The AI Policy Template That Fits on One Page
Most firms that adopted AI never wrote the rules down, and the thirty-page policies that exist go unread. This is the six-rule policy I have watched hold across 150+ service units: one page, spells POLICY, and it covers the one thing no downloadable template does, agents.
Most firms that adopted AI never wrote down the rules. In McKinsey's 2023 global survey, only 21% of organizations using AI had a policy governing it. A Littler survey of 399 employers put it at 37%. Meanwhile, 98% of organizations have employees using unsanctioned AI applications.
Training does not close the gap either. In UpGuard's study of 1,000 employees, 40% remembered receiving AI training. The same share still used unapproved tools every day.
So the missing piece is not a policy document, and it is not a training day. The policies that exist are thirty pages of legal language that nobody reads under deadline pressure. An AI usage policy changes behavior when it fits on one page, sanctions the AI use that is already happening in your firm, and defines human review in terms that still work when the AI is an agent instead of a chatbot.
Here is that policy. Six rules, and they spell the word POLICY, so your team can recite them. Copy it, adjust the first rule, and ship it this week.
The one-page policy
P: Pick from the approved list. The firm provides AI tools covered by a data agreement. Work happens in those tools and nowhere else. [Name your sanctioned tools here.]
O: Off-limits. Client and third-party data never goes into public tools. This covers prompts, uploaded files, and sharing links. If a tool is not on the approved list, client data does not touch it.
L: Loop the firm in. Found a better tool? Propose it. It gets validated, then approved, or it stays out.
I: Irreversible actions need a person. Agents act alone only where their work can be checked and undone. Anything irreversible or client-facing goes through a human first.
C: Client terms decide disclosure. If a client's contract or a client's question touches our AI use, the answer is the truth.
Y: You own what you ship. Any AI output that reaches a client or a production system gets verified first by a named person who answers for it.
Never: use AI to impersonate a person, or let AI sign, commit, or bind the firm.
That is the whole thing. One caveat: this is a template from an operator, not legal advice, and your industry or your contracts may demand more. Have a lawyer read it once before you publish it internally. Then stop editing.
Why one page beats thirty
A policy earns its keep in one specific moment: a Tuesday afternoon when someone on your team is about to paste a client file into a free chatbot to hit a deadline. The thirty-page version was written for a different moment, the lawsuit that might arrive in two years. It protects the firm on paper and changes nothing on Tuesday.
BlackFog found that 60% of workers using unapproved AI would knowingly take the risk to finish a project on time. And the employment lawyers at FRB Law put the failure mode plainly: "Employees quickly learn which rules are real and which are decorative." A rule nobody can recite is decorative by definition.
Six rules that spell POLICY fit in a person's head. That is the entire argument for the format.
Two rules carry most of the weight
The P and O rules do the heavy lifting because they remove judgment calls. Most policies ask employees to decide what counts as "sensitive" before pasting it somewhere, and employees get that call wrong constantly. A developer does not think of a stack trace as client data. It often is.
"Pick from the approved list" skips the classification problem. Nobody has to grade the sensitivity of a document at 4pm on a deadline. The tool is either on the list or it is not.
Through a franchise network's delivery data, I have watched 150+ units run on exactly this design: one internal AI tool with a defined data privacy agreement, and a hard ban on putting work data into public tools like ChatGPT, Claude, or Gemini. The reason is the one your clients care about: data pasted into a public tool can leak, and it can end up as training data for the next model.
Sharing links count too, and this is the part most firms miss. In August 2025, OpenAI pulled ChatGPT's discoverable-sharing feature after thousands of shared conversations showed up in Google results. In July 2026, Axios found Claude artifacts in Google search too, including business plans and clinical trial documents their owners had shared by link. And the canonical prompt leak is still Samsung in 2023: engineers pasted source code into ChatGPT three separate times in under three weeks.
The rule no template has: agents
Every AI policy template you can download says some version of "review AI output before use." That rule was written for chatbots. You prompt, you read the draft, you decide. Review happens in the gap between generation and use.
Agents erase that gap. An agent that handles your inbox has already sent the email. The one wired into your billing system has already issued the invoice. "Review before use" is not a rule an agent can break, because for an agent the rule has no defined moment. Most policies in circulation were written before agentic tools existed and have nothing to say here.
The workable line is the I in POLICY: agents act alone only where their actions can be checked and undone. An agent that drafts replies and leaves them queued for approval sits on the safe side. Its work is checkable and reversible. The same agent with send permissions on client email sits on the other side, and needs a person in front of it. The question to ask about any automation is never how smart it is. Ask what happens when it is wrong, and whether you can take the action back.
Your clients will ask before your lawyer does
The pressure to have this policy is arriving from the client side, in writing. Corporate legal departments now add AI clauses to their engagement terms, and the Association of Corporate Counsel publishes sample guidelines their members use as a starting point. The requirements repeat across clients: disclose or get consent before using AI on their work, keep their data out of public models, have a human verify anything AI produced, and do not bill full rates for hours AI compressed.
Law firms felt this first, and legal is the leading indicator here, not the exception. The American Bar Association's Formal Opinion 512 already requires lawyers to disclose relevant AI use, verify output, and bill honestly for AI-assisted work. The same expectations are flowing into consulting and agency contracts through procurement, one questionnaire at a time.
Read the six rules again with that questionnaire in mind. O, C, and Y are the answers to it. A firm running this policy responds to a client's AI clause by describing what it already does.
This policy legalizes what is already happening
If your firm has no policy today, the policy's real job is to sanction AI use, not to introduce it. Your team already runs AI daily, unsupervised, and the newest hires run it most. On the delivery model ladder, this is the move from Assisted, where individuals use AI in private with no rules, to Enhanced, where the firm knows what runs where and client work has a gate.
The L rule, loop the firm in, keeps the policy true over time. A hard ban with no path around it recreates the shadow use it banned, because the person who found a better tool has nowhere to bring it. The valve I have seen hold in production is a real proposal path: anyone can suggest a tool, and it always gets validated before anyone uses it on work. People stop hiding tools when asking works.
Be clear with yourself about what this policy does not do. It will not make your firm AI-native, redesign a workflow, or move the numbers that tell you whether AI is working. It closes the gate on the most expensive failure mode, client data walking out through a free chatbot, so the real rebuilding work can start from a safe floor.
Copy the six rules. Put your tool names in the P rule. Give a lawyer twenty minutes with it. Then tell your team the true reason it exists: you are already using AI here, and we would rather make that safe than pretend it is not happening.
Common questions about AI usage policies
What should an AI usage policy include?
Six things: which tools are approved, what data stays out of public tools, who verifies output before it ships, what agents may do alone, when to disclose AI use to clients, and how someone proposes a new tool. That is the whole POLICY template above. Anything longer belongs in your lawyer's drawer, not in front of your team.
Is an AI usage policy the same as an AI acceptable use policy?
Same document, different name. Firms also call it a generative AI policy or a corporate AI policy. Whatever the label, the test does not change: can the people it governs recite it?
Do I need an AI policy if my team seems careful?
Yes. 98% of organizations have employees using unsanctioned AI, per Varonis, and careful people still misjudge what counts as sensitive data. Your clients are also starting to require a written answer in their engagement terms, so the policy pays for itself the first time a questionnaire arrives.
Can I put this in the employee handbook?
Yes. Paste it as a one-page section, name the person who owns the approved-tools list, and date it. Review it quarterly, because tools change faster than handbooks.
Does a generative AI policy need separate rules for agents?
One rule, yes. "Review output before use" fails for agents because the action has already happened by the time you would review it. Give agents autonomy only where their actions can be checked and undone, and keep a person in front of anything irreversible or client-facing.
