Your account manager has a client contract open in one window and ChatGPT in another, and she’s asking it to “make this sound friendlier” before the deadline. She isn’t being careless, just efficient. But if your business has no AI usage policy for employees, that quick favour to herself may have just handed your client’s confidential terms to a third-party server you’ll never see.

At ZIA Networks, we see it in nearly every company we audit, and it rarely comes from bad intent. Banning AI won’t fix it. What works is a short, clear policy, approved tools that protect your data, and simple rules people can follow in the moment.

This guide covers what shadow AI is, where the real risk sits, and what happens to data once it’s pasted. It also shows how to write rules your team will actually follow. It’s written for business owners, IT managers, and operations leads who want practical answers, not panic.

What Is Shadow AI?

Shadow AI is the use of AI tools such as ChatGPT, Gemini, or Claude by employees without the knowledge, approval, or oversight of IT or security teams. It is the AI version of shadow IT, and without an AI usage policy for employees, it leaves blind spots around where company and client data goes. 

Consider the following example. Ten years ago, workers uploaded their working documents to Dropbox due to the poor performance of the corporate system. The problem was noticed several months later by IT departments when there were already issues. Shadow AI is the same thing but happening way quicker because the tool is free, opens in ten seconds, and is used like chatting with an intelligent coworker.

The numbers are way larger than what leaders expect them to be. The Microsoft and LinkedIn 2024 Work Trend Index states that 78 percent of all AI users use their personal AI tools at work. Your workers did not wait for an official launch of the application.

Shadow AI vs. Sanctioned AI

Why Smart Employees Use Unapproved AI

Employees use unapproved AI because it saves real time, the approved options are missing or clumsy, and nobody told them where the line is. A clear AI usage policy for employees fixes the last two. Most shadow AI use comes from people trying to be productive, not careless or malicious. 

This is significant in terms of your approach to solving the problem, since when you consider the employees guilty, they will act as such. However, if you consider them responsible grownups who have not been guided, they will cooperate.

Here are the usual reasons we see in the field:

• Speed. Summarising a 40-page report takes minutes instead of an afternoon.
• No alternative. The company has no approved tool, so people pick their own.
• Unclear rules. Nobody said “don’t,” so people assume it is fine.
• Peer pressure. Colleagues are doing it and getting praised for efficiency.
• Invisibility. A browser tab leaves no obvious trail, so it feels harmless.

Notice what is missing from that list: malice. This is a leadership and process gap, which is good news, because gaps can be closed.

Shadow AI Risks: What Can Actually Go Wrong

The key AI shadow risks include confidential data falling into the wrong hands, violations of the clients’ contractual or privacy terms, inaccurate results being used in the clients’ processes, and IP loss. These risks can all be dealt with, but only when identified.

Let us break them down in plain terms.

1. Client Data Leaves Your Control

The text pasted into the public chatbot will go to a third-party’s server. The text can then either be kept, monitored by staff members for security purposes, or utilized to enhance future generations of the model.

A real-life case in point is that of Samsung in 2023. Reports emerged about how some of their engineers used ChatGPT to paste their proprietary source code and minutes of their meetings. Samsung banned the use of any form of generative AI in their company’s devices. The point here isn’t that engineers were negligent. It was that there were no guardrails in place.

Research from security firm Cyberhaven, based on usage data from its customers, found that about 11% of the data employees pasted into ChatGPT was confidential. That is a small percentage of each paste, but multiplied across a whole team and a whole year, it adds up quickly.

2. Contract and Compliance Breaches

Many client agreements include confidentiality clauses, data processing terms, or restrictions on sub-processors. Pasting client data into an unapproved tool can breach those terms without anyone realising.

Then there are the laws. Depending on where you operate, you may be dealing with GDPR in Europe, where fines can reach 4% of global annual turnover or €20 million, or India’s Digital Personal Data Protection Act, 2023. Italy’s data protection authority fined OpenAI €15 million in December 2024 over how personal data was handled, a reminder that regulators are actively looking at this space.

If you handle health, financial, or legal information, the stakes are higher still.

3. The Financial Hit

IBM’s 2025 breach cost research found that about 20% of organizations had a security incident tied to unsanctioned AI use, and those incidents raised the average cost of a breach by roughly $670,000. Consider the exact numbers as a benchmark, not a predictor for your organization, but the trend is unmistakable – unmanaged AI leads to greater costs.

4. Wrong Answers in Client Work

These tools have a way of sounding very certain even when they’re incorrect. They create their own statistics, use examples that aren’t real, and misconstrue nuance. Should an employee drop this information in a client document without checking it, you’ll be at fault.

5. Lost Intellectual Property

Your pricing models, proposals, source code, and strategy documents are assets. Once pasted into a public tool, you have limited ability to retrieve or delete them. Some of these security concerns fall under the broader category of generative AI security risks, which also include prompt injection, data poisoning, and unsafe plugins.

ChatGPT Data Privacy at Work: What Happens After You Hit Enter?

When an employee pastes data into ChatGPT, it is sent to OpenAI’s servers and handled under that account type’s terms. Personal and free accounts can leverage conversations to train models unless opted out, whereas business and enterprise accounts do not consider customer data for training unless opted in. Your AI usage policy for employees should name which accounts are allowed.

This is the section that most employees will never read, while most managers will not even inquire about it. The concept behind data privacy on ChatGPT in the workplace can be summed up in one line: the type of account determines almost everything.

Below is a general table made based on the descriptions of the plans by major vendors.

Two practical takeaways:

• Opt-out is not the same as safe. Even with training turned off, data still passes through a third party. Retention periods and human review policies still matter.
• Personal accounts leave no trail for you. If an employee leaves, their chat history, including your client’s data, stays in their personal account.

Microsoft Copilot Security Risks: When the Approved Tool Is Still a Problem

Microsoft Copilot respects your existing Microsoft 365 permissions, which means it can surface any file a user can already access. If your SharePoint and OneDrive permissions are messy, Copilot can expose sensitive data to people who should never have seen it.

Many leaders assume that buying an enterprise AI tool solves the problem. It solves part of it. The Microsoft Copilot security risks that matter most are not about Microsoft’s servers. They are about your own house.

Imagine a new junior employee who asks Copilot, “Summarise all documents about executive compensation.” If an old HR folder was shared with “Everyone in the company” back in 2019, Copilot will happily find it. The AI did not break a rule. It followed the permissions you set and then forgot about.

AI assistants may also be compromised, as revealed by security researchers. EchoLeak (CVE-2025-32711) is one such bug that was discovered in the Microsoft 365 Copilot in 2025. The bug revealed how data can be leaked via a tailor-made email without the user’s interaction. Even authorized AI assistants need monitoring.

Before deployment of the AI assistant, conduct a quick test for readiness:

• Check who has access to the private SharePoint and Teams sites.
• Stop “Everyone” and “Anyone with the link” access to confidential information.
• Label your client and financial information with sensitivity tags.
• Enable auditing of what the assistant accessed.

What Should an AI Usage Policy for Employees Include?

An AI usage policy for employees is a short written document that tells staff which AI tools are approved, what data may never be entered into them, how outputs must be reviewed, and what happens when rules are broken. It turns vague caution into clear daily decisions.

A good  AI policy template is not a 30-page legal document that nobody reads. It is two to four pages, written in plain language, that a new hire can understand in ten minutes. Here are the nine elements every policy should cover.

1. Purpose and scope. Explain why the policy exists and who it covers, including contractors and freelancers.
2. Approved tools list. Name the exact tools and account types allowed. “ChatGPT” is not specific enough; “ChatGPT Enterprise through the company login” is.
3. Data classification rules. What is acceptable for entry and what is not (see the table below traffic light system).
4. Client confidentiality. State that client data may only be used in AI tools when the contract allows it and the tool is approved.
5. Human review requirement. AI content that needs to be sent to the client requires verification by a specified individual.
6. Disclosure. Decide when you will tell clients that AI was used.
7. Request process. Give people a fast, easy way to ask for a new tool to be approved. If the process is slow, shadow AI returns.
8. Incident reporting. Make it easy for people to say, “I think I pasted something I shouldn’t have.” Blame-free reporting will resolve issues faster.
9. Consequences and review date. Spell out what happens when someone breaks the rules, and commit to revisiting the policy every six months, since AI tools and risks change faster than most documents can keep up. 

The Traffic-Light Data Table

This is the single most useful part of any policy. It lets employees decide in seconds.

AI Policy Template: A Starter You Can Adapt Today

An effective AI policy template contains six parts: purpose, approved tools, data rules, review requirements, reporting steps, and enforcement. Start with this skeleton, then customise it to your industry, client contracts, and legal obligations.

Use the following as a working draft. Replace the bracketed items with your own details, and have legal counsel review it before you publish it.

ZIA Networks Generative AI Acceptable Use Policy

1. Purpose. This policy helps us use AI to work faster while protecting client, employee, and company information.
2. Approved tools. Employees may only use the AI tools listed, accessed through company accounts. One should never use personal accounts for AI at work.
3. Data rules. Do not enter Red data into any AI tool. Amber information can only be accessed using authorized business applications.
4. Review. AI products should not be considered the final answer. A named employee must verify facts, figures, and references before anything goes to a client.
5. Reporting. If you believe restricted data was entered into an unapproved tool, report it to [contact] within 24 hours. Honest reports will not be penalised.
6. Enforcement. Repeated or deliberate violations may lead to disciplinary action. This policy is reviewed every six months.

A template alone will not protect you. It is the starting line, not the finish.

How to Roll Out Your Policy in 30 Days

In your company, it takes about one month to create a successful AI utilization policy through a company audit of the utilization pattern, selecting the proper AI applications, setting up a company-friendly policy, and training employees. Shadow AI is quickly developing its own practices, which shouldn’t be taken lightly.

This is a realistic time frame for a small to mid-sized business.

 Week 1: Children will learn about the design process.  

Hold a survey anonymously regarding what tools are used and why.Promise no punishment. You will learn more in one honest survey than in a month of guessing.

At the same time, ask IT to review web traffic and browser extensions for AI tool usage.

Week 2: Choose Your Tools and Draft the Rules

Select one or two tools, ideally enterprise-grade tools, that cover all scenarios. Ensure that they are aligned with the findings of the survey you conducted on the needs of your team. Create the policy based on the example provided above.

Week 3: Train, Do Not Just Announce

An emailed policy will be ignored by everyone. Host a 30-minute workshop that includes examples from the past. Highlight the before-and-after difference in prompts: this is a risky prompt; this is how we should have gone about it.

Example of a risky prompt: “Rewrite the following email to our client Priya Sharma at Northwind Ltd regarding the outstanding payment of INR 4,50,000.”

Example of a safe prompt: “Rewrite the following email to a client regarding an outstanding payment.” 

The safe version gets the same help with zero exposed data. Teaching this one habit- remove names, numbers, and identifiers before prompting– prevents a large share of problems.

Week 4: Monitor, Measure, and Improve

Check whether usage has shifted to approved tools. Collect feedback on what is slowing people down. Adjust the tool list or the rules accordingly.

How to Detect Shadow AI Without Becoming Big Brother

You can detect shadow AI through network and browser monitoring, cloud access security tools, data loss prevention rules, and open conversations with your team. Pair technology with transparency, because an AI usage policy for employees is only as strong as your ability to see whether it’s being followed.

Without visibility, a policy is nothing but an aspiration. Below are some layers of detection in order of increasing complexity:

• Conversation and surveys. The cheapest method and often the most revealing.
• DNS and web filtering logs. Show which AI domains are being visited and how often.
• Browser extension audits. Many AI add-ons request broad permissions to read page content.
• Data loss prevention (DLP). Can flag or block sensitive patterns, such as ID numbers or card details, before they leave the network.
• Cloud access security broker (CASB). Gives deeper control over sanctioned and unsanctioned cloud apps.
• Endpoint management. Lets you restrict AI apps on company devices.

Warning on culture. Deploying the monitoring system without any reason leads to loss of trust, and people find ways to get around this system, for example, by copying data from their phones.

Common Mistakes Businesses Make

After working with companies across different industries, the same errors appear again and again.

• Complete prohibition of AI. Prohibitions make its use go into personal devices where you have no visibility.
• Writing for lawyers instead of staff. When an employee cannot put down the policy in one sentence, then it is too complex.
• Forgetting contractors. The freelancers and agencies deal with client information without any training.
• No permissions cleanup. The use of Copilot without solving the problem of permission creep is like installing a lock on a door with no walls.
• Never updating. AI features arrive in existing software constantly. Your tool list from January may be outdated by June.

FAQs

Is it illegal to paste client data into ChatGPT?

It is not inherently illegal, but it may be. There are instances where this could be considered a violation of client contract agreements or data protection regulations like the GDPR and DPDP of India.

Can ChatGPT see my company data?

The provider’s systems process whatever you submit. On free and personal accounts, conversations may be used to improve models unless you opt out. Most enterprise plans will not include your information for training purposes, although you should check what the existing agreement says.

Do small businesses need an AI usage policy for employees?

Yes. Small teams often face higher risk because they have fewer controls and more informal habits. A two-page policy and one approved tool can remove most of the exposure at very low cost.

How do I find out if my employees are using shadow AI?

Start with an anonymous survey, then review web traffic logs and browser extensions. Add data loss prevention tools if you handle regulated or highly confidential data.

Should I forbid AI applications from being used on the job?

Not usually, since this will only encourage employees to use them behind their supervisors’ backs. It is much more effective to allow good tools, establish data policies, and train employees on how to use the tool responsibly.

Can Microsoft Copilot handle my confidential data?

Yes, provided that permissions, sensitivity labels, and logging are set up correctly. Copilot will provide users with the information that they have access to.

Get Your AI Risk Review from ZIA Networks 

It is a today problem, and the people causing it are usually your most motivated employees. Banning the tools will not work, and ignoring them will eventually cost you.

Here is your action plan:

• For this month: Choose one appropriate tool and write a clear policy on the use of AI in your organization following the above format.
• For this quarter: Restructure the Microsoft 365 permissions, train all employees, and start monitoring.
• Every six months: Review the policy and tools list because AI technologies keep developing.

This does not mean that you should conduct everything alone. ZIA Networks supports companies in identifying the use of AI technology, creating a practical policy, and securing software like Microsoft Copilot.

Ready to see where your client data is really going? 

Contact ZIA Networks today to book an AI risk review and turn shadow AI into a safe, productive advantage.