Shadow AI at work: your team already uses ChatGPT without rules
Shadow AI at work
Shadow AI at work is already in most offices that “do not have an AI strategy yet”. Someone has a private ChatGPT account. Customer copy lands in the prompt because the draft is faster that way. There is no allow-list. Leadership still thinks the topic is outside the door. It is on the desk.
You do not need a monitoring project to notice it. People say it out loud: “I ran the reply through ChatGPT”, about a mail that included the client’s name. That sentence is the whole diagnosis. Private ChatGPT at a company, with no written rule, is shadow IT in the form of an AI account.
The AXA SME labour-market study, cited by DeepCloud, points the same way: conscious AI use in Swiss SMEs from 22 percent (2024) to 34 percent (2025). Translation and correspondence are the most common jobs. Internally the tool is often still not approved. That is not our survey. It explains why shadow AI appears: the benefit is obvious on a Tuesday afternoon, and the approval never arrived.
How to spot shadow AI without surveillance
You find shadow AI in behaviour, not in keyloggers. Someone says in a meeting that ChatGPT wrote the paragraph. A mail reads unusually polished on a topic that same person normally struggles with. A colleague asks whether the free account is “enough” because the firm has no company plan yet. Finance sees no new SaaS invoices, but wording suddenly looks uniform.
These are signals, not proof. They are enough for a leadership conversation: which tools run today, and which data classes land there?
What happens if you only ban it
A ban without a replacement invites the phone. The prompt leaves the company laptop for a private account. Then you have no contract, no company log of which tool was used, and no agreed place for the data class. You also have not stopped the draft. You have moved it somewhere you will hear about later, if at all.
“We will buy Copilot for everyone” with no rule for which file may go where is the same gap with a purchase order. A licence names a product. It does not say whether a sick note, a salary, or a mandate file may be pasted into it. People will assume yes, because the firm paid.
What Swiss data protection means for customer data in ChatGPT is in ChatGPT and customer data. This piece is the organisational hole in front of that: nobody has said which tool counts.
The minimum, on one page
An AI policy for an SME can be one page. Three decisions, signed by the managing office, walked through with the team in about 30 minutes. Not a 40-page document, and not something you wait to start until you hire an IT lead you do not have.
- Allowed tools. A company chat, and Copilot inside the Microsoft 365 you already have. Nothing else for personal data.
- Data classes. Public copy. Internal notes with no people in them. Personal data. If you are unsure, treat it as personal and keep it out of any chat that is not the contracted one.
- A human reads anything that goes to a customer. The model may draft. It does not send.
Write the page in the words the office already uses. Name the tools by the product people have open, not by a category (“generative AI solutions”). Name one person who can add a tool to the list later, so the page does not freeze the firm in last year’s vendor. Keep customer files, payroll, and health details off the free tools entirely. The data-protection article linked above is the place for that traffic light. This page should point at it, not restate the Act.
Allow-list after the one page
The one page is the start. Many firms next need a short allow-list: which accounts the company pays for, which free logins are banned, and who reviews a new tool. Approve AI tools in five steps walks through that path. Shadow AI does not end with a slide titled “AI strategy 2027”. It ends when someone knows which chat is allowed for the client letter.
What you can skip
You can skip a strategy deck that never names a tool. You can skip a prompt library. You can skip ten pilots. None of those tell a colleague, on Thursday, whether the client email in front of them may go into the chat they already use.
You can also skip a technical lock-down as the first move. Blocking a website on the office laptop does not cover the phone in the same person’s pocket, and it does not give them a legitimate way to draft the letter that was the reason they opened the chat. The replacement and the rule belong together. The rule without a replacement is the ban. The replacement without the rule is a licence and a shrug.
If the only pain is one repeating workflow that currently lives inside that private chat, rules alone will not build it. The same report, the same follow-up list, the same inbox sort belongs in a defined flow instead. The one-page policy stays one page. The next step for the workflow is below.
What leadership should ask in the next meeting
Three questions are enough for 15 minutes: which tool saw the last client mail before send? Is there a company account for it, or only private logins? Who may add a new tool to the list tomorrow if marketing wants to test another product?
Put the answers on one page, not in a strategy paper with no names. If nobody can answer the third question, shadow AI is already built in.
Two next steps, not the same one
Rules for the whole team, an allow-list, and half an hour so people hear the same three lines: the AI in Business workshop.
A recurring workflow that today gets done by pasting it into a private chat, and should run as a defined flow instead: the Process Check. One workflow, half an hour. If you only want to clarify one workflow, the offer page is often enough before you book.
The fork in full, including who should not book which one, is workshop or implementation.
What is shadow AI in a small company?
Staff use a private ChatGPT account, or another unapproved tool, on work that includes customer text, and nobody has written an allow-list. The practical minimum is one page: which tools are allowed, which data may go there, and a human check before anything is sent to a customer. A ban without an allowed tool moves the same prompt onto a phone.
Written by
Aurum Avis Labs
Builds and ships at Aurum Avis Labs. Writes here about what we learn working with founders and SMEs in the DACH region.
Related Articles
You might also be interested in these articles
ChatGPT and customer data: what Swiss data protection means for SMEs
May staff paste customer data into ChatGPT? Public/internal/personal traffic lights, Free vs Business, FDPIC. Not legal advice.
Workshop or implementation: which one you actually need
AI workshop or Process Check: when the team needs shared rules, and when a single workflow should be built.