a controlled deliverable. The assessment can observe data handling, tool choice, source verification, uncertainty, escalation, AI interaction, and revision. It does not certify an organization's policy, provide legal advice, monitor live systems, or make employment decisions.
Sources and limits
The template and worked example are original. Obtain jurisdiction- and sector-specific review before adoption. Version 1.0, prepared 15 November 2026.
- GOV.UK: Artificial Intelligence Playbook for the UK Government — current public-sector guidance on secure, lawful, controlled, skilled, and assured AI use.
- NIST AI RMF Core — voluntary outcomes for roles, context, third-party risk, oversight, documentation, measurement, and monitoring.
- ICO: Governance and accountability in AI — data-protection governance prompts, including policies supported by operational procedures.
Sources: [1] [2] [3]
Sources
- 1.Artificial Intelligence Playbook for the UK Government · UK Government
- 2.NIST AI RMF Core · National Institute of Standards and Technology
- 3.ICO: Governance and accountability in AI · Information Commissioner's Office