How to Write an AI Policy (Template)
This content was developed with AI assistance and is regularly reviewed for accuracy.
Snapshot: October 2026. Model names, prices, and product features change monthly; What's New in AI is the maintained chronology. Regulatory dates below are as reported in October 2026; confirm them with counsel before you publish a policy that cites them.
Every organization using AI needs a short, enforceable policy that protects the business and tells people what they are allowed to do. In 2026 that policy has to cover more than a chat window: assistants now read shared folders, connect to email and calendars, run scheduled tasks while nobody is watching, and act as agents inside products you already pay for. A policy written for "don't paste secrets into ChatGPT" leaves most of today's risk uncovered.
This page walks through the eight sections a complete acceptable use policy needs, gives you template language to adapt, and includes the data classification table most policies skip until something goes wrong. It is the written rulebook that Govern: Frameworks, Policy, and Compliance assumes you have.
Learning focus:
- Explain to leadership why a policy speeds adoption instead of slowing it
- Draft all eight sections from adaptable template language
- Classify data into four tiers and map each tier to AI rules, including connectors and agents
- Roll the policy out with training, a grace period, and a review cadence
Why every organization needs an AI policy
Your people already use AI. About half of US employees used it at work by mid-2026 and 15% used it daily (Gallup, Q2 2026, reported), often with personal accounts and no guidance. Without a policy you are exposed in four ways.
Data leakage is the most immediate. Employees who paste customer records, financial data, or proprietary code into consumer AI tools may be sharing it with a third party under terms nobody at your company reviewed. Samsung's 2023 source-code exposure through ChatGPT is still the standard example. In 2026 the same leak can happen without any pasting: a connector that grants an assistant read access to a whole shared drive exposes everything in it.
Inconsistent use creates quality and brand risk. When one team uses AI for customer replies and another bans it, you get uneven quality, conflicting messages, and no way to audit what was generated.
Unbounded actions are the new category. An agent with write access can send, delete, purchase, or publish. The April 2026 PocketOS incident, in which a coding agent deleted a production database and its backups in about nine seconds using an over-privileged key it found in an unrelated file, involved no attacker. It was a permissions failure, and a policy is where permissions start.
Compliance and liability round it out. The EU AI Act's transparency obligations under Article 50 (telling people they are interacting with AI or seeing AI-generated content) are reported to have applied from August 2, 2026, with the machine-readable marking requirement for systems already on the market reported as moved to February 2027. The Act's high-risk obligations, originally due August 2, 2026, were delayed by the Digital Omnibus on AI, which reached political agreement on May 7, 2026: stand-alone high-risk systems are reported to apply from December 2, 2027 and AI embedded in regulated products from August 2, 2028. One July 2026 commentary cites the amending act as Regulation (EU) 2026/1744; that number was not verified against the Official Journal. Vendor terms add obligations of their own: Anthropic's usage policy effective November 12, 2026 requires a qualified human reviewer and notice to the affected person where AI use affects health, legal rights, finances, livelihood, or essential services. Because these dates have already moved once, cite "applicable law" in the policy text and keep the dates in a separate, dated compliance note.
A good policy does not slow adoption. It is what lets cautious people start, because they finally know what is allowed.
The eight sections
A complete AI acceptable use policy covers eight areas, each building on the last.
| Section | What it settles |
|---|---|
| 1. Purpose and scope | What the policy covers and who must follow it |
| 2. Approved tools, connectors, and agent features | What people may use, and with which access |
| 3. Acceptable use | Encouraged, conditional, and prohibited uses, including what agents may do |
| 4. Data classification for AI use | What information may go into which tool or connector |
| 5. Disclosure | When to say AI was involved |
| 6. Quality and accountability | Who checks the output and who owns it |
| 7. Training | How people learn the rules, and the record that proves it |
| 8. Review and updates | How the policy stays current |
Replace every bracketed placeholder in the template language below with your own details, and match the tone of your existing policies.
Template: AI acceptable use policy
Section 1: Purpose and scope
This section frames everything else. Err on the side of inclusion: AI is now embedded in email, documents, browsers, and coding tools, and the policy should cover those ambient features as well as standalone assistants.
1.1 Purpose
This policy sets out how [Organization Name] employees, contractors, and authorized third parties may use artificial intelligence tools and services. It is designed to enable productive use of AI while protecting organizational data, maintaining quality, and complying with applicable laws, regulations, and vendor terms.
1.2 Scope
This policy applies to all employees regardless of department or seniority; contractors, consultants, and temporary staff with access to organizational systems or data; and third parties who process organizational data using AI.
It covers general-purpose AI assistants (for example Claude, ChatGPT, Gemini, and Microsoft Copilot); AI features embedded in existing software; connectors and integrations that give an AI tool access to organizational systems such as email, calendars, file storage, chat, or databases; agents, scheduled tasks, and automations that act without a person approving each step; and any custom AI application built by or for [Organization Name].
1.3 Effective date and authority
This policy is effective as of [Date], approved by [Approving Authority], and supersedes prior guidance on AI use. Violations may result in disciplinary action.
Section 2: Approved tools, connectors, and agent features
An approved list prevents shadow AI and gives IT and security visibility. In 2026 the list needs a column for connectors and their scopes, because the same assistant is low-risk with no connectors and high-risk with write access to your file system.
2.1 Approved tools
[Organization Name] maintains an approved AI tools list at [internal link]. Only listed tools, on organization-managed accounts, may be used for organizational work.
| Tool | Approved uses | Highest data class allowed | Approved connectors and scopes | Agent features allowed |
|---|---|---|---|---|
| [Tool 1] | [e.g., drafting, research, summarization] | [e.g., Internal] | [e.g., Drive: read-only, Team folder only] | [e.g., none] |
| [Tool 2] | [e.g., code assistance] | [e.g., Confidential with approval] | [e.g., GitHub: one repository, no admin] | [e.g., agent mode on branches; merges by humans] |
| [Tool 3] | [e.g., weekly reporting] | [e.g., Internal] | [e.g., Sheets: read-only] | [e.g., one scheduled task, owner named] |
2.2 Requesting a new tool, connector, or agent feature
Submit requests to [designated team] using the AI Tool Request Form. Requests are evaluated on the provider's data handling and retention terms, security certifications, the scopes the connector requires, what actions the agent can take and whether they can be undone, cost and licensing, and business justification. Evaluation will be completed within [timeframe, e.g., 15 business days].
Review the list quarterly at minimum. Vendors change retention terms, add agent features, and rename products often enough that a tool approved six months ago may not be the same tool today. Evaluate and Measure gives you the scoring method for new requests.
Section 3: Acceptable use
This is the operational heart of the policy. Three tiers (encouraged, conditional, prohibited) let people place their intended use in seconds instead of reading a long list of rules. The fourth subsection covers what is new in 2026: work that an agent does on someone's behalf.
3.1 Encouraged uses (no approval needed)
Drafting and editing internal communications, reviewed before sending; brainstorming and planning; summarizing public research; first drafts of non-sensitive documents such as agendas and project plans; learning unfamiliar topics.
3.2 Conditional uses (manager approval required)
Customer-facing content, including marketing, support replies, and product documentation; analysis that informs decisions above [budget threshold]; connecting an AI tool to a new organizational system; processing Internal data (see Section 4); any scheduled task or automation that runs without a person present.
3.3 Prohibited uses
Entering Confidential or Restricted data into any AI tool without written approval from [CISO or equivalent]; generating contracts, regulatory filings, or compliance documents without legal review; presenting AI-generated content as original human work where disclosure is required; using AI to make or inform employment decisions without HR involvement and human review; surveillance of employees or customers; bypassing security controls to reach unapproved tools.
3.4 Agents, connectors, and automated actions
When an AI tool can act on your behalf, the person who set it up is accountable for what it does. Grant the narrowest scopes that do the job, and prefer read-only. Do not store credentials or access keys in files an agent can read. Any action that sends external communications, spends money, deletes or overwrites data, changes permissions, or touches production systems requires human approval before it happens. Every scheduled task or unattended agent must have a named owner, a written purpose, and a spending or usage limit, and must be listed in the AI inventory. Do not point browser or computer-use agents at banking, legal, medical, or HR systems.
Section 3.4 is the clause most 2024-era policies lack. It encodes the lesson vendors themselves now teach: contain the agent first, then rely on approvals. Approvals alone wear thin; Anthropic found users approving about 93% of permission prompts in its coding agent before it introduced a classifier-based auto mode. Agent Safety and Guardrails explains the controls behind each sentence.
Section 4: Data classification for AI use
Data classification is the single most important safeguard in the policy. Without it, people have no way to decide what may go into which tool, or which folders a connector may reach.
| Classification | Description | AI usage rules | Examples |
|---|---|---|---|
| Public | Intended for public consumption or already public | Any approved tool | Published posts, press releases, public product specs, open-source code |
| Internal | For internal use; limited harm if disclosed | Approved enterprise tools with a data processing agreement; manager approval; connectors read-only and scoped to the relevant folder | Internal memos, meeting notes, project timelines, training materials |
| Confidential | Disclosure could cause material harm | CISO approval before any AI processing; enterprise tools with contractual data protections only; no consumer or free-tier accounts; no agent write access | Customer lists, forecasts, unreleased plans, compensation data |
| Restricted | Protected by law, regulation, or contract | Never sent to external AI tools; only internally hosted systems with appropriate controls; never in scope for any connector | Health records, government IDs, payment card data, trade secrets, material under NDA or legal hold |
The rules matter because AI tools handle data in ways that are not obvious. Consumer tiers may use conversations for model training unless the user opts out, and retention periods differ sharply between consumer and enterprise plans of the same product. Connectors add a second path: a file never pasted into a chat can still be read if it sits in a folder the assistant can reach.
Before using data with AI, people should answer three questions. What class is this data? Is this tool, and this connector, approved for that class? Have I obtained the approvals the class requires? When unsure, treat the data as Confidential and ask. The AI Data Privacy and Safety Checklist turns these rules into a one-page routine.
Section 5: Disclosure
Transparency protects credibility and keeps you aligned with transparency rules such as Article 50. Expectations vary by context, so the template sets four.
5.1 Internal documents
Work shared internally should note substantial AI assistance ("Drafted with AI assistance") in a footnote, header, or metadata. Minor editing and formatting do not need a note.
5.2 External communications
Customer-facing, partner-facing, or public content substantially generated or informed by AI must carry appropriate disclosure, following applicable law and industry norms. People interacting with an AI system on our behalf, such as a support chatbot, must be told they are dealing with AI. When in doubt, disclose.
5.3 Regulated and high-stakes outputs
Content subject to regulatory oversight, and any AI-assisted recommendation that affects a person's health, legal rights, finances, employment, or access to essential services, must follow the disclosure and human-review requirements in applicable law and in our vendors' usage policies. [Legal/Compliance] maintains the current list.
5.4 Code and technical work
AI-generated or AI-assisted code must be noted in commit messages, pull request descriptions, or project documentation.
Reference "applicable law" in the policy rather than naming statutes, and keep the named statutes in the dated compliance note so the policy does not go stale every time a deadline moves.
Section 6: Quality and accountability
AI tools produce confident output that can be wrong, outdated, or subtly biased, and agents can report success on work that did not succeed. The policy needs a review rule and an ownership rule.
6.1 Human review
A qualified person reviews AI-generated content before it is finalized, shared, or published, checking factual accuracy, consistency with our voice and standards, absence of bias, and correct handling of any data. For agent and scheduled-task output, the reviewer checks the result itself, not the tool's status report.
6.2 Fact-checking
Statistics, citations, legal or historical claims, and technical specifications in AI output must be verified against authoritative sources. AI tools can fabricate references.
6.3 Accountability
The person who submits, publishes, or distributes AI-assisted work, or who configured the agent that produced it, is responsible for its accuracy and appropriateness. Using AI does not transfer or reduce that responsibility.
Section 6.3 prevents "the AI did it." Courts have already sanctioned professionals for filing AI-fabricated citations, and the person who signed the filing carried the penalty, not the tool.
Section 7: Training
A policy nobody understands is a policy nobody follows. Training is also the evidence a regulator looks for: under Article 4 of the EU AI Act, organizations in scope must take measures to build AI literacy among staff who operate AI systems, and the record of who learned what is what proves it.
7.1 Required training
All staff complete the following within [30 days] of this policy's effective date or their start date: an overview of this policy; data classification for AI use; safe use of connectors and agents; and role-specific training from their department.
7.2 Ongoing education
Staff who use AI tools [weekly or more] complete at least [2 hours] of AI-related learning each year, from vendor academies, internal sessions, or approved courses.
7.3 Records
Completion is recorded in [learning system]. Managers ensure their reports meet requirements, and completion is reviewed at [quarterly check-ins].
Make the training modular and role-specific: a marketer, an engineer, and a finance analyst meet different situations. Build an AI Literacy Program shows how to design it and keep the record.
Section 8: Review and updates
AI products change faster than most policy cycles. Build in a schedule and a set of triggers.
8.1 Scheduled reviews
This policy is reviewed at least [every six months] by [AI owner or steering group], with input from IT Security, Legal, HR, and business units.
8.2 Triggering events
The policy is also reviewed promptly after a significant AI-related incident; new law or regulation affecting our AI use; material changes to an approved vendor's terms, retention, or agent features; or adoption of a new AI tool or agent capability at scale.
8.3 Proposing changes
Anyone may propose a change via [channel], stating the change, the reason, and any supporting information.
8.4 Version control
All versions are archived with effective dates, change summaries, and approvals. The current version is available at [location].
Six months is a sensible default. In a regulated industry, or in your first year of heavy agent use, review quarterly.
Implementation checklist
Drafting is half the work. Roll the policy out deliberately.
- Draft with input from IT, Legal, HR, Security, and business units
- Complete legal review against applicable law and employment rules
- Secure leadership approval and a named executive sponsor
- Publish the approved list with connector scopes and agent permissions
- Communicate through several channels and deliver the required training
Then set up a channel for questions and change proposals, and put the first review date on the governance calendar. Give people a 30- to 60-day grace period with active support before strict enforcement, so the policy reads as a guide rather than a trap. After that, enforce consistently, and design the incident path so that reporting a mistake is never punished; the checklist covers what to do when data goes somewhere it should not.
Key takeaways
- Every organization using AI needs a policy, and in 2026 it must cover connectors, agents, and scheduled tasks, not only chat
- Data classification is the linchpin; map each tier to tools and to connector scopes
- Encouraged, conditional, and prohibited tiers give people fast, usable guidance; a separate clause governs what agents may do without approval
- Accountability stays with the person who publishes the work or configured the agent
- Cite "applicable law" in the policy and keep moving regulatory dates, such as the Digital Omnibus deadlines of December 2027 and August 2028, in a dated compliance note
Next steps
Continue to: AI Data Privacy and Safety Checklist
- Need the governance structure the policy sits inside? Read Govern: Frameworks, Policy, and Compliance
- Writing the agent clause and want the technical controls behind it? Read Agent Safety and Guardrails
- Deciding whether a new tool belongs on the approved list? Use Evaluate and Measure: Tools, ROI, Impact