Key takeaways
- Employees already use AI on their own, often through personal accounts nobody in IT approved, so a policy is catching up to a decision that has already been made.
- A workable policy answers twelve questions, from which tools are approved to who owns the document, in language people will actually read.
- The right stance, open, balanced or strict, follows from how sensitive your data is and how much review already happens, not from a template.
- A personal account and a company-provided account from the same AI vendor are different products for data purposes, even when the branding looks the same.
- Have legal counsel and HR review the policy before you adopt it, especially the sections on disclosure and decisions about people.
This guide is for owners and IT, HR and operations leads at companies of roughly 20 to 200 people, mainly in the US. Employees did not wait for a policy before they started using AI. Someone pasted a contract into ChatGPT to summarize it, asked Copilot to draft a client email, or ran a spreadsheet through Gemini looking for errors, often through a personal account nobody in IT approved or even knows about. A policy written after that happens is still worth writing. One written before the next tool shows up is worth more, because it gives people a clear yes instead of a guess.
Why do you need an AI acceptable use policy now?
Generative AI reached most employees through their own initiative, not through a company rollout, and that gap has a name. Microsoft and LinkedIn's 2024 Work Trend Index found that 78 percent of AI users bring their own AI tools to work, a pattern the report calls Bring Your Own AI, and warns that it puts company data at risk. When employees pick the tool, the company has no say in where a customer's details, a draft contract or an unreleased plan end up.
The risk is not AI itself. It is that nobody decided which tools are approved, what data may go into them, who checks the output before it is used, and who employees ask when a new tool shows up. Without that, IT finds out about a tool from a support ticket, HR finds out about an AI-drafted policy from an employee complaint, and legal finds out about a data question from a client. A short, specific policy replaces those surprises with a standing answer.
What should an AI acceptable use policy cover, section by section?
You can draft one interactively with the AI acceptable use policy builder, which turns answers to a short set of questions into a policy tailored to your tools, data and review habits. The sections below explain what each part of that draft is doing and why it belongs there.
A workable policy covers twelve areas. Each is a question an employee, or a lawyer, will eventually ask, and the policy should answer it before they do.
| Section | What it must answer |
|---|---|
| Purpose and scope | Who this applies to, and what counts as using AI at work |
| Approved tools | Which AI tools may be used, and what happens with a company-provided plan versus a personal account |
| What you may and may not enter | Which kinds of data may go into an approved tool, and what is off limits entirely |
| Checking AI output | How closely output must be reviewed before it is used, sent or relied on |
| Disclosure | When AI use must be disclosed to a customer, partner or colleague |
| Decisions about people | Whether AI may be used in hiring, performance, discipline or similar decisions, and what a human must still do |
| Building AI into company systems | Who may connect an AI tool, agent or automation to company systems, and what approval that needs |
| AI coding assistants | Whether developers may use AI coding tools, and what still goes through code review |
| Confidentiality and intellectual property | How to treat trade secrets, client confidences and other people's copyrighted material |
| Training | What training is required or recommended before someone gets access |
| Reporting problems | How to report a wrong or concerning AI output, and what happens when someone does |
| Ownership and review | Who owns the policy, approves new tools and sets the review date |
What does a sample AI acceptable use policy say?
A sample AI acceptable use policy has one plain sentence for each of the twelve sections above. Copy the lines below into your own document, replace the bracketed text and adjust the rules to your stance. Have legal counsel and HR review the result before you publish it.
- Purpose and scope. "This policy applies to every employee, contractor and intern who uses an AI tool for company work, whether or not the company provided the tool."
- Approved tools. "Employees may use [the company's approved AI tools]. A personal AI account may be used for public information only, never for company, customer or employee data."
- What you may not enter. "Do not enter a customer's contact details, an unpublished financial result, source code, a password or an employee record into any AI tool that has not been approved for that purpose."
- Checking output. "The person who sends or relies on AI output is responsible for it. Check facts, figures, names and citations before anything leaves the company."
- Disclosure. "Tell the customer or colleague when AI drafted a substantial part of a deliverable, and always when a contract or client requires it."
- Decisions about people. "AI may help summarize a resume or draft interview questions. A named manager makes every decision about hiring, pay, performance or discipline and records that they reviewed the AI's input first."
- Building AI into company systems. "Only [IT or an approved partner] may connect an AI tool, agent or automation to company systems, email or files, and each connection gets only the access its task needs."
- AI coding assistants. "Developers may use [approved assistants]. AI-written code goes through the same review and testing as any other code."
- Confidentiality and intellectual property. "Do not enter client confidences, trade secrets or other people's copyrighted material into a tool that retains or shares what you enter."
- Training. "Complete the AI training module before using a company AI tool, and again when this policy changes."
- Reporting problems. "Report a wrong, harmful or concerning AI output to [owner] as soon as you see it. Reporting early is never treated as a violation."
- Ownership and review. "[Role] owns this policy, approves new AI tools and reviews the policy every [6 to 12] months."
Naming the plan in line 2 resolves most of the shadow AI problem, because it gives people a clear substitute rather than a blanket no. Line 7 follows least privilege: an integration that uses one shared login can see everything that login can see, so give any AI connection only the access the task needs, never a standing administrator account. Name a person in line 12, not a department, because a policy with no named owner tends to go stale the first time a new tool appears.
How do you choose between an open, balanced and strict policy?
The right stance follows from two things: how sensitive the data your people handle actually is, and how much review already happens before work goes out the door. There is no universal answer, and a policy copied from another company's website usually gets this wrong in one direction or the other.
| Stance | Fits a company that | Typical rule |
|---|---|---|
| Open | Handles mostly public or internal information and already reviews work carefully | Any approved tool, review left to employee judgment |
| Balanced | Handles some customer or employee personal data, with several teams using AI differently | Approved tools plus limited personal use, review required before anything leaves the company |
| Strict | Handles regulated data, such as health, financial or legal records, or AI touches decisions about people | Company-provided tools only, review everything, IT approval for any new connection |
A company that bans every tool outright when its data is mostly public ends up with the shadow AI problem it was trying to avoid, because employees will use AI on their own devices when the company version is not there. A company that allows any tool with sensitive client files ends up with a data problem it does not know it has yet. Match the stance to what is actually at risk, not to how the last headline about AI made someone feel.
Illustrative example: Consider a 120-person marketing agency that lands on a balanced stance. Employees may use the company's approved tools, plus a personal account for public research only. Anything reaching a client goes through review first, and the agency discloses when AI drafted a substantial part of a deliverable. AI may help summarize meeting notes and draft first passes, but it plays no role in hiring or performance decisions beyond that summarizing. The whole policy fits on two pages, and every employee signs an acknowledgment when they join.
ChatGPT, Copilot, Gemini and Claude: how do free and business plans differ on data use?
Free and business plans from the same AI vendor differ on data use, so an approved tools section has to name the plan, not the brand. Terms change, so treat the table as a snapshot of each vendor's own documentation and check the linked page again before you rely on it.
| Provider | Consumer product | Business or commercial plans | Admin controls | Source |
|---|---|---|---|---|
| OpenAI (ChatGPT) | Free and Plus: conversations may be used to improve models unless the user turns this off in data controls | ChatGPT Business, Enterprise and the API: not used for training by default | Qualifying organizations can configure how long OpenAI retains business data | OpenAI enterprise privacy, Data controls FAQ |
| Microsoft Copilot (formerly Microsoft 365 Copilot) | Personal Microsoft accounts use consumer products that this documentation does not cover | Prompts, responses and data accessed through Microsoft Graph are not used to train foundation models | Copilot only surfaces content the signed-in user can already view; admins set retention through Microsoft Purview | Microsoft Learn |
| Google Gemini | Personal Google accounts are outside the Workspace commitments on the linked page | Workspace does not use customer data for training models without the customer's prior permission or instruction | Admins can disable the Gemini app and set how long conversations are kept | Google Workspace privacy hub |
| Anthropic Claude | Free, Pro and Max: new chats are used to train models when the user's model improvement setting is on | Claude for Work and the API: inputs and outputs are not used for training by default | Owners can turn off thumbs up and down feedback for the organization | Anthropic privacy center, consumer terms update |
Three things fall out of this table. First, a personal account and a company-provided account from the same vendor are not the same product for this purpose, so "we use ChatGPT" or "we use Gemini" is not a specific enough answer to a data question. Second, every major vendor now defaults business and commercial plans away from training on your data, which is the strongest argument for the approved tools section pointing people toward the company-provided plan rather than leaving the choice to them. Third, admin controls, not the model itself, are usually where retention and visibility actually get decided, so whoever owns the policy should also own those admin settings, or know who does.
For the wider data question, including prompt injection, logging and access control, see data security and governance for AI in the enterprise. This guide covers the policy that tells people what to do; that one covers the systems that enforce it.
How do you roll out an AI acceptable use policy?
Rolling out an AI acceptable use policy takes three steps, because a policy nobody has seen is not a policy.
- Train before access. A short session or module covering the policy, the approved tools and one or two examples of output that needs a closer look is usually enough, and it matters most for anyone whose role involves decisions about people or building AI into company systems. New hires complete it before they get a company AI tool, and existing employees complete it once, on a set date.
- Collect a signed acknowledgment. Even a simple one turns a document into something each employee has agreed to, and gives HR a record if a problem comes up later.
- Keep it in one place people already look. Use the employee handbook, the intranet or wherever onboarding paperwork lives, not a one-time email that scrolls out of view. Mention it again whenever the company adds or changes an approved tool.
How do you enforce and review an AI acceptable use policy?
Enforcement works best as a named person and a known route, not a monitoring dashboard. This is the part of AI governance a policy cannot do on its own. Someone, often in IT, security or HR depending on the company, owns the policy, approves new tools and answers questions about edge cases. Reporting a wrong or concerning AI output should go to that person by name, and reporting a problem early, even one the employee caused, should not by itself be treated as a violation. Ignoring a known problem should be.
Review the policy on a fixed schedule, every 6 or 12 months is common, and sooner if the company adopts a materially different AI tool or a new contract or client requirement changes what is allowed. Organizations wanting a broader structure for this work often reference the NIST AI Risk Management Framework, which organizes AI risk work into four functions: Govern, Map, Measure and Manage. This policy is the governance layer that framework assumes already exists.
What mistakes do companies make with an AI use policy?
- Banning every tool outright. This does not stop AI use; it moves it to personal devices and accounts where the company has no visibility at all, which is the exact risk a policy is meant to reduce.
- Writing a policy nobody reads. Ten pages of legal language gets skimmed once and forgotten. Two pages in plain language gets followed.
- No named owner. Without one, nobody approves a new tool request, and employees either wait indefinitely or go around the policy.
- No route for requesting a new tool. Employees who find a genuinely useful tool need somewhere to ask, or the policy becomes a reason to hide the request rather than raise it.
- Treating the policy as finished. AI tools change faster than most policies do; a policy with no review date will quietly fall out of date.
How Kastling approaches AI acceptable use policies
An AI acceptable use policy is a starting point, not a finished control. It sets the rules; the systems around it, permissions, logging, approval steps, have to actually enforce them, which is where most of the technical work sits. As part of AI & Cloud Infrastructure, we help set the admin controls, data access and approval points a policy like this assumes are already in place, so the rule on paper matches what the systems actually allow.
We are not a law firm or an HR function, and this guide is not legal advice. Have legal counsel and HR review any policy before you adopt it, particularly the disclosure and decisions-about-people sections, since the right answer depends on your contracts, your state and your workforce.
AI acceptable use policy builder
Answer questions about the AI tools your people use, the data they handle and who reviews what, and get a draft policy your team can adapt.
Questions
Can employees use ChatGPT or Claude at work on a personal account?
It depends on what the policy allows and what the account is used for. A policy with an open stance might permit a personal account for public information; a strict one might rule out personal accounts entirely for work purposes. What should never be allowed, whatever the stance, is company, customer or employee data going into an account the company does not manage.
Does a small business need an AI acceptable use policy?
Yes, because the risk comes from what people paste into AI tools, not from company size. A company of 20 people can have the same shadow AI problem as one of 2,000, and a two-page policy is cheaper to write before a data question comes up than after. Keep it short enough that everyone reads it.
Should we ban ChatGPT at work?
A blanket ban usually moves AI use to personal devices and accounts, where the company has no visibility at all. A better default is to approve a company-managed plan, say what data stays out of any tool, and give people a route to request new tools.
Who should own an AI acceptable use policy?
One named person, commonly in IT, security or HR depending on the company, who approves new tools, answers questions and runs the review. A policy owned by a committee with no single name attached tends to slow down every new request without making anyone safer.
Do we need a lawyer to review the policy before we adopt it?
Yes. This guide and the AI acceptable use policy builder can produce a strong first draft, but employment rules, disclosure requirements and contractual obligations vary by state, country and industry. Have legal counsel and HR review the draft, especially the sections on disclosure and decisions about people, before you publish it.
Sources
- NIST: AI Risk Management Framework
- OpenAI: Enterprise privacy
- OpenAI Help Center: Data controls FAQ
- Microsoft Learn: Data, privacy, and security for Microsoft Copilot
- Google Workspace: Generative AI privacy hub
- Anthropic Privacy Center: Is my data used for model training?
- Anthropic: Updates to our Consumer Terms
- Microsoft and LinkedIn: 2024 Work Trend Index