The best LLM policy example is a short document your staff can follow in one sitting. It names the approved tools, says which data may go into each one, requires a person to check every output, and has an owner who keeps it current. Below are sample clauses, industry examples for Wisconsin manufacturers, law firms, dental practices and financial firms, and a rollout plan. Policy wording works only when your systems enforce it, so we also show where IT has to back the policy up.
TL;DR
- A usable AI policy covers scope, approved tools, data rules, human review, an owner and an update schedule. Governance decides who approves tools. The usage policy tells employees what to do each day.
- Approve tools by configuration, not by brand name. Connectors, agents and file access can each need separate review.
- Start from a published sample, then rewrite it for your own data, customers and contracts. Published samples say they are starting points.
- There is no comprehensive federal AI statute in the U.S., but states and agencies apply existing law. A policy reduces risk. It does not give you legal protection.
- A policy that no one enforces technically is a wish. Permissions, monitoring and tool controls make the rules real.
Table of Contents
- What an LLM policy is, and how it differs from AI governance
- What a workable policy contains
- Sample clauses you can adapt
- Examples by industry
- Build an approved-tool register, not just a list
- Where U.S. law fits in
- Roll it out in six steps
- Make the policy enforceable
- How Collett Systems helps
- Our perspective
- FAQ
- Sources
What an LLM policy is, and how it differs from AI governance
An LLM policy, also called an AI acceptable use policy, tells employees how they may use tools such as ChatGPT, Copilot or Claude at work. Governance is the layer above it. It assigns accountability and decides how tools and risks get approved and reviewed. The usage policy turns those decisions into daily rules. Both exist in a healthy organization, and for a small company they can sit in the same two or three pages.
ISACA's guidance on organizational generative AI policies separates a policy from a standard. A policy states what is required and prohibited, and why. A standard is a detailed, mandatory rule set underneath it. ISACA also says AI policies should not sit alone. Your data privacy and information security policies are tied to them and need to agree with them.
There is also an audience split. Saul Ewing's AI policy playbook distinguishes internal workforce policies from external, customer-facing ones. The internal one governs staff use and review. The external one explains to customers how you use AI and what safeguards apply. This article is about the internal policy.
What a workable policy contains
Published guidance converges on the same building blocks. Kohler Law Group's October 2026 guide says a company policy should define purpose and scope, list approved tools, set data rules for confidential, personal, customer and source-code information, and require human review. It should also address ownership of outputs, disclosure, security and employment issues, and name an owner who keeps it current.
The Information Technology Alliance's policy starter kit adds practical questions a policy should answer:
- Which tools are approved, approved for limited use, or prohibited?
- What data may go into each tier, and what may never go in?
- When must AI use be disclosed, to whom, and in what form?
- Who approves a new tool or use case, and how long does that take?
- How are prompts and outputs retained?
- Who do staff ask when the policy does not cover a situation?
That last question matters more than it looks. People bring in new tools faster than any policy can list them. A named person to ask keeps staff from guessing.
Sample clauses you can adapt
These are original examples in plain language, built on the elements above. Adjust the names and categories to your business.
Approved tools. "Employees use only the AI tools on the company's approved list for company work. Personal accounts on public AI services are not used for company information."
Data tiers. "Public information may go into any approved tool. Internal information may go only into tools approved for internal data. Customer, patient, client, financial, employee and trade-secret information may go only into tools the approval list marks for that category, and only the minimum needed for the task."
Human review. "AI output is a draft. The employee who uses it is accountable for its accuracy, tone, confidentiality and security before it is sent, filed or acted on."
Prohibited uses. "No AI tool makes a final decision about hiring, discipline, pay, credit or a customer's account. No one pastes credentials, source code or contract terms into an unapproved tool."
Disclosure. "When AI materially contributed to a deliverable for a customer, the manager decides whether and how that is disclosed, following the company's disclosure rule."
Reporting. "Staff report a suspected leak, a wrong output that reached a customer, or an unapproved tool to the policy owner right away. Good-faith reports are not penalized."
Ownership and review. "The policy owner reviews this policy at least once a year and whenever a new tool, regulation or contract requirement changes the risk."
The National Association of REALTORS® brokerage sample shows how specific a good clause can get. It requires human review before AI-generated listing content, client communications, transaction documents, disclosures or valuations are used. It says client, financial and transaction data should not be entered without prior approval. It also recommends minimum-necessary data, multi-factor authentication and encryption for systems that hold brokerage data.
Examples by industry
A sample from another trade is a starting point, not a finished policy. The National Council of Structural Engineers Associations says this itself about its engineering-firm sample. Platform considerations vary with a firm's technology, needs and use cases. Here is how the same skeleton changes by trade.
Manufacturers
The sensitive material is process knowledge: drawings, bills of materials, quality records, supplier pricing and customer specifications. A manufacturer's policy should say that none of these go into a public chatbot. It should also say which approved tool can summarize a work instruction or draft a maintenance log. Add a rule for operational technology: no AI tool gets a connection to the plant network or production systems without review. Manufacturers that need this decided before rollout can read our guide to AI adoption in manufacturing.
Law firms
Client confidentiality drives everything. Seyfarth's May 2026 briefing notes that AI data is now subject to legal hold and discovery, and that AI use may waive attorney-client privilege. A firm policy should therefore name exactly which tool may touch client matters, forbid client facts in any other tool, require attorney review of every output, and set how prompts and outputs are retained. Have your own counsel or ethics advisor confirm the wording against Wisconsin professional rules.
Dental and medical practices
Patient information is the category that should almost never leave an approved system. Your policy should list what counts as patient information and name the tools cleared to handle it. It should ban pasting charts, images or notes into anything else, and require staff review of AI-drafted patient communications. For HIPAA questions about a specific tool, confirm with your compliance advisor.
Financial firms
Client financial data, account details and anything that feeds a recommendation belongs in the restricted tier. Your auditors and examiners will ask how AI is governed, so the policy has to be documented and the approval list kept current. The federal court conference paper reports that the SEC made AI-related disclosures a 2026 examination priority. It also reports that the DOJ updated its corporate compliance guidance in September 2024 to expect compliance programs to address AI governance.
Build an approved-tool register, not just a list
The most useful single artifact in an AI policy is a register of approved tools. The 2026 NCSEA program recommends recording permitted data classes, enabled features, connectors, agent capabilities, owners and required controls. It also says approval should attach to a tool's configuration and capabilities, not just the vendor name. Connectors, agents, code execution and local-file access may each need their own review.
A starting layout looks like this:
| Tool and plan | Business purpose | Data allowed | Features enabled | Owner | Next review |
|---|---|---|---|---|---|
| Company-managed assistant | Drafting, summaries | Public, internal | No agents, no external connectors | IT lead | Quarterly |
| Meeting transcription tool | Internal meetings | Internal only | Recording off for client calls | Operations | Semiannual |
| Industry-specific tool | Quoting, scheduling | Customer data, minimum necessary | Named integrations only | Department head | Annual |
| Public chatbot, personal account | None | Nothing from the company | Not approved | n/a | n/a |
The same tool can be approved on one plan and unacceptable on another, because retention and training settings differ by contract. That is why the register records the plan and the enabled features.
Where U.S. law fits in
A policy is not a substitute for legal review, and the legal picture changes. As of White & Case's June 2026 tracker, the U.S. has no comprehensive federal AI law. The only AI-specific federal statute it lists is the TAKE IT DOWN Act, which covers non-consensual intimate imagery including deepfakes. States have filled the gap with their own laws, including those in Colorado, Texas and California.
A December 2025 executive order on a national AI policy framework directs federal action against conflicting state AI laws. That adds uncertainty about which state rules will last. Baker Botts' January 2026 update describes the NIST AI Risk Management Framework as voluntary in general, although some state statutes tie incentives to it. Obligations depend on your jurisdiction, sector and contracts.
Existing law still applies to how you use AI. The court conference paper reports that the FTC treats false claims about AI capabilities as deceptive practices and that the EEOC applies Title VII disparate-impact doctrine to AI-driven employment tools. For a Wisconsin employer, that means an AI screening tool in hiring is a high-risk use that needs human decisions and documented review.
Seyfarth adds one more caution. A strong governance framework helps manage risk but does not create a legal safe harbor. It has to be operationalized through procurement, documented decisions and consistent disclosure. For deeper reading on the data side, see our piece on LLM data privacy and security for Wisconsin businesses.
Roll it out in six steps
- Inventory current use. Ask department heads what staff already use, including personal accounts. Seyfarth recommends inventorying where AI touches people and ranking use cases by risk, with hiring, promotion and pay at the top.
- Assign an owner. One named person keeps the policy and the register current.
- Classify your data. Three or four tiers are enough: public, internal, confidential, restricted. Every tool then gets a ceiling tier.
- Draft from a sample. Use the clauses above and a published sample from your trade. Cut anything you will not enforce.
- Train and publish. Walk staff through real examples from their own work. Keep the document short and dated. The ITE kit advises balancing readability against the risk of constant revisions that overwhelm staff.
- Review on a schedule. Review at least annually and whenever tools, laws or customer contracts change.
Before any tool that reads your files goes live, check who can see what. An AI assistant can surface anything its user has permission to open, which is why we wrote fix your permissions before you roll out AI. To rank risks before you approve a tool, use our AI risk assessment framework. For the policy text itself, our AI acceptable use policy guide for employers goes deeper. Vendor questions belong in your vendor risk management process.
Make the policy enforceable
A written rule that nothing checks is easy to ignore. Enforcement takes four things that a policy document cannot supply:
- Identity and access controls, so staff use company accounts and multi-factor authentication.
- Permissions that match job roles, so an assistant cannot expose files an employee should not see.
- Monitoring that shows which AI services are in use on company devices.
- Backups and incident response, for when something goes wrong.
This is the gap between a policy and a program, and it is IT work.
How Collett Systems helps
We provide managed IT for small and mid-sized businesses in West Bend and Southeastern Wisconsin at a fixed cost per user. That includes 24/7 monitoring, proactive support, security management and compliance documentation, and we have worked with more than 150 local organizations. For manufacturers we also offer AI strategy and governance, and for other businesses AI consulting. If you want to know where your environment stands before you write a policy, start with the IT & Security Assessment.
If you want us to run the systems your AI policy depends on, see our Managed IT Services.
Our perspective
Most AI policies we read are fine on paper. The trouble starts on Monday morning, when an employee under deadline pastes a customer spec into whatever tool is open. A policy that depends on everyone remembering it will fail sooner or later.
We treat AI policy as infrastructure. It sits on top of permissions, monitoring, backups and vendor review, and it works only when those are in place. That is why we run IT as a standardized, fully loaded stack with a fixed cost per user: the controls a policy assumes are already part of the service, not add-ons.
For Wisconsin manufacturers, law firms, dental practices and financial firms, the policy should be short, dated and specific to your data. Then back it with technical controls and a person who owns it. Our team handles strategic IT planning, audit readiness and automation as included work, so the policy and the systems stay in step.
Dustin Collett
FAQ
Is there one LLM policy template that works for every business?
No. Published samples, including the engineering and real estate ones above, are starting frameworks that each firm must tailor to its technology, data and use cases. Keep the structure, then rewrite the data tiers, prohibited uses and review duties for your own trade.
How long should an AI usage policy be?
For a small or mid-sized business, two to four pages plus the tool register is usually enough. The ITE starter kit stresses readability and answering the questions employees actually ask. If staff will not read it, it will not protect you.
Do we need both an AI policy and an AI governance framework?
Yes, but they can be light. Governance assigns an owner and decides how tools and risks get approved. The usage policy gives staff day-to-day rules. In a small company, one person and one short document can cover both, as long as the approval path is clear.
How often should we update the policy?
Review it at least once a year and whenever you add a tool, change a vendor plan, or a law or customer contract changes. Date every version. Tracking tool approvals and renewal dates in your register makes these reviews quick.
Sources
- ISACA: Key Considerations for Developing Organizational Generative AI Policies
- Saul Ewing: AI & LLM Use Policy playbook
- Kohler Law Group: AI Use Policy Guide for Companies
- Information Technology Alliance: Generative AI Policy Starter Kit, Governance
- National Association of REALTORS®: AI Use Policy Template for Brokerages
- NCSEA: Generative Artificial Intelligence Program
- NCSEA: Generative AI Program, 2026 revision
- Seyfarth: AI in HR briefing, May 2026
- U.S. District Court conference paper: Your AI Is on Display
- White & Case: AI Watch, U.S. regulatory tracker
- Federal Register: Ensuring a National Policy Framework for Artificial Intelligence
- Baker Botts: U.S. AI law update, January 2026
