AI use policy template: one page your team will actually read.
Three in four knowledge workers already use AI at work and most bring their own tools; 57% hide it from their employer. A generative AI policy is not about banning ChatGPT — it is about deciding, in writing, which tools, which data, who signs off and who you call when it goes wrong. Below: a free editable Word template, what each section is for, and the facts behind why an SMB needs one this quarter.
- Free .docx download
- Mapped to NIST AI RMF & ISO/IEC 42001
- Covers EU AI Act Articles 4 and 50
- Not legal advice
Statistics verified against primary sources: October 2026. Have counsel review the policy before it becomes binding on staff.
of employees who use AI at work bring their own tools — 80% at small and medium companies (Microsoft & LinkedIn Work Trend Index, 2024).
of employees say they hide their AI use; only 40% say their workplace has any gen-AI policy or guidance (KPMG / University of Melbourne, 47 countries, 2025).
organisations reported a breach caused by shadow AI; high shadow-AI use added $670,000 to the average breach cost (IBM Cost of a Data Breach, 2025).
of organisations had a formal, comprehensive AI policy in 2026 — up from 15% in 2024 (ISACA AI Pulse Poll).
Download the AI use policy template
Twelve short sections, bracketed fields to replace, a sign-off line. Written for a company of 5 to 250 people that uses ChatGPT, Copilot, Claude or Gemini and may have a chatbot on its website. Adapt it in an afternoon, review it every six months.
Download the template ↓Word .docx · 12 sections · version 1.0, October 2026 · free, no email required
Prefer we fill it in with you? Start with the free AI readiness assessment
What the twelve sections do, and where they come from
Each clause maps to a control in NIST’s AI Risk Management Framework, its Generative AI Profile (NIST AI 600-1) or ISO/IEC 42001 — so the one-pager stands up when a client, auditor or regulator asks what it is based on.
1–2 · Purpose and scope
Why the policy exists and what counts as an AI tool — general assistants, AI features inside tools you already pay for, and chatbots on your own website. Tool-agnostic on purpose: SHRM’s 2026 survey found 54% of organisations with policies consider them too tied to specific tools.
3 · Approved tools and how to add one
NIST AI 600-1 action GV-6.1-007 says it plainly: “establish approved GAI technology and service provider lists”. A table of tool, approved use, data allowed and account type — plus a one-email route to add a tool, because the alternative is people adding them silently.
4 · Data rules
What never goes into a public tool: customer personal data, credentials, NDA’d code, unreleased financials. The UK NCSC’s 2023 advice still holds: do not submit queries that would cause a problem if they became public. Cyberhaven measured 34.8% of data shared with AI tools as sensitive in 2025, up from 10.7% two years earlier.
5 · Disclosure and customer-facing AI
Chatbots tell customers they are AI at first contact, AI-generated media is marked. In the EU this is Article 50 of the AI Act, in force since August 2, 2026; NIST GV-5.1-002 recommends the same. The clause keeps you honest with customers and compliant in one sentence.
EU AI Act Article 50 guide →6 · Human review and accountability
A named person is accountable for every AI-assisted decision; decisions affecting people are never made by AI alone. Harvard’s staff policy phrases it as remaining “fully responsible” for AI-assisted work. The court in Mata v. Avianca called it lawyers’ “gatekeeping role” — and fined them $5,000 for skipping it.
7 · AI inventory
NIST GOVERN 1.6: mechanisms to inventory AI systems. A living list of every tool, who uses it, what it sees and the vendor’s data terms. Gartner found 69% of organisations suspect or know staff use prohibited public AI tools; the inventory turns suspicion into a list.
8 · Risk assessment for new use cases
Six written questions before AI is used for something new: what could go wrong, who could be harmed, what data, is it regulated, how would we know it failed, how do we roll back. Low-risk uses get a yes in a day. This is NIST’s Map function at SMB scale.
9 · Training and AI literacy
Everyone who uses AI tools completes a short induction on the policy and on how the tools fail — hallucination, bias, prompt injection. For EU users this meets Article 4 of the AI Act (since February 2025, enforced since August 2026): the Commission says no mandatory course or certificate, just proportionate measures and an internal record.
10 · Incidents
Report a suspected leak or a harmful output that reached a customer within 24 hours, and reporting is never penalised. NIST GOVERN 4.3 and ISO/IEC 42001 control A.3.3 both require a channel for concerns; IBM found 63% of breached organisations had no finished AI governance policy.
11 · EU AI Act notes
Transparency duties since August 2, 2026; prohibited practices and literacy since February 2025; high-risk obligations deferred to December 2027 for recruitment, credit and similar uses. The SME measures lower the fine cap — they do not exempt anyone from Article 50.
12 · Review
Every six months and whenever a new tool or use case is adopted. ISO/IEC 42001 control A.2.4 is literally “review of the AI policy”; NIST GOVERN 1.5 asks you to decide the frequency. Six months is our synthesis, not a legal minimum — the tool list changes faster than the policy.
What the numbers say about unmanaged AI use
The case for a policy is not hypothetical. These are the studies we would cite to a sceptical owner, with their dates.
Usage ran ahead of policy years ago.
Microsoft and LinkedIn surveyed 31,000 workers in early 2024: 75% used AI at work, 78% of them with tools they brought themselves. Gallup put US usage at 40% by mid-2025 — with only 30% of employees saying their organisation had guidelines or a formal policy.
People hide it, and they skip the checking.
KPMG and the University of Melbourne (48,000 people, 47 countries): 57% hide their AI use and present AI output as their own, 66% rely on output without evaluating accuracy, 56% have made mistakes because of it. A policy that says “tell us, and check it” fixes more than a ban would.
The data leaves in pieces.
Harmonic Security analysed 22.5 million enterprise prompts and uploads in 2025 and found 579,000 instances of sensitive data, 71% of it through ChatGPT and 17% through personal accounts. Netskope’s 2026 report counted an average of 223 monthly attempts per organisation to put sensitive data into generative AI tools — more than double the year before.
It shows up in breach costs.
IBM’s 2025 Cost of a Data Breach: one in five organisations reported a breach due to shadow AI, and those with high shadow-AI use paid $670,000 more per breach on average. 63% of breached organisations either had no AI governance policy or were still writing one.
Policies are spreading, mostly at large companies.
Littler’s 2026 employer survey found 68% with a formal AI policy, up from 38% a year earlier — but the sample skews large, and only 55% had a formal process for approving tools. ISACA’s broader poll puts comprehensive policies at 38%. For SMBs the honest estimate is lower still.
Regulators now expect a paper trail.
The EU AI-literacy duty is enforceable since August 2026; the Commission’s guidance is to keep an internal record of measures. US states added their own: Texas’s AI law gives a safe harbour for substantial compliance with the NIST framework, Illinois requires notice when AI is used in employment decisions, Colorado’s rewritten law starts January 2027.
Does the vendor train on your data? Depends on the account.
The single most common gap in SMB policies: staff on personal accounts. Vendor defaults as published in their own documentation, checked October 2026 — confirm against the current terms before relying on them.
| Vendor | Personal / consumer plan | Business, Enterprise or API |
|---|---|---|
| OpenAI (ChatGPT) | Conversations may be used to improve models unless the user turns off “Improve the model for everyone”; Temporary Chat is excluded. | Business, Enterprise, Edu and API: not used for training by default; API data retained up to 30 days for abuse monitoring. |
| Anthropic (Claude) | Free, Pro and Max: opt-out model since September 2025, with five-year retention if opted in. | Team, Enterprise, API, Bedrock and Vertex AI: not used for training under the commercial terms. |
| Google (Gemini API) | Unpaid tier (AI Studio, free quota): Google uses content to improve products and human reviewers may read prompts. | Paid services: prompts and responses are not used to improve products; brief logging for abuse detection. |
| Microsoft (Copilot) | Consumer Copilot: separate consumer terms. | Copilot with a work account: prompts, responses and Graph data are not used to train foundation models; inherits sensitivity labels and retention. |
Summarised from each vendor’s published data-usage pages. Policies change; the template’s approved-tools table has a column for exactly this.
What happens without a policy
Each of these would have been stopped by one sentence in the template.
Samsung, April 2023
Engineers entered sensitive internal data, reportedly including source code, into ChatGPT. Samsung banned generative AI tools company-wide within weeks. The data rule in section 4 is the whole lesson.
Mata v. Avianca, June 2023
Lawyers filed a brief citing cases ChatGPT invented and were sanctioned $5,000. The database of court decisions involving AI-hallucinated material passed 2,200 cases in October 2026. Section 6: a human checks before it leaves the building.
Air Canada, February 2024
The airline’s chatbot promised a bereavement refund that did not exist; a tribunal held the company liable and rejected the argument that the chatbot was “a separate legal entity”. Section 5: your bot’s words are your words.
Deloitte Australia, October 2025
A A$440,000 government report contained fabricated references and an invented judicial quote; Deloitte disclosed its use of GPT-4o and repaid the final instalment. Section 6 again — and section 8, for a use case that deserved a risk check.
Trade secrets, 2026
US courts are beginning to treat information pasted into a public AI tool as voluntarily disclosed, which can defeat a trade-secret claim. Whether or not your jurisdiction follows, section 4 removes the question.
The quiet one
No headline: a sales rep uploads the customer list to a free tool on a personal account to “clean it up”. Harmonic found 17% of sensitive exposures go through personal accounts. Section 3’s account-type column exists for this.
From template to policy in a week
The template is the easy part. These five steps are what makes it real, in the order that avoids rework.
- 01
Name the owner
One person, not a committee — the policy, the inventory and the approvals route to them. In a 20-person company this is usually the founder or the operations lead, not IT.
- 02
Ask, do not audit
A two-question survey: which AI tools do you use for work, and for what. Promise no consequences for honest answers — 57% of people hide their use, and you need the list more than you need the moral high ground.
- 03
Fill the approved-tools table
For each tool people actually use: approved, approved with limits, or replaced by a business-plan equivalent that does not train on your data. Most SMBs end up with three to five tools and a rule about personal accounts.
- 04
Run the 45-minute induction
Walk through the policy, show one hallucination and one prompt-injection example, record who attended. That record is your EU AI-literacy evidence and your defence when something goes wrong anyway.
- 05
Put the review date in the calendar
Six months out. The policy will survive; the tool list will not. If you build a customer-facing bot before then, that is a new use case and triggers section 8 immediately.
Governance is one of six dimensions.
The free AI readiness assessment scores your policy, inventory and AI-literacy evidence alongside use cases, data, skills, infrastructure and EU AI Act exposure — and tells you what to fix first.
Get the free AI readiness assessmentAI readiness checklist·EU AI Act Article 50 guide·AI maturity model explained
Questions owners ask
- What is an AI use policy?
- A short internal document that says which AI tools staff may use, with which data, who is accountable for AI-assisted work, how customers are told when they are dealing with AI, and what to do when something goes wrong. A generative AI policy is the same thing scoped to tools like ChatGPT, Copilot, Claude and Gemini. NIST’s Generative AI Profile recommends “transparent acceptable use policies” as a baseline governance control.
- Why not just ban ChatGPT?
- Because the ban does not work and the data says so: 78% of AI users bring their own tools, 57% hide their use, and Gartner found 69% of organisations suspect staff use prohibited tools. The UK’s National Cyber Security Centre advised in 2023 to let people experiment in ways that do not put organisational data at risk. A policy with an approved-tools list and a personal-accounts rule achieves what a ban pretends to.
- Is this template legally sufficient?
- It is a sound starting point mapped to NIST and ISO/IEC 42001 controls and to the EU AI Act duties that touch SMBs. It is not legal advice and it does not know your jurisdiction, sector rules or employment contracts. Have counsel review it before it becomes binding — most lawyers will spend an hour on a one-pager and a day on a forty-page one, which is the second reason it is short.
- Does the EU AI Act require an AI policy?
- Not in those words. It requires measures to support AI literacy (Article 4, since February 2025, enforced since August 2026) and, for AI that interacts with people or generates content, transparency (Article 50, since August 2026). The Commission says there is no mandatory training or certificate and recommends keeping an internal record. A policy plus an induction log is the proportionate way to show both.
- Who owns the content AI generates for us?
- Check the vendor terms for the contractual answer — most business plans assign outputs to you. The copyright answer is narrower: the US Copyright Office (January 2025) says prompts alone do not create protectable authorship, while human selection, arrangement and modification of AI output can. Treat AI drafts as raw material a person shapes, and say so in the policy.
- How often should we review it?
- The policy every six to twelve months; the approved-tools table whenever a tool is added, which in practice is quarterly. ISO/IEC 42001 asks for review “at planned intervals” and NIST asks you to set the frequency; no body mandates a number. We put six months in the template because the vendor data-training defaults in the table above changed three times in 2025 alone.
Sources: NIST AI Risk Management Framework 1.0 — Govern playbook · NIST AI 600-1 — Generative AI Profile (Jul 2024) · ISO/IEC 42001:2023 — AI management systems · Microsoft & LinkedIn Work Trend Index 2024 · KPMG / University of Melbourne: Trust, attitudes and use of AI (2025) · IBM Cost of a Data Breach 2025 · ISACA AI Pulse Poll 2026 · ISACA AI Pulse Poll 2025 · Cyberhaven 2025 AI Adoption & Risk Report · Harmonic Security: 22 million enterprise prompts (Jan 2026) · Gartner on shadow AI (via ITPro, Nov 2025) · Gallup: AI use at work nearly doubled (Jun 2025) · Littler Annual Employer Survey 2026 · UK NCSC: ChatGPT and large language models (Mar 2023) · European Commission: AI literacy Q&A · Harvard FAS staff policy on generative AI · Mata v. Avianca sanctions order (LawNext, Jun 2023) · AI Hallucination Cases database (Charlotin) · Moffatt v. Air Canada (CTV News, Feb 2024) · Deloitte refund to Australian government (Accounting Times, Oct 2025) · TechCrunch: Samsung bans generative AI (May 2023) · US Copyright Office: Copyright and AI, Part 2 (Jan 2025) · OpenAI: data usage for consumer services · Anthropic: updates to consumer terms (Aug 2025) · Google Gemini API terms · Microsoft Copilot enterprise data protection