Your team is using secret AI. Your customer data may go with it

Secret use starts with time pressure

The first instance of shadow AI at a business is rarely the result of a grand strategy. It is a quick shortcut between meetings. Someone asks a tool to make an email sound friendlier, summarise a long document or draft an initial answer to a customer query. The working day gets a little easier, which is exactly why the tool catches on.

At a small business, it might be a single browser chat. A larger company may have several services, departments and even in-house applications. The underlying problem is much the same: management lacks a complete picture, making it difficult to judge where information is going or whether outputs are being reused without review.

The figure from Germany comes from a survey cited by Heise in its coverage of an SAP report. A separate survey by SAP and Oxford Economics found that 69 per cent of the companies surveyed reported occasional to frequent use of unapproved AI tools. These are different surveys, but both point to the same thing: an honest conversation about shadow AI is more useful than a reassuring shrug. What do you hear when you ask?

A ban only moves the problem

A blanket AI ban does not automatically mean less AI. Staff may switch to personal accounts or other services, while the work that prompted them to use AI still needs doing. If someone loses 20 minutes every day to a repetitive writing task, the appeal of a shortcut is easy to understand. That does not make the use safe, but it does explain why it may continue despite a ban.

So do not ask only, “Who broke the rule?” Ask, “What was the person trying to do, and why was there no approved way to do it?” Perhaps a template is out of date, nobody knows who owns the task, or the answer to a customer query is scattered across several folders. In cases like these, shadow AI also points to a process problem.

A survey by Okta and Apprize360 found that 52 per cent of employees use unapproved AI tools. In the same survey, 58 per cent of leaders reported an AI-related incident or near miss in the previous year. These figures describe the people surveyed, not every business. But they show why guidance and clear alternatives need to go together. How can you create a route that people will actually use when the day gets busy?

The risk is often in the prompt

A prompt can contain more sensitive information than the finished text. Someone who pastes a customer email in for editing may also share names, contract details or internal views. In HR, even a CV can contain sensitive information. Whether a particular use is lawful depends on the circumstances. The first practical question for a business is simpler: which information may staff enter into which tool?

An investigation does not have to begin with monitoring. Ask teams to describe the tasks they use AI for and the applications they use. Ask for examples, but do not demand private chat histories. Record whether staff handle customer data, internal documents or job applications, and whether someone checks the output before it is shared.

When decisions about people are involved, convenience is no reason to approve a tool. An Indeed survey of 300 HR decision-makers in Germany found that 28 per cent of respondents had no software for assisted screening and therefore sometimes turned to freely available tools such as ChatGPT, despite this being unauthorised. Heise describes this as prohibited under the GDPR. Which data or decisions in your business need a clear boundary first?

Recruitment can leave people out of the picture

A job application quietly entered into an AI chat is more than an attempt to save time. It contains information about a real person and may be used in a process that affects their chances. In the same Indeed survey, 11 per cent of HR decision-makers said applications rejected by AI would not then be reviewed by a person. Heise describes this as prohibited under the GDPR.

The report also points to an upcoming requirement: from 2 December 2027, companies must document AI-supported job applications and ensure a final human review. Anyone responsible for recruitment should not file this away as a problem for later. Check now where a tool sorts, assesses or recommends candidates, and who actually makes the final decision.

Even a small business should record which steps are supported by automation and who reviews the results. In a larger organisation, HR may need to coordinate with data protection and IT teams. The question is not just whether a system produces an assessment. Can anyone understand what it is based on, and who will stop it if it looks clearly wrong?

What this means for your business

You do not need a complete AI strategy before taking a useful first step. Start with a stocktake that does not put anyone on the spot. In a team meeting or a short anonymous survey, ask which tools come up in day-to-day work, what tasks they handle and where people are unsure. Make clear that the aim is to make processes safer, not to catch people out.

Then shape your rules around a few clear decisions. Which applications are approved? Which data must stay out? When should someone check the output before it goes to a customer, applicant or business partner? Who should staff contact when they need a new tool? At a small business, the answers may fit on one page. At a larger organisation, they may need to vary by department and data access.

Make the rules easy to find and use. A phrase such as “no confidential data” is little help if nobody knows whether an internal price list counts as confidential. Give practical examples, explain how to get approval and name someone people can ask. If the safe option is more cumbersome than a personal chatbot, the rule will be put to the test the next time things get hectic. Who will help your team then?

• Create an inventory: record the tools, tasks and data involved.
• Check higher-risk cases first: recruitment decisions, customer information and internal documents.
• Set out the approval route: put responsibilities, reviews and permitted data in writing.

Controls need to fit the workflow

An intranet policy makes little difference if it does not show up in the flow of work. Staff need an approved way to handle the tasks they actually do. They also need clear limits on what an AI system can access and when a person must check its output. This matters especially when applications use internal information or carry out tasks on their own.

The governance figures in the SAP report show how wide the gap between AI use and oversight can be. Only 11 per cent of companies rated their AI governance capabilities as fully adequate. Just 18 per cent said the necessary processes and frameworks were in place. These figures do not prove that every company is operating without controls. They are a reason to check your own position rather than assume it is fine.

Here is a useful test for next week: ask someone in customer service where they find a reliable answer to a common question. If the information is spread across documents, emails and different systems, Klarframe OS can bring those sources together, show answers with visible evidence and take access permissions into account. That is one possible way of working, not a prerequisite for a stocktake. What matters is that staff have an approved route they can follow and understand. Which recurring task could you make safer first?

Frequently asked questions

Related solutions

• Klarframe OS →

  Bring knowledge together, create transparency and act instead of react. Connect knowledge from documents, email, Notion, Slack and CRM with evidence and an interactive knowledge map.

More articles

• Use casesMultilingual guests, one wrong answer: the AI trap for tourism
• Use casesAI sends visitors. Your website loses the lead
• Models & releasesOne AI task cost 20 cents or $32.08. Which are you buying?