Surveys this year put unsanctioned AI use near half the workforce, and higher in law and finance. The tool your associate quietly adopted is handling the same confidential material your engagement letter promised to protect.
Your firm has an AI vendor. You did not choose it, sign anything, or read its terms. Your staff chose it for you, one browser tab at a time, and it is very likely a consumer chatbot running on a free tier that trains on whatever it is fed.
The numbers landed hard this month.
| Datapoint | Source |
| About 45% of workers use unsanctioned AI tools at work | TechTimes, June 2026 |
| About 67% of employees use AI tools at work | Salesforce 2026 Workforce AI Survey |
| About 27% of employees have pasted confidential data into a public AI tool | 2026 shadow-AI surveys |
| About 18% of organizations have a formal AI security policy | 2026 shadow-AI surveys |
The exact figures move around by survey, and the regulated fields run higher, with law and financial services sitting at or above these averages in the same 2026 reports. Round the whole thing off and the picture holds. Most of your people already use AI at work, a quarter of them have handed it something confidential, and fewer than one in five firms wrote down a rule about it.
Why this is a sharper problem for you
A retailer with shadow AI risks a leaked promo calendar. A law firm, an accounting practice, or an advisory shop risks the one thing it sells, which is confidentiality. When an associate pastes a client contract, a deal memo, or a quarter of financials into a consumer tool, that text can leave your control, enter a training pipeline, and come to rest on a server governed by terms nobody at your firm has read. The engagement letter still promises discretion. You can no longer prove you delivered it.
The part most policies miss
Shadow AI is a demand signal. Your people reach for these tools because the tools help them finish the work. A hard ban drives that use underground and leaves you blind to it. The firms handling this well give staff a sanctioned tool that is genuinely good, then make the safe path the easy one.
| Move | Why it works |
| Give people a sanctioned model with a real data boundary. A licensed enterprise tier with a written no-training clause. The free consumer tab does not qualify. | Removes the reason to open the shadow tool. Safe becomes the convenient option. |
| Write one page of rules people will actually read. What is fine to paste, what is never fine, and who to ask when unsure. | Fewer than one in five firms have any policy. A clear single page beats a forty-page document nobody opens. |
| Turn on visibility first. Find out which AI domains your network already talks to before you write another rule. | You cannot govern what you cannot see. Detect first, then set policy against what is actually happening. |
| Name the off-limits categories out loud. Client identifiers, privileged material, financials, health records, source code. | Specific warnings change behavior at the keyboard. Vague ones get scrolled past. |
The takeaway
The instinct is to send a stern email and consider it handled. That email will not reach the associate who pasted a merger agreement into a free chatbot at 11 p.m. because the sanctioned option was slow, locked down, or nonexistent. Give people a tool that is both better and safe, and the shadow version loses its job on its own.
At Ayvede, shadow AI is one of the first things a review looks for. We help high-trust firms swap the unsanctioned free tier for a governed tool their people actually prefer, with the confidentiality terms vetted before anyone pastes a single word.
AI, Properly is the Ayvede field report.
Advise. Vet. Deliver.
Ask a question in writing