Acceptable Use Policy for AI: Data Tiers, Roles and Guardrails
An AI acceptable use policy has to name approved tools, sort company data into permitted, restricted and prohibited tiers, and give someone the job of enforcing it.
An acceptable use policy defines how people may use company technology and what information they may put into it. For generative AI, it must name approved tools, set data restrictions by role, and explain monitoring and enforcement. A signed memo is not enough. The rules need controls behind them.
Suppose an audit finds managers summarizing confidential client contracts in personal chatbot accounts, the policy has failed at the point of use. Give staff an approved way to do the work, then restrict the alternatives. A ban without a workable replacement encourages use outside company visibility.
Why a ban would not stop a manager on a deadline
A manager facing a contract deadline has a reason to use AI. Prohibiting the tool does not remove that deadline. Without an approved alternative, the work can move to a personal account or device where company controls provide little visibility.
Start by checking whether your approved tool can handle the actual task. If it cannot, resolve that gap before treating every workaround as a disciplinary problem. Approving a tool does not mean approving every document for upload. A client contract can remain prohibited even when the tool itself is sanctioned.
What the signed agreement must answer before a tool touches company data
“Enterprise version” is a product label, not proof of protection. Check the signed agreement for the exact plan against five questions:
- Are prompts and outputs used to train models, and is that the default or an opt-in?
- How long is submitted content retained, and where is it stored?
- Who at the vendor can access it, and under what circumstances?
- Which subprocessors handle the data?
- How does deletion work, and how long does it take?
A personal login on a company laptop does not become a company-controlled account. Name personal accounts explicitly in the policy. Recheck vendor terms at renewal and when the service changes.
What a defensible AI acceptable use policy must contain
The broader security policy assigns responsibility and defines required controls. The AUP tells staff what those requirements mean in daily work. NIST SP 800-53 Rev. 5 control PL-4 addresses rules of behavior; NIST CSF 2.0 control GV.PO-1 addresses policy governance.
For AI use, cover the following without turning the document into a technical manual:
- Purpose and scope, including personal devices used for work
- AI features inside existing software as well as standalone chatbots
- Approved tools by role
- Data classifications and permitted uses
- Explicit prohibitions, including unapproved personal accounts
- User responsibilities and acknowledgment
- Monitoring notice
- Consequences for violations
- A review date and named owner
Pair acknowledgment with scenario-based training. Ask staff to decide whether a real work document belongs in an approved tool. That tests understanding more effectively than asking whether they read the policy.
How to sort company data into permitted, restricted and prohibited tiers
Classify the data before choosing the tool. An approved vendor does not make every upload acceptable. Use this starting model, then have counsel check it against contractual and sector-specific obligations.
| Tier | Data types | Where it may go | Example |
|---|---|---|---|
| Permitted | Public or already-published material | An approved tool | A published product description |
| Restricted | Internal, non-public material not classified as prohibited | An enterprise tool with approved terms and settings | An internal process document |
| Prohibited | Client personal information, unreleased financials, source code under NDA, privileged legal material, credentials | No generative AI tool under this policy | A payroll file or API key |
Keep confidential client contracts out by default. Any proposed exception needs a review of the specific content and the vendor terms before upload. A manager’s deadline is not an exception process.
Writing the rules role by role
Finance needs a rule for budget drafts. Engineering needs one for code. “Use AI responsibly” gives neither team a usable boundary. Adapt this example to the tools and data your organization actually handles:
| Role | Approved tool class | Restricted data allowed there | Never enters any tool |
|---|---|---|---|
| Finance | Approved enterprise AI or spreadsheet features | Reporting templates without sensitive figures | Unreleased financials, bank details, payroll records |
| Engineering | Approved enterprise code assistant | Non-sensitive code and internal architecture notes | Source code under NDA, credentials, customer data |
| Legal | Enterprise AI with reviewed confidentiality and no-training terms | Non-privileged internal policy drafts | Privileged material, confidential client contracts, litigation files |
| HR | Enterprise AI reviewed for the intended use | Job description drafts and general policy text | Personnel files, health records, compensation data |
| Sales | Approved CRM and email AI features | Internal preparation notes without customer-sensitive content | Client personal information and non-public contract terms |
Public filings and published pricing belong in the permitted tier, not the restricted tier. Review each role’s row whenever an approved tool or its data handling changes.
Enforcing the policy with network, browser and identity controls
A signature does not prevent an upload. Technical controls must support the written restrictions, including AI features embedded in software the company already licenses.
What DNS and gateway logs can tell you
DNS and secure web gateway logs help identify AI domains reached from company devices. Where identity mapping is available, they can also associate activity with users. Review a few weeks of records rather than treating a single snapshot as the full inventory.
What encrypted traffic hides and how to close the gap
A domain visit does not prove that someone uploaded a contract. Encrypted traffic usually conceals the prompt’s contents from ordinary network logs. Use managed browser profiles and supported upload restrictions to limit what can leave. Block personal AI logins on managed devices where the controls allow it.
Identity controls that make the approved path the easy one
Require single sign-on for approved AI services. Disable separate local accounts where the vendor supports it, and tie access to managed devices. Test these restrictions against the actual service: a rule that blocks the approved workflow while leaving personal access open defeats its purpose.
Auditing the policy after it is signed
Start with existing DNS and gateway records, then compare them with SaaS discovery reports and subscription expenses. Ask department heads which tools their teams actually use. An initial disclosure window without penalties can surface uses that logs miss.
If restricted or prohibited data has entered an unapproved tool, involve counsel before deleting accounts or messages. Preserve available logs and establish what was sent. Assess notification duties before deciding how to address the employee’s conduct.
Review the policy at least quarterly. Check vendor terms and embedded AI features, then measure how much use is visible and how consistently staff follow the approved route. Document length is not a useful measure of control.
Apply consequences consistently, including to executives. Notice and acknowledgment matter, but selective enforcement undermines the policy. Decide the process before the first difficult case.
What Texas privacy law adds when staff paste customer information
The Texas Data Privacy and Security Act took effect July 1, 2024. For covered processing, its sensitive-data consent requirements make data classification more than an internal preference. Health information and biometric data require particular scrutiny before anyone puts them into an AI service.
Consent to collect information does not automatically authorize every later AI use. For Fort Worth and North Texas firms handling Texas residents’ data, counsel should assess the Act’s applicability and the proposed data flow. Do that before approving uploads, not after a manager has pasted customer information into a prompt.
Who runs the policy once it is in force
Assign an owner with authority to approve tools, resolve classification questions, and run the quarterly review. IT can configure access, but it should not have to decide legal exceptions or business risk alone. Our fractional Chief AI Security Officer service covers AI inventories and acceptable-use policies for organizations adopting AI.
Total 360 Security is based in Fort Worth and runs security programs across all 13 risk domains, with direct board reporting. We are Texas DPS-licensed and hold recognized professional security credentials. AI use belongs within that wider operating model, not in a document disconnected from the security program.
For a policy that needs an accountable owner, book a 30-minute risk discussion through the consultation page, or call 817-677-0515. Bring your current policy and the list of AI tools staff use, including personal accounts.
Frequently asked questions
What is typically in an acceptable use policy?
An acceptable use policy sets out what staff may do with company technology and what they may not do. It also says how the company monitors use and what happens when someone breaks the rules. An AI-specific policy adds approved tools by role and explicit rules for what data may enter them.
Should an AI acceptable use policy ban AI tools outright?
For a mid-market company, a blanket ban without a workable alternative encourages use outside company visibility. Approve tools for defined tasks, then restrict other services and personal accounts. Keep prohibited data out even when the tool is approved.
Who should own the AI acceptable use policy?
Assign someone with authority across IT, legal, and business operations. That owner must coordinate tool approvals, resolve exceptions, and run at least quarterly reviews. A fractional security leader can fill that role; drafting the document alone does not fulfill it.
Does the monitoring language need a lawyer's review?
Yes. Counsel should review the notice and proposed monitoring practices for the states where employees work. The AUP should explain what the company monitors, but notice alone does not make every form of monitoring lawful.
What should we do if client data already went into a personal AI account?
Involve counsel before deleting accounts or messages. Preserve available logs, identify the content sent and the service that received it, and assess notification obligations. Address employee conduct after establishing the facts and containing further exposure.
Total 360 Security provides virtual Chief Security Officer, vCISO, and Enterprise Security Risk Management programs for mid-market organizations. They design, run, and report security programs covering all 13 risk domains rather than just offering advisory decks or point solutions. The firm is Texas DPS-licensed with CPP, CISSP, and CISM credentials.
Serving Texas, California, Nevada, New Mexico.
More from the blog
10 min read
Identity Access Management Software: What to Govern First
A practical order for identity access management software: define privileged access, name an owner, inventory raw exports and review each quarter before buying a tool.
Read article7 min read
SOC 2 Audit: What Type II Requires and Who Owns Each Control
A SOC 2 audit is an independent CPA examination of your controls against the AICPA criteria. Here is what a Type II report requires, who owns each control and what evidence to keep.
Read article7 min read
Compliance Monitoring: What to Check Monthly and Quarterly
Compliance monitoring is the ongoing check that your operations follow the rules that apply to you. This guide sets a monthly evidence set, a quarterly sampling routine and clear ownership for every exception.
Read article