An AI usage policy your company can implement right now

Aug 26

If your employees use computers connected to the internet, AI is already touching your business data. Surveys consistently show most workers use the “big three” AI chatbots at work without telling anyone, with one recent report indicating: “[a] clear majority of office professionals (88%) have shared work-related information with public AI tools such as ChatGPT, Claude, or Gemini” and “[m]ore than two in five (43%) have entered emails or other work-related correspondence” into the same tools.[1]

And most worryingly: 66% of workers who use AI on the job “have used AI tools or services at work even though they believed they were not allowed under company policies.”[2]

It is happening, one way or another, so what should you do about it?

Banning AI does not work. As my colleague, Seth Pavsner of Cuddy & Feder LLP, has described, AI risks “do not vanish simply by banning AI use; instead, such bans merely push the use of AI tools out of the employer’s sight, making it extremely difficult to monitor and potentially catastrophic… .”[3]

You need to enable use of the AI tools that your employees are using anyway, allowing them to skill up into more sophisticated workflows and better tooling.

For this, your company must do three things: (1) define a small group of employees who have responsibility for triaging current AI use in your organization (AI chatbots? Meeting transcription? Spam filters?), reviewing and approving new AI tooling (using a defined framework), and escalating instances where management needs to step in; (2) get employees off personal AI subscriptions; and (3) adopt an AI Policy that governs those efforts.

This piece walks through how to do it, including the most important decisions your company needs to make, and includes a free, complete template at the end.

Put someone in charge of AI use

You must understand the landscape of AI at your company before you can intelligently adopt any AI policy. Step 1 is to put someone in charge: assemble an AI governance team of two to three individuals with different perspectives (e.g., technical, legal, business). This should not be a large committee: if ownership is too diffuse, the group will become ineffective.

Their first job is not writing the AI Policy; it is surveying AI use at your company and categorizing that use into buckets. For example, the tools your company has purchased (a coding assistant, a meeting transcriber); the tools employees bring in themselves on personal accounts (the shadow use in the survey numbers above); and the AI features your existing software switched on by default (the spam filter, the CRM’s draft-an-email button, the resume screener inside your HR platform).

When triaging, the volume in the third bucket may initially be a little shocking. AI features seem to be in every new software release. But that is the reason the survey comes before the policy: your company cannot write a reasoned policy to govern an inventory you have not taken.

Keep a list of what your team has found, which will form the foundation of your AI Register. A shared Excel file works fine and lets you add columns as you begin to understand how your business interacts with AI risk.

Once you have grasped and documented the landscape of AI at your company, you should triage the risk of the tooling in each bucket, ideally along the axes below:

Accuracy and reliability — What are the consequences of wrong output in this use case? Any known failure modes (hallucination, staleness)? Are errors detectable by the user in the ordinary course of work?
Bias and fairness — Does AI use affect hiring, employee evaluation, or customer treatment? Is vendor bias testing available?
Data sources — Which data was the model trained on? Does the training data create IP/infringement exposure?
Privacy exposure — What personal (human) data might be processed by AI? What are the retention and training-use terms? What deletion rights do we have? Where does the data reside (with us, AI vendor, subprocessor)?
Security and system integration — What systems does the tool connect to? What scope of access is permitted? Is there any risk of prompt-injection for tools that read untrusted content? What do we know about vendor breach history?
Regulatory exposure — Is the use regulated (e.g., employment rules, export controls)? What are the associated disclosure obligations and record-retention duties?

Evaluate risk however you prefer to do so at your company. If you are not sure, let me suggest the following heuristics:

Risk should be viewed as the composite of (a) the negative impact or magnitude of a potential harm and (b) the likelihood of that occurrence (adapted from NIST’s AI Risk Management Framework 1.0)[4]. For each dimension above, rate the tool or use Low, Medium, or High. Low, Medium, and High ratings should be applied in the context of your company’s risk tolerance, or its readiness to bear risk in order to achieve objectives.

For “Low” risk AI use: Review vendor terms on data training/retention, security basics, and account model. One team member may approve.

For “Medium” risk: Do all of the above (Low), plus review vendor security documentation, DPA where personal data is processed, and model documentation. Possibly conduct a small group trial before rollout.

For “High” risk: Do all of the above (Low + Medium), plus conduct a legal review of contract terms, document a failure-mode analysis, evaluate whether sandboxing makes sense, and where relevant, perform bias testing. The full team must approve. This three-level structure follows the FS-ISAC Generative AI Vendor Evaluation & Qualitative Risk Assessment.[5]

For every approved tool, document your decision in the AI Register. Ideally, include:

Intended use: what the tool is approved for, which data tiers, which user groups.
Known failure modes: how this tool goes wrong in this use (e.g., fabricated citations in research use; subtly insecure code in coding use; stale answers in RAG use; over-broad actions in agentic use).
Mitigations: the specific controls that address each failure mode (mandatory human review points, verification steps, scope limits, logging), and who is responsible for each.
Out-of-scope uses: the nearest tempting uses that are NOT approved.
Sandboxed use: whether the tool runs isolated from the rest of your business data, from the public internet, or both. Decide when you will default to sandboxed use.

AI cleanup

After triaging AI use, GET ALL EMPLOYEES OFF PERSONAL AI SUBSCRIPTIONS, with special emphasis on ChatGPT, Claude, and Gemini.

Why?

The frontier labs (OpenAI, Anthropic, Google) train their models on consumer inputs, and their models eat data like Cookie Monster eats cookies. The controls vary by provider, and they change often. OpenAI trains on ChatGPT Free and Plus conversations by default, and both free and paid users can switch that off in Data Controls.[6] Google will route a subset of inputs to human reviewers, and retain reviewed chats for up to three years even after the user deletes the conversation.[7] Anthropic currently uses consumer Claude chats for training if the user affirmatively opts in.[8] Among the diversity of options, there is next to zero chance every employee has made the right choice, or even knows the choice exists.
Disclosing confidential information to a public AI might be a breach of your company’s agreements (NDAs, or other agreements with confidentiality provisions). Or it may cost you trade secret protection you have spent significant resources to create.
Prompts are discoverable business records. In the consolidated OpenAI copyright MDL, plaintiffs (including the NYT) moved to compel production of ChatGPT conversation logs, and the court affirmed an order requiring OpenAI to produce twenty million of them.[9] The order applied to essentially all consumer users of ChatGPT, but it did not impact those who were on Enterprise or Edu accounts, or API customers who had signed a “zero data retention” agreement with OpenAI (again corporate users, not consumers).[10]

In other words, the account tier settings dictated whether customer records were produced. Organization-wide retention controls are therefore critical.

Building the AI Policy

In the next couple of sections, I am going to cover how to build your own policy, at a high level. If you cannot tell from my other articles, I do enjoy building things, so if you’re like me, come along.

Otherwise, as I mentioned up at the top, I have attached a template AI policy at the end of this article. If you are just interested in a template, feel free to scroll down, download, and read it over. I will not be offended (or even know you did it)!

First, if you have not already, decide which chatbot provider (ChatGPT, Claude, and Gemini recommended) your organization will use. Since your employees already use chatbots, this is the best and easiest way to generate employee buy-in to the policy while enabling productivity gains. At a minimum, remember to change privacy and retention settings at the admin level to options that prevent providers from training on your data. Some organizations will also want zero data retention (or as little retention as possible) to avoid dealing with issues like those in the OpenAI MDL (discussed above).

Second, have your AI governance team draft a simple, readable chart explaining to your employees which AI uses are allowed. In my draft, I use the classic stoplight colors (Green, Yellow, Red — we have all played “Red Light, Green Light” at some point in our lives, right?). It is not a mandate, though; feel free to use another system if it makes sense to you. Here is a rough outline of mine:

Green (Go). No permission needed. Employees can use AI here so long as they remain inside their usage limits. This category includes already approved tools for already approved purposes.
Yellow (Caution). Ask before proceeding. Employees must submit a one-page form to the AI governance team, and the default response should be “yes,” allowing the use. This category includes items that are a bit more sensitive, so they may require some kind of safeguard in the corporate setting, for example: use with contracts and pricing, financial/board/investor materials, core technical information, personal data, any new tool, integration, or agent, or anything the company will make public.
Red (Stop). Ask before proceeding, but expect a “no” answer unless there is strong business justification. These items present high risk, for example: trade secrets, privileged/litigation material, export-controlled data, passwords and API keys. Approval requests here might be escalated to the GC if the AI governance team feels the business need is strong enough.

Generally, the red list should be short and memorable. Most employees will genuinely try to comply, but it is easy to forget items on a long list. A word of caution: if you leave items off the red list but block them using admin controls (thinking you are achieving the same goal), be aware that some employees will return to their personal accounts (where the controls have not disabled access). Employees need to know and understand why an item is in the red category, or they will not buy-in.

Third, the AI policy and training should highlight reliability issues. AI systems make mistakes for a number of reasons. Thus, employees who use AI output should get into the habit of fact-checking. Both AI and humans make mistakes, but when a colleague makes a mistake, blame usually falls on the colleague. When an AI makes a mistake, blame falls on the user. It seems obvious, but some employees will need it to be stated in the policy.

Fourth, watch out for how default AI tool rules can impact information flow inside of your business. For example, in many companies, the finance team does not have access to R&D systems and data, and vice versa. Those teams do not need the access, and smart organizations often implement permission-based systems so each employee touches the information they need, but not what they do not.

By default, AI tooling connected to your data can indirectly expose your knowledge base across your entire organization. Accounting can ask a question about R&D and, through the tool, receive an answer. The AI policy and the AI governance team need to recognize that AI tools, wherever possible, should enforce permissions, and that some kinds of information should never be indexed (privileged communications, trade secrets).

Hiring

I am splitting this into a separate category because of its special regulatory exposure. There is currently significant statutory movement at the state level around the proper use of AI in hiring. Several states impose restrictions, for example requiring companies provide notice to job candidates whenever AI is being used in the job application process, mandating proof of bias auditing, and ensuring humans are appropriately in the loop. Watch out for these.

The law is changing fast, but a good baseline as of writing (mid 2026): inventory your tools, keep a human decision-maker with real authority, give notice, get bias-testing documentation, retain records for at least four years.[11]

If you can avoid AI in hiring (some companies cannot), or at least wait until the landscape is a little more settled, you might avoid headaches down the road.

Implementing the policy while getting employee buy-in

Though it may seem simple enough, the act of actually implementing an AI policy can lead to tension that ultimately spills over into public discourse,[12] if you are not careful. As with many things, openness and transparency are critical. Work with not against your employees; they are already using AI in their personal lives. The policy should facilitate AI use in a non-threatening way, rather than curtail or force it. This is, of course, a topic worthy of its own freestanding article. But, at minimum, if you encounter hard resistance, take a step back, open a dialogue, and incorporate feedback.

Below is an example playbook to achieve buy in. Note deadlines are extended over several weeks, not days, and the instances where the company can say “yes” occur before the periods when you may need to curtail some more risky use revealed in your audit.

Weeks 1-2: Select the AI governance team and plan the internal AI use survey (including amnesty for self-reporting).
Weeks 3-4: Survey tools employees are already using, without punishing shadow AI. Take no immediate action during the survey period unless the risk is too high to justify waiting.
Weeks 5-6: Add a set of primary AI tools on enterprise agreements, for example one of ChatGPT, Claude, or Gemini, as well as others that make sense for use in your business. Migrate any employees who are using personal AI tools off their personal accounts and onto enterprise accounts. Do not shut down unapproved AI use just yet.
Weeks 7-10: With primary tools in place, ask your GC or outside counsel to review state law requirements and contract terms that touch on AI use. In the meantime, prepare forms and procedures allowing employees to request AI tooling or use outside of the primary tooling. Do not circulate them just yet. Develop triggers for escalating requests to GC/board where needed.
Weeks 11-14: Announce the full AI policy, highlighting again with what tools are allowed. Employees should be aware an AI policy is taking shape, as they have already been migrating from personal to enterprise tools. Explain there will be a phase out of any unapproved tools unless the employee submits a request for use, which will be evaluated with enough time to prevent disruption so long as it is submitted on time. Conduct a short training.

After the rollout, you should have an approved set of core AI tools in place, an AI register and decision log, a working policy, and an AI governance team.

The Template: Take it!

I am linking a template you can freely take and adapt here. Substack only allows me to upload PDF files (not DOC or DOCX), but I am happy to share the original word version. Just email me or reach out on Substack!

AI Usage Policy Template - Strain PLLC (PDF, 320KB)

And of course, if you would like the finished version, tailored to your business, data, contracts, and where you operate, that is work I do for technology companies, alongside patent and IP work. Please feel free to contact me anytime.

Author Note

Stephen G. Nagy is a patent attorney and engineer at Strain PLLC. This article is for general informational and educational purposes only, is not legal advice, and does not create an attorney-client relationship. It reflects the author’s views, not necessarily those of Strain PLLC. Legal authorities change; verify currency before relying on anything here. For advice on your specific situation, consult qualified counsel.

Notes

[1] PagerDuty, Shadow AI in the Workplace (survey conducted by Wakefield Research, Apr. 9–20, 2026; released June 11, 2026), https://www.pagerduty.com/newsroom/shadow-ai-workplace-survey-2026/; PagerDuty, Survey: Office Professionals Used AI Tools at Work Despite Not Being Allowed, https://www.pagerduty.com/resources/ai/learn/survey-office-professionals-used-ai-tools-at-work-despite-not-being-allowed/. Sample: 1,250 non-IT office professionals at organizations with at least $500 million in annual revenue, in the United States (500), United Kingdom (250), Australia (250), and Japan (250); margin of error ±2.8 points overall.

[2] Id. (“Among those who have ever used AI as part of their work responsibilities, two-thirds (66%) have used AI tools or services at work even though they believed they were not allowed under company policies.”).

[3] Seth M. Pavsner, How the Adaptation of Artificial Intelligence Tools Is Impacting the Practice of Law, 9 J. Robotics, A.I. & L. 377, 379 (Sept.–Oct. 2026), https://www.cuddyfeder.com/wp-content/uploads/2026/07/Pavsner.pdf. The full sentence reads: “Risks of client confidentiality and privilege waiver (not to mention potential plagiarism, even if unintentional) do not vanish simply by banning AI use; instead, such bans merely push the use of AI tools out of the employer’s sight, making it extremely difficult to monitor and potentially catastrophic if the attorney using the lesser AI tool is just as cavalier about accepting AI-generated output as they are about using it despite being forbidden from doing so.”

[4] Nat’l Inst. of Standards & Tech., U.S. Dep’t of Commerce, NIST AI 100-1, Artificial Intelligence Risk Management Framework (AI RMF 1.0) § 1.1 (Jan. 2023) (”[R]isk refers to the composite measure of an event’s probability of occurring and the magnitude or degree of the consequences of the corresponding event.”), https://airc.nist.gov/airmf-resources/airmf/1-sec-risk/.

[5] FS-ISAC, Generative AI Vendor Evaluation & Qualitative Risk Assessment (Feb. 2024), https://www.fsisac.com/hubfs/Knowledge/AI/FSISAC_GenerativeAI-VendorEvaluation&QualitativeRiskAssessment.pdf (”The recommended due diligence plan is based on the highest risk level selected for all domains.”; Level 1 is “a basic set of questions for lower risk engagements” and Level 3 is “the most comprehensive set of questions”).

[6] OpenAI, How Your Data Is Used to Improve Model Performance, https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance (“When you use our services for individuals such as ChatGPT and Codex, we may use your content to train our models”; “By default, we do not train on any inputs or outputs from our products for business users, including ChatGPT Business, ChatGPT Enterprise, and the API.”); OpenAI, Data Controls FAQ, https://help.openai.com/en/articles/7730893-data-controls-faq (opt-out via Settings → Data Controls → “Improve the model for everyone,” available to signed-in and signed-out users).

[7] Google, Gemini Apps Privacy Hub, https://support.google.com/gemini/answer/13594961 (“A subset of chats are reviewed by human reviewers . . . to help improve Google services”; “Chats reviewed by human reviewers . . . are not deleted when you delete your activity. Instead, they are retained for up to three years.”; users may “turn off your Keep Activity setting”).

[8] Anthropic, Is My Data Used for Model Training?, https://privacy.claude.com/en/articles/10023580-is-my-data-used-for-model-training (last updated Mar. 16, 2026) (consumer Claude Free, Pro, and Max chats are used to improve models where “[y]ou choose to allow us to use your chats and coding sessions to improve Claude”).

[9] In re OpenAI, Inc., Copyright Infringement Litig., No. 25-md-3143 (SHS) (OTW) (S.D.N.Y. Jan. 5, 2026) (Stein, J.) (overruling OpenAI’s objections to Magistrate Judge Wang’s Nov. 7, 2025 order and requiring production of 20 million de-identified ChatGPT conversation logs, subject to the existing protective order), https://cdn.arstechnica.net/wp-content/uploads/2026/01/NYT-v-OpenAI-Order-1-5-26.pdf.

[10] OpenAI, How We’re Responding to The New York Times’ Data Demands in Order to Protect User Privacy, https://openai.com/index/response-to-nyt-data-demands/ (retention obligation applies “if you have a ChatGPT Free, Plus, Pro, and Team subscription or if you use the OpenAI API (without a Zero Data Retention agreement)”; “This does not impact ChatGPT Enterprise or ChatGPT Edu customers”; “This does not impact API customers who are using Zero Data Retention endpoints under our ZDR amendment.”).

[11] Cal. Code Regs. tit. 2, § 11013(c) (as amended eff. Oct. 1, 2025) (employers must preserve personnel and employment records, including automated-decision system data, for at least four years); Colo. S.B. 26-189 (signed May 14, 2026; eff. Jan. 1, 2027) (three-year retention, advance notice to applicants, post-decision explanation within 30 days, and meaningful human review).

[12] Samsung banned generative AI outright after an internal code leak and faced backlash. Mark Gurman, Samsung Bans Staff’s AI Use After Spotting ChatGPT Data Leak, Bloomberg (May 2, 2023), https://www.bloomberg.com/news/articles/2023-05-02/samsung-bans-chatgpt-and-other-generative-ai-use-by-staff-after-leak; Coinbase mandated AI coding tools, gave engineers a weekend to onboard, and fired those who did not; it too faced backlash. Maxwell Zeff, Coinbase CEO Explains Why He Fired Engineers Who Didn’t Try AI Immediately, TechCrunch (Aug. 22, 2025), https://techcrunch.com/2025/08/22/coinbase-ceo-explains-why-he-fired-engineers-who-didnt-try-ai-immediately; see also Fortune (Aug. 25, 2025), https://fortune.com/2025/08/25/coinbase-ceo-brian-armstrong-ai-coding-assistants-mandate-tech.

Stephen Nagy