By James Gort. Originally published on LinkedIn on June 3, 2026.
Most businesses we talk to have no AI policy. Most of their people use AI anyway. Here is how to close that gap without slowing your team down.
There is a comfortable assumption I hear in a lot of firms: "We have not really adopted AI yet." It is almost always wrong. Not because leadership is fooling itself on purpose, but because adoption did not wait for a decision. Your people started using these tools on their own, to get through the day faster, long before anyone wrote a rule about it.
The numbers back this up. In a late 2025 study by the security firm UpGuard, more than 80% of workers reported using AI tools their employer had not approved, and fewer than one in five said they stuck only to approved tools. IBM’s 2025 Cost of a Data Breach report found that most organizations still have no AI governance policy in place or are only now writing one. So the real question for most leaders is not whether to allow AI. It is whether you know how it is already being used and whether you have made it safe.
What shadow AI actually looks like
It is rarely dramatic. A staff member pastes a long document into a free chatbot and asks for a summary. An associate drops in a draft and asks it to tighten the language. A bookkeeper feeds in a spreadsheet to track down an error faster. Good people, trying to do good work quickly.
The problem is where that information goes. Paste something into a consumer AI tool and you often have no idea whether it is stored, whether it trains the next version of the model, or whether it could surface in someone else’s output later. The data left your control the moment it was pasted.
For a law firm, this is the sharpest version of the problem. A paralegal pastes a privileged document into a public chatbot to summarize it before a hearing, and in that one move, client confidentiality is at risk. This is not hypothetical. Lawyers have already been sanctioned for filing briefs built on cases an AI tool invented and no one verified, and the ABA issued formal ethics guidance on these tools in 2024.
Do not read this as a law firm problem, though. Swap the privileged document for a client tax return, and it is a CPA problem. Swap it for a patient record, and it is a medical practice problem. Swap it for your pricing model or intellectual property, and it is every business’s problem. Law firms just have the highest-stakes version of a risk that sits in almost every office.
Why banning it does not work
The instinct, once a leader sees this, is to ban AI outright. It feels responsible. It does not work, for three reasons.
- First, your people are already using it, and a ban you cannot enforce is not a control. It pushes the usage further into the shadows, where you have even less visibility than you do now.
- Second, training alone will not save you. The same UpGuard study found that the employees who best understood security risks were among the most likely to use unapproved tools because their confidence in judging the risk went up. Knowing the rules and following them turn out to be two different things.
- Third, AI is delivering real productivity, and your competitors are capturing it. A blanket ban does not protect you. It just means you carry all of the risk of shadow AI and none of the benefit of governed AI.
A policy is not how you say no to AI. It is how you say yes to it safely.
What a real policy contains
A usable AI policy is shorter and more practical than most people expect. Five pieces do most of the work.
Approved tools. There is a real difference between a free consumer chatbot, which may train on what you give it, and an enterprise tier that contractually commits to no training and no retention of your data. Name the specific tools your firm allows and the version or tier.
Prohibited data. Draw a bright line around what must never go into a general AI tool. Client financial records, patient information, privileged material, credentials, and anything you would not post publicly. Make the line concrete enough that a new hire could follow it.
Human verification. Nothing an AI tool produces leaves your firm as work product until a person has checked it. The sanctioned-citation cases are what happens when that step gets skipped.
Disclosure. Decide when AI use needs to be disclosed, whether to a client or to a court. The specifics here belong with your counsel, not with us.
A fast way to say yes. People follow a policy they understand and that gives them a quick path to get a new tool approved. If approval is slow, shadow AI comes right back.
Paper is not a control
A policy that enforces nothing is a wish. The technical side is where it becomes real. Deploying the approved enterprise tools so people have a sanctioned option. Blocking or monitoring the consumer ones. Data loss prevention that catches sensitive information on its way out the door. Logging, so you actually have visibility into what is in use. That technical layer is the part we build and operate.
The interpretation layer is not ours. Whether a specific use waives privilege, what your state bar or licensing body requires, and how HIPAA applies when patient data touches a tool, those are conversations for your legal counsel. We build the guardrails. The legal judgment stays with you.
What good looks like
A business that has done this work is not a business that banned AI. It is a business where people use it openly, in approved tools, with the sensitive material walled off. The productivity gain is real, and the quiet exposure is gone. Nobody is hiding anything, because they have no reason to. And when a client or an insurer asks how your firm governs AI, you can answer in a sentence.
The quieter alternative
A staff member pastes the wrong thing into the wrong tool. You find out months later, when a client asks why their confidential information turned up somewhere it should not have, or when opposing counsel notices a citation that does not exist. By then it is not a policy gap. It is an incident, and incidents cost far more to clean up than policies cost to write.
If you do not know what your team is using right now, that is the normal place to be in 2026. The next step is finding out and building the guardrails before the gap becomes the incident.