Over the past two posts, we’ve looked at generative AI from two angles: as a tool that can draft a solid first-pass contract, and as a risk that most NDAs were never written to address. Both of those posts point to the same underlying fact — your employees are already using AI tools, today, with your company’s data. The question this raises is a practical one: what does your company actually allow, and does anyone know what that policy is?

For many firms, the honest answer is that there is no policy. There’s a general sense that people shouldn’t do anything “reckless,” and maybe the IT department has endorsed Copilot. That’s not a policy; it’s an absence of one, and it leaves every employee to make his or her own judgment call about what’s safe to type into a chatbot. Some of those judgment calls will be fine. Some won’t be, and you won’t find out which until after the fact.

“Don’t Use AI” Isn’t a Policy. The instinctive response for some companies, particularly in the early days of the AI rollout, was to simply prohibit AI tools altogether. In practice, this doesn’t work. Employees have found these tools genuinely useful, and a blanket ban tends to just push the behavior out of sight.  Employees will use personal accounts on personal devices, and the company loses all visibility into what the employee base is doing.  

Not All Data Is the Same.  The most important element of a good policy is a simple way to classify information, because the right answer for a marketing draft is not the right answer for a client’s financial projections. Most companies can work with three practical buckets.

1. Public or fully generic.Information that’s already public, or that carries no competitive or confidentiality risk, such as a press release you’ve already issued, a request to draft a form thank-you email. No real restriction needed here.

2. Internal, but not subject to an obligation owed to a third party. An HR manager asking an AI tool to help word a difficult performance review, a manager summarizing internal meeting notes into action items, or someone tightening the language on an internal reorg announcement all fall here. None of that is information you’d want reaching a competitor,  but it isn’t information you’ve contractually promised a client, vendor, or partner you’ll keep confidential. Using an approved, business-tier AI account is a reasonable choice for this bucket.

3. Covered by an NDA, a client engagement, or genuine trade secret protection. This bucket works differently from the other two, and the policy should say so explicitly: the default here should be no AI use at all.  An approved business-tier account addresses the training and  security issues but it doesn’t answer the question that actually controls in this bucket. For NDA- or engagement-covered information, that question is contractual: does the agreement governing this information permit it to be put into any AI tool in the first place? For trade secrets, it’s even more fundamental: trade secret protection depends on the owner taking reasonable measures to maintain secrecy, and feeding the information into a third-party AI tool may itself undercut that status, agreement or no agreement. The default posture for information in this bucket is that it doesn’t go into an AI tool unless the governing agreement (or the law, in the case of trade secrets) doesn’t prohibit it, and someone with actual authority to make that call has approved that specific use in writing.

Not All AI Tools Are the Same.   Not all AI products are the same from a data-handling standpoint, and the policy should say so explicitly. Business or enterprise-tier accounts from major AI providers typically come with contractual commitments that customer inputs won’t be used to train the underlying models, along with better retention and deletion controls. Free, consumer-facing versions of the same tools frequently don’t offer those protections or bury them in terms of service most employees never read. A workable policy names the tools (or categories of tools) that are approved for company use, ties that approval to the data tier it’s meant for, and makes clear that free consumer accounts are not an acceptable substitute for an approved business account even when the underlying model is the same one.

The Output Still Needs a Human.  An AI tool’s output is a draft, not a finished product, and someone needs to own the job of reviewing it before it goes anywhere external. A good policy should say plainly that AI-generated summaries, emails, analyses, and documents get reviewed by a person before they’re sent to a client, filed with a regulator, or relied upon in a business decision. This is the same discipline any competent business already applies to work product, regardless of who or what produced the first draft.

 Have a Plan for When Something Goes Wrong.  Even a well-designed policy won’t prevent every mistake. Someone will eventually paste something they shouldn’t have. A useful policy says what happens next: who gets told immediately, whether the tool involved is on a business-tier account where the entry can actually be deleted from storage before it’s used in any training cycle, and whether the incident needs to be evaluated against any contractual notice obligations.

This Isn’t a New Problem. An AI use policy isn’t really a new category of corporate governance. It’s the functional equivalent of your company’s email policy, your data retention policy, and your acceptable use policy for company laptops,  The businesses that get ahead of this issue now aren’t doing anything groundbreaking. They’re just writing down, clearly and specifically, what they’d otherwise be leaving to chance.

If your company doesn’t have one yet, this is a good week to start.