Transparency notice: ownvia is a product of SATRI Consulting. This article is intended to provide professional information and does not constitute legal advice.

Generative AI gives the sentence a new meaning. Employees translate quotes, instructions and emails every day using ChatGPT and other AI tools. In doing so, they teach the AI how their company communicates: which technical terms apply, what tone is appropriate for the customer and what a product is called. However, this knowledge usually remains in the personal chat history and does not become part of a shared knowledge base.

The short answer: Shadow AI in translation is not just a data protection risk, but above all a knowledge problem. Company knowledge flows out into private AI chats, but does not come back. Every correction, every technical term and every successful wording remains with the individual employee. Bans rarely solve this. A shared environment is more effective, in which terminology, approved translations and guidelines are maintained centrally and benefit everyone.

How widespread is shadow AI in companies?

Shadow AI refers to the use of AI tools outside the applications provided or approved by a company. Two recent surveys show that it is part of everyday life in German companies.

Taken together, 42% of the companies surveyed by Bitkom therefore report or suspect shadow AI. “Not provided” does not automatically mean “prohibited”. However, it does show that actual AI use and official processes are diverging.

One finding from the IW is surprising: shadow AI is most common where employers promote the development of AI skills – 44.5% of employees who receive strong support use tools that have not been provided, compared with 21.3% of those who perceive no support. The IW refers to this as a provision gap: employees’ willingness is growing faster than the official offering. According to Bitkom, only 26% of companies provide their employees with generative AI. Shadow AI is therefore less a sign of rule-breaking than a sign of unmet demand.

This is particularly significant for translation. Unlike a short search query, the complete document is usually submitted to the AI – including specialist knowledge, customer data and product details.

What employees teach AI every day

Anyone who translates with AI is not merely entering text. Good results are only achieved when employees explain to the AI how their company communicates. “Teaching” here means instructions and context in the respective chat, not training the AI model. Typical instructions from everyday work:

• Terminology: “Here, we call the component a drive unit, not a motor module.”
• Tone of voice: “We address customers in Sweden by their first names.”
• Product knowledge: Explanations of technical relationships to ensure the translation is technically accurate.
• Corrections: “This sounds too promotional; please make it more factual.”
• Context: uploaded previous versions, reference documents and earlier translations.

This is exactly the knowledge that companies otherwise maintain at considerable effort in terminology databases, style guides and Translation Memories. Employees generate it alongside their everyday work, every day and often to a high standard. The problem is that it is not captured centrally anywhere.

Practical example: One component, three translations

A medium-sized mechanical engineering company uses AI in three departments. Sales translates quotations, Service translates maintenance instructions, and Technical Documentation translates operating manuals. Each department uses its own accounts and its own prompts.

The result: The same component appears in English as “drive unit”, “motor module” and “actuator”. The customer receives a quotation, a manual and a service note with three different terms for one component. Each department has taught its AI something correct – but each has taught it something different.

The one-way street: Knowledge flows out, but not back

Shadow AI functions like a one-way street: company knowledge flows into private AI chats, but not back into the company. Employees add specialist terms, correct wording and provide reference documents. This improves their own results. But without a shared process, this knowledge does not become part of a centrally maintained knowledge base – what is corrected in a personal chat today is not available to a colleague for their next assignment. The consequences become apparent in everyday work:

This makes the classic knowledge management problem even worse. In the past, knowledge remained unused in people’s heads. Today, it is actively written down — but in a place to which the company has no systematic access.

Paraphrasing the well-known saying: If companies knew what their employees had already taught AI, they would have had the foundation for a consistent corporate language in their hands long ago.

And data protection? The other side of the same coin

The same one-way street has another side: data leave the company along with the knowledge. If a document contains personal data, the GDPR applies. If an external service is used as a processor, a contract under Article 28 GDPR is generally required. Business secrets and confidentiality obligations towards customers also come into play, regardless of whether personal data is involved.

Personal AI accounts are therefore generally not a suitable solution for confidential or personal company documents. Business offerings provide more control: OpenAI, for example, does not use data from ChatGPT Business and Enterprise for model training by default and provides a data processing addendum (OpenAI Enterprise data protection). But even then, the following applies:

Important: “No model training”, “no storage” and “processing exclusively in the EU” are three different statements. In ChatGPT Business, for example, the selected region controls where content is stored—not necessarily every processing step (OpenAI: Storage location for ChatGPT Business).

A data processing agreement is an important prerequisite, but not complete proof of compliance. The legal basis, data minimisation, transfers to third countries and, where applicable, a data protection impact assessment remain the company’s responsibility. We describe in detail how companies establish the rules for this in our article AI translation in the enterprise: Governance in four steps.

Reclaiming knowledge in five steps

Shadow AI emerges because the private chat is faster and more convenient than the official route. Bans alone achieve little: according to IW, they would primarily affect committed employees. More effective is an official route that works better than the private chat.

1. Make usage visible – without blanket suspicion. Ask which departments are using AI to translate, which texts, languages and volumes are involved, and which prompts have proved effective. The aim is understanding, not control.
2. Gather existing language knowledge. Glossaries, proven instructions, approved translations and style guides are often already scattered across departments. They form the foundation for a shared basis.
3. Provide an approved environment that convinces users. It must be at least as easy to use as a private chat and offer more: better results because the company’s knowledge is already built in.
4. Maintain knowledge centrally and apply it selectively. Terminology, Translation Memories and specifications are maintained once and assigned to the relevant job. Approved corrections and translations can be incorporated into the shared knowledge base and reused for future jobs. This gradually turns a one-way street into a cycle.
5. Establish rules for review, approval and AI competence. Define which texts require subject-matter review and who gives approval, and provide employees with practical training. This also supports the obligation under Article 4 AI Act to take measures to promote AI literacy.

How ownvia retains language knowledge within the company

ownvia is the AI infrastructure for professional translation processes. The underlying principle: your language and knowledge remain the constant, while technology continues to evolve with ownvia. Instead of getting lost in private chats, language knowledge grows in one place that everyone with the appropriate access rights can use.

Storing knowledge is not the same as applying knowledge in the right context. A central knowledge base alone is not enough. A job involving machine control requires different technical terms and reference texts than a marketing brochure. That is why the relevant company knowledge must be assigned specifically to the respective job. Only then can the AI take the appropriate information into account without being burdened by contradictory or irrelevant specifications. That is precisely what Translation Templates in ownvia are for.

The difference from a general chat assistant is not that it could not process glossaries or style guides. What matters is whether knowledge is maintained centrally, applied consistently to every assignment and linked to defined review steps. In 2026, ownvia was included in Slator’s ‘Language AI 50 Under 50’ list.

Checklist: Does your language knowledge stay within the company?

• We know which AI tools employees use for translations.
• Proven instructions and glossaries from the departments have been collected.
• Mandatory terminology is maintained centrally and applied automatically.
• Approved translations are systematically reused.
• There is a defined process for incorporating reviewed and approved corrections into the shared knowledge base.
• It is regulated which documents may be processed with which applications.
• Contractual and data protection requirements have been reviewed.
• Review and approval steps for quality-critical texts have been defined, and employees have been trained.

The more points remain open, the more language knowledge your company develops every day—and then loses again. The checklist is intended as guidance and does not replace data protection or legal review.

Conclusion: Shadow AI is a signal, not an enemy

The fact that employees use AI to translate is initially good news. It shows demand, initiative and a wealth of language knowledge that is generated anew every day. The problem is not the use itself, but that this knowledge ends up in private chats and data leaves the company unchecked.

Companies that systematically make this knowledge usable can reduce data protection risks and establish a consistent corporate language that continuously evolves through reviewed and approved translations.

The decisive question is therefore not: “Are our employees allowed to translate with AI?” But rather: “How do we ensure that what they teach the AI benefits the entire company?”

Talk to us about your translation process – we will show you ownvia in a demo.

Frequently asked questions about shadow AI in translations

What is shadow AI in translations?

Shadow AI refers to the use of AI tools outside the applications provided or approved by the company. In translations, this typically means translating business documents using personal ChatGPT accounts or other private AI services.

Why is shadow AI a knowledge problem?

When translating, employees teach the AI specialist terminology, tone of voice and product knowledge. Without a shared process, this knowledge does not become part of a centrally maintained knowledge base. Colleagues cannot use it, the company cannot maintain it, and it is lost when employees leave.

Should companies ban private AI use?

A ban alone rarely solves the problem because the need remains. Institut der deutschen Wirtschaft concludes that outright bans would primarily affect committed employees. A more effective approach is an approved solution that is at least as easy to use as a private chat and delivers better results through built-in company knowledge.

Is ChatGPT GDPR-compliant for translations?

A blanket answer is not possible. Business offerings with a data processing agreement and deactivated model training create important prerequisites. Personal accounts are generally not suitable for confidential or personal data-containing company documents.

Is it enough to deactivate training with my data?

No. Deactivation only affects use for model training. Questions about the legal basis, confidentiality, storage location and access remain – as does the knowledge problem.

How can terminology remain consistent in AI translations?

By maintaining specialist terms centrally and assigning them to the relevant translation order, rather than explaining them in individual prompts. Translation Memories also ensure that approved wording is reused.

What does the AI Act require for AI translations?

Article 4 of the AI Act requires companies to take measures to promote the AI literacy of their employees. Since the amendment introduced by Regulation (EU) 2026/1744 of 27 July 2026, they no longer have to guarantee a specific level of competence. Practical training on the opportunities and limitations of AI translation is a useful component.