Back to Guides
Enterprise Software 15 min read

Anatomy of an AI acceptable-use policy that staff will follow

Anatomy of an AI acceptable-use policy that staff will follow

In a study of 48,340 people across 47 countries, the employees most likely to upload sensitive company data to a public AI tool were the ones whose employer had a policy. A policy guiding use was associated with a 56% rate, a ban with 67%, and no policy at all with 33%. Hold that while drafting, because it means the document is not the control. The parts of it a person can act on, unaided, are the control.

The source is Trust, attitudes and use of AI: A global study 2025 from the University of Melbourne and KPMG, with fieldwork from November 2024 to January 2025. One caveat the authors would make themselves: companies that write policies are also the ones using AI most, so this is association, not proof your policy caused the behavior. It still kills the comfortable reading. A signed file buys no lower risk rate, and the version people follow differs, part by part, from the version people acknowledge.

What the document can decide

Two things share the name. A vendor’s usage policy, from OpenAI or Anthropic, governs what anyone may do on their platform and reaches you by contract. Your AI acceptable use policy governs what your staff may do with any AI tool on company work. Most writing blurs them, which matters because you cannot edit the first and you own the second outright.

Inside that scope it settles four things and nothing else: which tools are in play, which data may enter them, which decisions a person must still make, and what happens when something goes wrong. What it cannot do is replace the conditions that produce the behavior. In the same study, 70% of employees reach AI at work through free public tools against 42% through tools their employer provides, and 56% have used AI without knowing whether it was allowed. A rule aimed at people with no company account, who cannot tell whether their use is permitted, is aimed at the wrong variable. That gap is the subject of our safe AI use playbook.

The eleven parts and what each has to settle

Eleven parts earn their place. The test for each is harsher than the published drafting checklists: can a person apply this at 4pm on a Tuesday, from memory, without asking anyone? If not, the part is prose rather than policy.

Part What it settles Followability test
1. Scope Which people, which work, which accounts Can a contractor tell whether it covers them?
2. Definitions What counts as an AI tool, features inside paid tools included Can someone classify the AI summary in their meeting app?
3. Approved tools The named products, by tier Can someone name the sanctioned tool for their task?
4. Data rules Which categories never enter which tools Can someone recite the categories without the file?
5. Permitted uses What is explicitly fine Does it name a task they did last week?
6. Prohibited uses What is out, each with the route instead Does every “no” carry a “do this”?
7. Human review Where a person must check output Is the trigger a document type, not a judgment call?
8. Disclosure When AI involvement must be stated, to whom Can someone tell from the audience alone?
9. New-tool path How a tool gets added, by whom, how fast A name and a turnaround time?
10. Reporting Where a mistake goes, and what happens first Is reporting cheaper than concealing?
11. Ownership Who owns the file, who owns the controls, next review date Two names and a date on page one?

Part 11 is left blank most often, and the cause is rarely a turf fight between operations and IT. It is vacancy, where each assumes the other owns the file, and a file with no owner never gets revised when a vendor changes a default.

Parts 3, 9 and 11 decide whether the policy survives next quarter. The Security Industry Association publishes its adopted policy openly, a rare chance to read a real one rather than a template, and it shows the standard defect: the approved-tool list is “subject to change without notice” and the review cadence is “periodically,” with no date. The list is good. The clock is missing.

The three parts that fail most often, rewritten

Three parts carry most of the risk, and all three fail the same way. They are written to be complete rather than to be applied.

Part The version that gets ignored The version that gets followed
Data rules “Do not enter confidential information, such as customer lists, source code, or financial records.” Five categories, each tied to the account it is allowed in: customer-identifying records, anything under someone else’s confidentiality terms, credentials, unreleased financials, personnel matters. Company tenant only.
Human review “All AI output must be reviewed for accuracy before use.” Triggered by the artifact, not the method: anything client-facing, anything with a number in it, anything entering a system of record, any text naming a person.
Prohibited uses “AI may not be used for hiring, performance management, or legal advice.” The same list, each line carrying the route: resume screening stays manual, and here is the tool for the job ad instead.

Data rules deserve the most attention, because the usual construction asks every employee to reason by legal analogy under deadline. Examples generalize badly: someone holding a vendor contract absent from the list concludes the list is complete. Five categories survive recall under pressure in a way twenty examples skimmed once do not.

Human review fails differently. “Review for accuracy” is an instruction to care more, which nobody can act on, and it implies the trigger is AI. Make the trigger the artifact. Client-facing work gets checked whether a person or a model drafted it, which is the rule you want on the day someone forgets to say they used a tool.

Every prohibition needs a route

This clause separates a file people follow from one they work around, and almost nothing published says so.

Each prohibition leaves a task that still has to be done. If the policy does not say how, the employee decides in private, and the study says what they decide: 57% report non-transparent use, presenting AI-generated work as their own or avoiding saying they used a tool. A prohibition with no route does not remove the behavior, only your knowledge of it.

Two lines aimed at the same risk:

Employees may not use AI tools to process customer data.

Customer records stay in the company AI tenant, where our retention and training settings apply. If you need AI on customer data, work there. Nothing customer-identifying goes into a personal or free account, ever.

Only the second tells someone with a customer spreadsheet open what to do next, and only the second leaves the work where you can see it. That is why the ban figure, at 67%, is the highest of the three: a flat no with nowhere to go routes the work out of sight.

The same logic governs tools. A short approved list with a named path for additions, a person and a turnaround time of five business days, turns the shadow tool into a request instead of a secret.

What to leave out

Length is the cheapest lever you have, and eleven parts does not mean eleven pages. The published guides grow the page count instead, adding sections that decide nothing, while 19% of employees in the study did not know whether their employer had any policy at all. That is a circulation failure, and a longer file makes it worse.

Four things belong somewhere else:

  • Framework mapping. Your alignment with the NIST AI Risk Management Framework or ISO/IEC 42001 belongs in a governance register for auditors, not in the file people read.
  • Risk analysis. The reasoning behind the rules belongs in that register too, and what it looks like at this size is in our anatomy of an AI risk register.
  • Tool configuration. Retention settings, connector permissions and tenant defaults change monthly. Reference the admin standard; do not copy it in.
  • Monitoring detail. Say what you can and cannot see, then stop. The instruments that enforce a policy without watching anyone are in our piece on enforcement without surveillance.

What remains fits on two pages, with the data categories and the reporting route on the first. Two pages people read beat forty pages people acknowledge, and the acknowledgment is all your records show either way.

Where the standards put the burden

Two reference texts name the policy directly, and both pair it with teaching rather than with inspection.

NIST’s generative AI profile, AI 600-1, names acceptable-use policies as a governance action twice. GV-3.2-003 asks you to define them for AI interfaces and human-AI configurations, “including criteria for the kinds of queries GAI applications should refuse to respond to.” GV-6.1-010 asks you to update them “to address proprietary and open-source GAI technologies and data, and contractors, consultants, and other third-party personnel.” That last clause is the one most policies miss, and the fix is in extending the policy to contractors and freelancers.

EU AI Act Article 4, applicable since 2 February 2025, asks providers and deployers to support AI literacy among staff and anyone operating AI systems on their behalf, measured against those people’s training and the context of use. It sets no required level for any one person. If you place products or AI output on the European market, the duty is a teaching one, and the policy is evidence of what you taught.

That is the practical reading of the 56% figure. Someone who knows why a rule exists applies it to the case the file never anticipated. Someone who signed an acknowledgment applies it to nothing, which is the argument for treating AI enablement as the control and the document as its record, made in the case against license distribution.

How to tell whether it is being followed

Four signals, none of which requires reading anyone’s work.

Can people say the data categories out loud? Ask five staff in different functions. Three correct from five is a policy in circulation. Blank looks mean a drafting success and a reading failure.

What share of use happens in the company account? Seat adoption against headcount, from your admin console. If half the team is active in the tenant and everyone says they use AI daily, the rest happens where you have no control.

How many new-tool requests arrive each quarter? Zero is the bad answer. It means the approval path is slower than the workaround.

How fast do mistakes surface? Measure time from incident to report, not incident count. Expect the count to rise in month one, because incidents are moving from invisible to recorded. Judge the trend from month three.

The Traliant AI Governance Gap Report, surveying more than 500 HR professionals in 2026, found 45% provide AI literacy training to all staff and 39% have a formal process to review AI output for bias, accuracy or legal risk. That is the honest mid-market baseline, and a policy shipped into a company with neither stays a document.

What this costs

Line item Market range Note
Drafting from a reviewed template plus counsel review Low four figures Two pages, not forty
Full external drafting engagement Mid four to low five figures For a regulated data class or a contract that demands it
Company AI seats Roughly $20 to $30 per seat per month Verify current list prices before budgeting
An applied AI enablement session Roughly $2,000 to $15,000 in the open market Varies with group size and custom material
Governance tooling and prevention of data loss Low to mid five figures a year Last in the order, not first

The ordering matters more than any figure in it. Seats and teaching come before tooling: a company that buys inspection before it issues accounts has paid to watch work happening elsewhere.

Two weeks covers the sequence: ask what is in use under an amnesty, buy the seats and set tenant defaults, draft the eleven parts to two pages, then teach it in ninety minutes against the team’s own documents. Run the recall test a month later.

Frequently asked questions

What should an AI acceptable use policy include?

Eleven parts: scope, what counts as an AI tool, approved tools by tier, data rules as categories, permitted uses, prohibited uses each with a route, human-review triggers, disclosure rules, a path for adding a tool, a reporting route, and named owners with a review date.

How long should an AI use policy be?

Two pages, with the data categories and the reporting route on the first. In the Melbourne and KPMG study, 19% of employees did not know whether their employer had a policy and 56% had used AI without knowing whether it was allowed. Both are reading failures that length makes worse.

How is an AI acceptable use policy different from a general acceptable use policy?

A general one governs company systems and devices. An AI policy governs what enters a third party’s model, which adds three questions the older file never asked: whether the input trains the model, whether a person must check the output, and whether the use must be disclosed. Most companies amend rather than replace.

Who should own the AI use policy?

Two named people: one for the document and the tool list, usually operations or legal, and one for the controls behind it, usually IT or security. The failure at 50 to 500 people is vacancy, not conflict, so put both names and the review date on page one.

Does the policy apply to AI features inside tools we already use?

It has to, and this is the scope question missed most often. The AI summary in a meeting app, a CRM’s drafting assistant and a document tool’s writing features handle the same data as a chat window. Define an AI tool by what it does with data, not by what the product is called.

Should we ban consumer AI accounts outright?

Ban them for company data and issue a sanctioned account the same week. In the study, uploads of sensitive company information were highest among employees whose employer had banned generative AI, at 67%. A ban with no route moves the work out of view rather than stopping it.

Does the policy apply to contractors and freelancers?

Yes, and NIST AI 600-1 names them in action GV-6.1-010, which asks organizations to update acceptable-use policies to cover contractors, consultants and other third-party personnel. The mechanism is a clause in the engagement terms plus an account decision, not a second file.

Is an AI use policy legally required in the US?

No federal law requires a private US company to have one today. Sector rules, state law and client contracts reach the same place in practice, and the EU AI Act’s Article 4 literacy duty has applied since 2 February 2025 to providers and deployers, including US companies placing AI output on that market.

Key takeaways

  • Risky AI use was highest where a policy existed, at 56%, and highest of all where AI was banned, at 67%, against 33% with no policy. The document is not the control.
  • Eleven parts earn their place, and each passes one test or fails it: can someone act on it from memory, unaided.
  • Write data rules as five categories tied to the account they are allowed in, not as examples inviting reasoning by analogy.
  • Give every prohibition a route. A no with nowhere to go produces concealment, and 57% already conceal their AI use.
  • Trigger human review on the artifact, not on whether AI was involved.
  • Keep it to two pages and move framework mapping, risk analysis and tool settings into a register. Both standards that name the policy pair it with teaching rather than inspection.

If you want a read on whether your policy is applicable as written, or what a ninety-minute applied session for your team would cover, book a free AI-readiness call. We will look at your data rules, your tool list, and the part people cannot repeat back.

Last Updated: Nov 15, 2026

DJ

Dirk Jan van Veen, PhD

SFAI Labs helps companies build AI-powered products that work. We focus on practical solutions, not hype.

Make your team fluent in AI — then automate what proves out

  • Hands-on training applied to your own documents and workflows
  • Often reimbursable through state small-business training programs
  • Built for 5–50 person firms with no IT department

Related articles