Shadow AI: how to make sure company data is not uploaded where it shouldn't be
Your employees already use AI, often with personal accounts and pasted exports. A ban rarely stops that. Here is how you as an employer take control of where the data ends up, and give the team a route that is better than the back door.
A controller has to explain a variance to the management team tomorrow. It is faster if an AI service looks at the numbers, so the profit and loss report is exported and pasted into a chatbot, on a personal account. The answer is good. But the report now sits with a service the company has no agreement with, and nobody at the company knows about it.
This is called shadow AI. The need behind it is almost always reasonable: employees want to do the job faster and better. The problem is the route. This guide covers what the studies show, where the risks actually are and what you can do as an employer.
What is shadow AI?
Shadow AI is when employees use AI tools the company has not approved, often with personal accounts and without IT or management knowing. It is the same thing as shadow IT, with one difference: in an AI tool you do not just send files, you paste the content straight in. Contracts, salaries, customer lists and financial reports become part of a conversation held somewhere else.
What the studies show
Cyberhaven, which measures how data moves through company systems, looked at data flows at 222 companies in its 2026 report. One in three employees uses AI tools through personal accounts, and almost four in ten AI interactions contain sensitive data 1.
Data ending up in the wrong place is not just a theoretical risk. Two widely reported examples:
- Samsung, 2023. Staff reportedly pasted internal source code into ChatGPT to get help. When it came to light, Samsung imposed a temporary ban on generative AI tools 2.
- Shared ChatGPT conversations in Google, 2025. Conversations that users had shared by link, and marked as discoverable, showed up in Google search results, and OpenAI removed the feature 3.
Neither case needed a breach. All it took was an employee who wanted to do their job, a tool without a company agreement and a setting nobody had thought about.
Why a ban is not enough
The most common reaction is to ban AI. It feels safe, but the need does not go away. Employees carry on with their private phones and computers, and the company sees even less of what happens. A ban without an alternative moves the risk out of sight.
What works is giving the team an approved route that is at least as good as the back door, and making it clear what may be used for what.
Six steps to protect company data
- Find out what is already in use. Ask the teams, without looking for someone to blame, which AI tools they use and for what. The answers show where the need is.
- Classify the data. Decide what is public, internal and sensitive. Personal data, salaries, customer data and unpublished figures are almost always sensitive.
- Write a short AI policy. One page is enough: which tools are approved, which data may be used in which tool and who to ask when unsure.
- Give the team an approved route. Choose tools with a company agreement, where the data is not used to train models and where you control who has access. For financial data: let AI work on connected data instead of pasted exports.
- Add technical controls where needed. Block sign-in with personal accounts in the most used AI services on company devices, or use tools that warn when sensitive data is pasted.
- Follow up and train. Review the policy when new tools arrive and show concrete examples of what is fine and what is not. Most people want to do the right thing when they know what it is.
Financial data is the most sensitive thing being pasted
In the finance team the need for AI is greatest and the data most sensitive. Profit, margins, salaries and receivables are exactly what you want help analysing, and exactly what should not leave the company. On top of that, the AI works on an export that was out of date the moment it was taken, and the answer cannot be traced back to the books.
So a good solution for finance is not only about security. It should also give better answers than the chatbot: on current data, with numbers that can be checked.
How it works in Fintower
Fintower connects the accounting system, such as Fortnox, Visma or Business Central, through secure read access. Nobody exports and nobody pastes. The team asks in plain language and the AI works on the connected data, and every number can be traced back to account and voucher.
Colleagues are invited to the same workspace and permissions control who sees what. Fintower reads the data and does not write to your source systems. The data is stored within the EU and is not used to train AI models. That makes it easy to write an AI policy where financial data has an approved route.