Shadow AI in a Mid-Sized Company: Finding It, Sizing the Risk and Giving Staff a Better Option
November 11, 2026 · 6 min read · Etzal Group Team
A finance manager pastes a customer list with outstanding balances into a free chat assistant to draft collection letters. A developer pastes a block of source code, including a database password, to find a bug. An HR officer uploads CVs to get a ranked shortlist. None of them asked anyone, because there was nobody to ask and no rule to consult. This is shadow AI, and in most mid-sized companies it is already routine.
Why it happens
- The tools are free or cheap, and they help.
- The company has offered no approved alternative.
- There is no rule, or the rule is "do not use AI", which everyone understands to be unrealistic.
- People are trying to do their jobs faster.
It is worth starting from that last point. Shadow AI is evidence of demand. The task is to meet the demand safely, not to punish it.
The real risks
Data leaving the company. Customer personal data, employee records, pricing, contracts, source code and credentials entered into a service the company has no agreement with. Depending on the service and plan, the content may be retained, reviewed or used to improve models.
Personal data law. Sending personal data to a third party without a lawful basis, proper agreements or attention to cross-border transfer rules can breach obligations under Indonesia's personal data protection law or Singapore's PDPA. The organisation is responsible, even if an employee acted alone.
Unchecked output. A contract clause, a calculation or a statement of fact produced by an AI and used without review. The error becomes the company's error.
Confidentiality owed to others. Client data covered by a non-disclosure agreement, pasted into a tool, may breach the agreement regardless of what the tool does with it.
Automated decisions about people. Ranking candidates or assessing staff with a tool nobody has evaluated raises fairness and legal questions.
No record. When something goes wrong, nobody can say what was entered where.
Step 1: find out, with an amnesty
You cannot manage what you cannot see, and people will not tell you if they fear consequences. Announce that for the next two weeks the company wants to know what AI tools people use, that nobody will be disciplined for past use, and that the goal is to provide proper tools.
Ask five questions:
- Which AI tools do you use for work, including features inside other apps?
- Personal account or company account?
- For which tasks?
- What kinds of information do you enter?
- What would you lose if it were taken away?
Add what IT can see: browser extensions, sign-ins with company email, expense claims for subscriptions.
The answers usually show a handful of tools, a dozen tasks, and one or two practices that need to stop today.
Step 2: classify data into three kinds
A rule people can remember has few categories.
- Red: never in any external AI tool. Passwords, keys and credentials. Identity numbers, health and financial details of individuals. Payroll. Anything under a non-disclosure agreement that does not permit it. Unreleased financial results.
- Yellow: only in approved tools. Internal documents, customer and supplier business information, source code, contracts, operational data.
- Green: any reputable tool. Public information, general questions, drafting from scratch with no company data, learning.
Give examples from your own business for each. People classify by example, not by definition.
Step 3: provide an approved tool
This step makes the rest work. If the approved route is worse than the unapproved one, people will keep using the unapproved one.
When choosing, check the current terms of the specific plan for:
- Training. Is content you enter used to train models? Can that be switched off, and is it off by default on business plans?
- Retention. How long are conversations kept, and can you set this?
- Location. Where is data processed and stored?
- Access control. Company accounts, single sign-on, the ability to remove a leaver.
- Administration. Usage visibility and audit logs.
- Agreements. A data processing agreement suitable for your legal obligations.
Then make it easy: accounts for everyone who needs one, a short guide, and a few worked examples for common tasks.
For sensitive workloads, options include a private deployment, models running in your own cloud account, or tools that work on documents in place. These cost more and suit specific cases, not general chat.
Step 4: a one-page rule
Long policies are not read. One page:
- You may use AI for work with the approved tools, signed in with your company account.
- Red data never goes into any AI tool. List of examples.
- Yellow data only goes into approved tools.
- You are responsible for the output. Check facts, figures and references before using them. AI-generated text sent to a customer or regulator is your text.
- Do not let AI make decisions about people such as hiring, discipline or credit without a process approved by management.
- Tell us about new tools you want to use, and we will assess them within a stated time.
- Report mistakes quickly. If you entered something you should not have, tell the named contact. Early reports are not punished.
Name a person who answers questions. Train everyone for thirty minutes with real examples.
Step 5: review every quarter
- Which tools are in use, approved or not.
- Requests for new tools and how fast they were answered.
- Incidents and near misses.
- Whether the terms of approved tools have changed.
- Whether the red list needs updating.
The field moves quickly. A rule written once and filed is out of date within months.
Technical measures that help
Rules come first, and controls support them.
- Company accounts with single sign-on for approved tools.
- Managed browsers that restrict unapproved extensions.
- Secret scanning in code repositories, so credentials are not in the code to be pasted.
- Data loss prevention rules for the most sensitive patterns, such as identity numbers.
- Network or browser warnings, more than outright blocks, on unapproved AI sites: a reminder that points to the approved tool.
Blocks without an alternative push usage to personal phones, where you see nothing.
What not to do
- Ban everything.
- Pretend it is not happening.
- Write a twenty-page policy.
- Approve a tool without reading its terms.
- Punish the first person who reports a mistake. Nobody will report the second.
Where this fits
Shadow AI is usually the first sign that a company is ready for a deliberate approach: approved tools, a short rule, and then proper projects for the tasks where AI adds most. We help companies with that sequence through our AI Consulting service. If you want an outside view of your current exposure, contact us.
Summary
Assume staff already use AI tools. Find out with an amnesty survey. Classify data as red, yellow or green with examples. Provide an approved tool whose terms you have read, and make it easier than the alternatives. Write a one-page rule, train on it, name a contact and review every quarter. Demand that is met safely stops being a hidden risk.
Frequently asked questions
What is shadow AI?
Shadow AI is the use of AI tools for work without the organisation's knowledge or approval, typically staff using personal accounts on public chat assistants, browser extensions or AI features inside other apps. It is usually well intentioned. The risk comes from what data goes in and from decisions made on unchecked output.
Should a company ban public AI tools?
A blanket ban is rarely effective. People who find a tool useful keep using it on personal devices, and the company loses visibility. A better approach is to provide an approved tool with proper data terms, state clearly which data may never be entered anywhere, and make the approved route easier than the unapproved one.
Is company data used to train AI models when staff paste it in?
It depends on the tool, the plan and its settings. Consumer plans and business plans often have different terms about training and retention, and terms change. Read the current terms of the specific product and plan, and do not rely on general statements, including this one.
Related service
WhatsApp AI Chatbot
We build, train, and run a WhatsApp AI agent that answers customers 24/7, qualifies leads, and follows up.
