Opens in a new tab
English

Useful Links

Why Does EU AI Act Data Governance Start in the First 60 Seconds of a Localization Request?

A legal team uploads 40 witness statements to its language partner’s portal. Before anyone at the partner opens the files, AI is already at work. This intake step rarely comes up in discussions of EU AI Act data governance, yet it is where client content first meets AI.

In under a minute, AI parses each file and counts the words. Next, it classifies the content by type and subject (e.g., medical or legal) and flags key terminology. In many workflows, it also runs a machine-translation pass and scores the output to estimate how much of the translation the AI can handle. Only then does a project manager prepare a quote.

By this point, confidential and sometimes personal data has entered an AI pipeline. In the legal, life sciences, and finance sectors, that data is often regulated. In many cases, a statement of work doesn’t yet cover this job. So where did the content go? Which systems saw it, and did any of them keep a copy?

Localization has long treated data governance as a delivery-stage concern. Now that AI reads and often translates content at upload, governance has to start there too.

Why does AI at intake change the data risk?

Intake, the work of scoping a job and preparing a quote, used to pose little risk to client data. Now, AI does much of that work, often before a statement of work is in place, which puts confidential files at risk.

For years, a client used a portal or an automatic connector from their content system to hand off files. A project manager opened the files, estimated the effort, and sent a quote. The content stayed with the language partner’s staff until the client approved the work. 

AI has changed that. To analyze the files, AI needs a copy of them, and that AI may run on another company’s systems. Depending on that company’s terms, it may keep the files or use them to train its models, without the client knowing.

Clients are aware of the risk. In Slator’s latest buyer survey, 18% of enterprise buyers chose compliance, auditability, and risk control as their top priority for language AI. Only output quality and accuracy (28%) and cost efficiency (20%) ranked higher. The results appear in our joint report, Trusted AI + Human Expertise.

The risk isn’t limited to the language partner. According to Slator’s 2026 Language Solutions & AI Market Report, language AI increasingly operates as a part of a company’s wider AI infrastructure, instead of as a separate localization layer. That makes the data risk at intake part of a company’s wider AI risk. The rules a company sets for its other AI should also cover localization data security.

Where does the EU AI Act stand in late 2026?

Enforcement of the EU AI Act’s transparency and general-purpose AI rules began on August 2, 2026, even though the EU postponed its rules for high-risk AI to 2027 and 2028. Much of the public discussion in mid-2026 focused on that delay, as the Cloud Security Alliance noted in August, but the enforcement date did not move.

Since that date, the European Commission’s AI Office and national authorities can enforce Article 50 transparency obligations, which cover AI-generated content. The AI Office can also enforce the rules for general-purpose AI models, the large models behind chatbots and similar AI applications. Breaches of these rules can lead to fines of up to €15 million or 3% of worldwide annual turnover.

The delay came from an amendment that took effect on July 27, 2026. Under this amendment, known as the Digital Omnibus, AI Act rules for high-risk AI uses, including many hiring tools, will apply from December 2, 2027. For high-risk AI built into regulated products such as medical devices, they will apply from August 2, 2028.

The next deadline is December 2, 2026. Providers of generative AI systems placed on the market before August 2, 2026, have until that day to mark AI output in a machine-readable format. Newer systems have had to comply since August 2. For localization, the impact is limited, because the Commission’s guidelines treat faithful AI translations as exempt from marking. 

Timeline of EU AI Act Deadlines

  • July 27, 2026: Digital Omnibus on AI takes effect
  • August 2, 2026: Enforcement begins (transparency rules and general-purpose AI rules)
  • December 2, 2026: Grace period ends for older generative AI systems to mark AI output
  • December 2, 2027: High-risk AI rules apply in areas such as hiring and credit
  • August 2, 2028: High-risk AI rules apply to regulated products such as medical devices

So where does that leave companies? Enforcement has started and the high-risk rules have fixed dates, so the delay is no reason to relax. Meanwhile, the General Data Protection Regulation (GDPR) still makes a company responsible for how its vendors handle personal data. For now, what protects a company’s data at intake is what it agrees on with its language partner. 

What does Article 50 of the EU AI Act mean for AI translation?

Article 50 requires some AI-generated content to be marked or labeled, but AI translation is often exempt. Under the Commission’s July 2026 guidelines, an AI translation that preserves the original’s meaning and style does not require a machine-readable AI marker. Human review can also remove the need for a visible AI label on public-interest text, as long as a named person or company takes editorial responsibility for it.

This table shows how the marking and labeling requirements differ, including who each one applies to.

Machine-readable markingVisible AI label
Who it applies toProviders (the companies that build and offer AI systems)Deployers (the companies that use AI under their own authority)
What it requiresMarking AI output in a format software can detect, such as a watermark or metadata tagLabeling AI-generated text that is published to inform the public on matters of public interest, such as public health or consumer safety information
How translation is treatedExempt if the translation does not change the substance, meaning, or style of the originalExempt if a qualified person reviews the translation and a named person or company takes editorial responsibility

The marking rule applies to content created by AI. A faithful AI translation creates nothing new because it says the same thing as the original, so the guidelines treat it like a spelling or grammar fix. The Act calls this standard editing, and it doesn’t need a mark. However, summaries and rewrites usually require a mark, and heavily adapted content, such as transcreated marketing copy, may require one as well.

The rules for the visible label are stricter. A published AI translation can go without one only if it meets 3 conditions.

  • A qualified reviewer checks what it says, including the facts. A spell check or a quick sign-off does not count.
  • A named person or company takes responsibility for the published text.
  • AI does not make meaningful changes after the review.

The guidelines even use translation as an example. They describe an AI translation of an article written by a person, which is reviewed by a human, as text that doesn’t need a label.

For machine translation, AI Act labels depend most on the third condition. For example, a company reviews and approves an English article on a public health topic, then uses machine translation to publish it in German. Because machine translation happens after the review, the English review doesn’t cover the German version. To publish the German version without an AI label, the company needs a qualified reviewer to review it too.

This means human review now matters for both compliance and quality. Some law firms already require it due to liability concerns. A former law firm translation manager describes the practice in Slator’s 2026 market report.

“There is no such thing as a non-post-edited or non-revised machine translation in a law firm,” the manager says. “This is because of liability issues.”

Does GDPR stop your content from training someone else’s AI model?

Not on its own. GDPR limits how a language partner can use personal data. Beyond that, whether an AI model can train on your content depends on your contract with the language partner and the terms of its AI providers. 

GDPR’s limits may change under a separate proposal known as the data omnibus, which would revise the rules on using personal data to train AI. On October 21, 2026, the Council of the EU agreed on its position, which keeps the current definition of personal data and adds safeguards when companies use sensitive data, such as health information, to develop AI.

The European Parliament still has to agree on its own version of the proposal before talks on a final law can begin. Until a final law is in place, the current GDPR applies.

Clients don’t need to wait for the final law. GDPR already requires a written contract with any vendor that processes personal data for a company. Clients can use that contract and their master services agreement to set where their language partner processes their data and whether the partner can keep data after the job.  They can also state whether any AI model, including the partner’s own, may train on it.

Redacting personal data before translation adds a technical safeguard, although people still need to check what AI redacts. Slator’s market report finds that regulated buyers use language AI more widely when they can limit risk, keep their data separate, and protect privacy.

What should clients ask their language partner about AI and data?

Before sending the first file, clients should ask how a language partner handles their content. Procurement and legal teams can start with these questions that cover AI translation data privacy, human review, and accountability.

  • Which AI systems process our files during intake?
  • Can you redact personal data before AI processes our files?
  • Do any third-party AI providers receive our content, and on what terms?
  • Where do you process and store our content, and for how long?
  • What happens to files we upload for a quote if we don’t go ahead with the job?
  • Can any AI model, including your own, train on our content?
  • Which content types receive human review, and who performs it?
  • Who holds editorial responsibility for translations that we publish?
  • Can you document your human review if a regulator asks for evidence?

What does EU AI Act data governance look like at intake?

At intake, data governance means the client and language partner agree on how the partner will handle content before uploading the first file. This includes which AI sees the content, where it is stored, and who reviews the results. The partner then builds these rules into how the AI localization pipeline runs. 

Slator’s market report finds that as text translation becomes a commodity, language partners differentiate themselves by how they integrate, govern, and secure AI, including deploying it in controlled environments. That makes secure AI translation and data governance part of what clients buy.

Our approach combines Trusted AI and Human Expertise, and VistatecMAP (Managed AI Pipelines) provides a practical framework for putting that strategy into action. Before production begins, we agree on a quality standard with the client for each content type. We then prepare the client’s translation memories, glossaries, and style guides for use with AI. Together with our technology partner Phrase, we configure and fine-tune the pipeline against that standard.

Once the pipeline is running, every translation goes through automated quality checks. Our linguists review the results and fix recurring errors at the source, so fewer translations need correction over time.

For companies that prefer to set up their own controls, our AI Governance services provide frameworks for compliant, auditable AI localization, with traceable records of AI decisions and human oversight. We also work with companies that build their own AI systems. Through Vistatec Data, we bring localization-grade quality to AI data.

To see how enterprises are redesigning review, quality, and governance for language AI, download Trusted AI + Human Expertise, our new research report with Slator. To discuss governed AI pipelines for your own content, talk to us. 

Frequently asked questions

Does the EU AI Act cover files we upload for a translation quote?

Not directly. The EU AI Act’s data governance rules cover the data used to train and test high-risk AI, and they won’t apply until December 2027 at the earliest. Files uploaded for a quote are subject to different rules. GDPR applies to any personal data in those files, and the contract with the language partner should cover confidential business information. Clients can agree on those terms now, without waiting for the AI Act’s remaining deadlines. 

Does AI-translated content need an AI label?

In most cases, no. Under Article 50 of the AI Act, AI Act labels apply only to text published to inform the public about matters of public interest, such as public health information. Even then, a translation doesn’t need a label if a qualified person reviews it, checks the facts, and a named person or company takes editorial responsibility for it. The review of the translation has to come last, with no meaningful AI changes afterward.

Note that a faithful translation doesn’t need a machine-readable AI marker, as long as it keeps the substance, meaning, and style of the original.

Do translated internal documents or product descriptions need an AI label?

Usually not. The visible AI label applies only to text published to inform the public about matters of public interest, such as public health or fundamental rights. Internal documents, such as company intranet posts, aren’t published to the public, so they don’t need a label. Product descriptions and ads don’t need one either, unless they make claims about matters such as health, consumer safety, or sustainability.

Who is responsible for AI labels and personal data, our company or our language partner?

For AI labels, it depends on who decides how AI is used. Under the AI Act, the company that makes that decision is the deployer, and the visible label is its responsibility. It stays the deployer even if contractors or freelancers run the AI on its behalf. A company that only orders translation and leaves AI decisions to its language partner is not the deployer. 

For personal data, a company stays responsible under GDPR, regardless of who uses the AI. Clients can use their contracts to state who decides on AI use and who takes editorial responsibility for published translations.

Do you want to know more?

Talk To Us