{"id":1653,"date":"2026-06-16T22:00:29","date_gmt":"2026-06-16T22:00:29","guid":{"rendered":"https:\/\/www.ata-divisions.org\/LTD\/?p=1653"},"modified":"2026-06-16T22:50:52","modified_gmt":"2026-06-16T22:50:52","slug":"ata-tektalks-languagecheck-ai","status":"publish","type":"post","link":"https:\/\/www.ata-divisions.org\/LTD\/ata-tektalks-languagecheck-ai\/","title":{"rendered":"ATA TEKTalks – The Review Bottleneck: When Your Translation Workflow Needs a Wingman – Is LanguageCheck.ai the Right Tool for You?"},"content":{"rendered":"\n

By Flor Russo<\/em><\/h3>\n\n\n\n

This post is based on the ATA TEKTalk 5.1 session, featuring Marco Baglioni and Greta Cavaliere of LanguageCheck.ai<\/a>, hosted by the ATA Language Technology Division on May 14, 2026.<\/p>\n\n\n\n

About the Lead Presenter:<\/h3>\n\n\n\n

Marco Baglioni<\/strong> is the chief executive officer and co-founder of LanguageCheck.ai and founder of the language technology company Aqrate. With 25 years of experience in the translation and localization industry, he has worked at the intersection of language services, workflow automation, and AI-driven quality evaluation. Marco focuses on developing technologies that help translators, language services providers, and enterprises scale translation quality in AI-assisted workflows. His work centers on practical applications of AI for translation review, risk detection, and quality scoring. Through LanguageCheck.ai, he is advancing new approaches to translation quality management that combine human expertise with AI-assisted evaluation.<\/p>\n\n\n\n

We’ve all been there. A big project lands in your queue, the machine translation or legacy TM has done its thing, and now you’re staring at thousands of segments wondering how on earth you’re going to review them all without losing your mind -or your margins. Marco Baglioni and Greta Cavaliere from LanguageCheck.ai<\/strong> sat down with over 200 attendees to talk honestly about AI in translation review: what works, what doesn’t, and why the human in the loop is still very much the star of the show.<\/p>\n\n\n\n

Spoiler alert. This is not about a rebellion of robots. It’s a story about a really good sidekick.<\/p>\n\n\n\n

How Review Became the Bottleneck of the Industry<\/strong><\/h3>\n\n\n\n

There is a brand new pain point for the industry, and it’s not even the translation itself anymore. With MT and project volumes growing, it turns out that quality assurance has now turned into the costly and tedious process no one has expected to see. LanguageCheck.ai was actually born from this exact frustration. Aqrate, the language services company behind the tool, invested two and a half years into creating their solution by running countless tests and learning what AI can and cannot be applied to.<\/p>\n\n\n\n

That kind of backstory matters. This is not just a new flashy tool created overnight thanks to a couple of prompts and fancy slides.<\/p>\n\n\n\n

Human in the Loop \u2014 Never Out of the Picture<\/strong><\/h3>\n\n\n\n

The philosophy at the heart of LanguageCheck.ai is called the “human-on-the-loop” model. Notice it’s in<\/em> the loop, not out<\/em> of the loop. The machine does the scanning; the human makes the calls.<\/p>\n\n\n\n

What did the LanguageCheck.ai team discover early on? AI proves itself as much more effective at evaluating preexisting content than generating it. As Marco put it during the TekTalk, it’s like a child who can’t cook a five-course meal but will very confidently tell you if the soup is too salty. The same technology that produces machine translation turns out to be quite good at detecting mistakes in it, including its own.<\/p>\n\n\n\n

The practical upshot? The tool categorizes every translation unit into one of three buckets: flawless, needs improvement, or incorrect meaning. Flawless segments\u2014 which the tool identifies with 98\u201399% reliability\u2014can be skipped entirely. That typically frees up reviewers to focus only on the 30% or so that actually needs human attention, delivering an average time and cost saving of around 60-70%.<\/p>\n\n\n\n

The machine filters the noise. The human decides what to do with what’s left.<\/p>\n\n\n\n

Fitting into Your Existing Workflow<\/strong><\/h3>\n\n\n\n

One of the more practical things about the tool is that it doesn’t ask you to abandon your existing setup. You upload bilingual files: XLIFFs from Trados, MemoQ, XTM, Phrase, and others, run the check, and then choose how you want to work with the results.<\/p>\n\n\n\n

Three options to consider when handling the results. You can download an .xlsx report with all the analysis performed offline-friendly and shareable to third-party parties. Or, you can get an annotated XLIFF file, which you can further edit in your CAT tool, with all the remarks and suggestions added. And lastly, you can post-edit directly on the LanguageCheck website. The editing screen enables you to filter segments by severity: start with incorrect meanings and major errors, and finish with everything else.<\/p>\n\n\n\n

When the Machine Stumbles: Knowing the Limits<\/strong><\/h3>\n\n\n\n

It’s important to note that even the best tool is prone to mistakes and has some limitations, which is why it’s necessary to know the ones this tool has.<\/p>\n\n\n\n

Short segments are a known weak spot.<\/strong> One or two-word strings, acronyms, abbreviations, spare part lists; when a segment doesn’t give the AI enough context to work with, you’re going to see false positives. It doesn’t mean the segment is wrong; it means the machine is guessing in the dark.<\/p>\n\n\n\n

Consistency is not automatic.<\/strong> Without a glossary attached, the tool may evaluate the same term differently across a document depending on context. Attaching terminology goes a long way toward solving this, but it’s worth going in with realistic expectations.<\/p>\n\n\n\n

Transcreation is a hard no.<\/strong> If you’re working on marketing copy, slogans, or anything where creative adaptation is the point, keep your human reviewer front and center. The tool is built around accuracy and meaning fidelity, not poetic flair.<\/p>\n\n\n\n

Paragraph or document-level analysis is not yet available.<\/strong> The tool works segment by segment, which is where it performs best today. Expanding to larger chunks tends to degrade performance \u2014 more on why in the next section.<\/p>\n\n\n\n

A Lesson in Effective Communication With AI<\/strong><\/h3>\n\n\n\n

Here comes a lesson the developers had to spend a significant amount of time to realize. Contrary to the intuition, giving additional information on how AI should behave and what the results should be doesn’t always produce better outcomes.<\/p>\n\n\n\n

Additional information in the form of stricter constraints, style guides, guidelines, and so on made the results of the AI deteriorate due to what the team calls the “butterfly effect”: correcting one issue in the prompt would affect another part of the work negatively.<\/p>\n\n\n\n

This holds true even for the scope of the work. Providing AI with a longer text to process resulted in deterioration of performance. Sometimes, it would also change the behavior of the AI making it alter the source text.<\/p>\n\n\n\n

The moral of the story here is quite simple: be concise in your communication with AI. The tool will do a way better job in handling structured information rather than text-heavy instructions. If your goal is to achieve a consistent terminology, don’t write long guides on how to handle it correctly. Simply create a glossary.<\/p>\n\n\n\n

Confidentiality: Your Data Isn’t Floating Around in the Cloud<\/strong><\/h3>\n\n\n\n

The data security aspect is especially important if your company works with sensitive data (for instance, legal, medical, or financial documents). In the current reality, confidentiality might be either a deal-breaker or a green light for your work.<\/p>\n\n\n\n

Fortunately, LanguageCheck.ai has obtained a GDPR certificate, and, when uploading a file, you decide from the start where it will be located (in a US- or European-based Google Cloud). The segments are sent to the machine-learning algorithm via API to perform the analysis, but crucially, the system doesn’t collect data to train the algorithm.<\/p>\n\n\n\n

Marco was clear during the session about why the “your data will improve our model” claim doesn’t hold up even from a technical standpoint: to make a meaningful difference in AI performance today, you would need billions of data points. A few translation projects from your clients simply aren’t going to move that needle, and any tool claiming otherwise deserves skepticism.<\/p>\n\n\n\n

It’s crucial to differentiate two approaches. Copy-pasting your confidential data into a publicly available AI interface is much worse than direct integration with the professional API service (such as what LanguageCheck.ai does).<\/p>\n\n\n\n

The former makes it possible for the model to hallucinate and alter your data instead of doing quality assurance, which leads to poor results. In turn, the latter allows you to get much more predictable results from a model that you are sure to handle correctly.<\/p>\n\n\n\n

How Does It Work in Practice<\/strong><\/h3>\n\n\n\n

Two scenarios for using the tool emerge as the most common ones:<\/p>\n\n\n\n

1) Quality assurance after a human translation. You can use it as the last step in your review procedure to catch the overlooked mistakes and improve overall quality.<\/p>\n\n\n\n

2) The most frequent scenario is to use the tool for filtering your MT before you start post-editing.<\/p>\n\n\n\n

Neither of them makes you lose your translator or reviewer. AI identifies possible issues, and a human decides what to do with them.<\/p>\n\n\n\n

A Hint on Evaluating AI Tools (Since Everyone’s Doing It These Days)<\/strong><\/h3>\n\n\n\n

During the live discussion of this presentation, a recurring question was how to assess different AI tools, given how many companies are selling a bright future based solely on some fancy demos. Marco’s recommendation was to try things, conduct trials, and not jump to conclusions.<\/p>\n\n\n\n

Specifically, the LanguageCheck.ai team tried numerous AI engines and concluded that in some cases, an older version of the model performs better than a new one. It’s because engines vary in their capabilities. In addition to that, certain engines may be better at processing certain types of documents.<\/p>\n\n\n\n

So, before trying any AI tool, evaluate it against your needs. Run tests on different types of files you work with: short or lengthy segments, technical or creative content. Don’t judge the results by some random file the vendor has shown to you as a demo. Judge the tool based on your work.<\/p>\n\n\n\n

The Bottom Line<\/strong><\/h3>\n\n\n\n

This is a tool for people or people who know that AI shines in boring, routine tasks, leaving complex and demanding jobs for humans to take care of.<\/p>\n\n\n\n

Review is the new bottleneck. But you don’t have to review everything. The machine can tell you where to look. What you do when you get there is still entirely up to you.<\/p>\n\n\n\n

\n\n\n\n

Q&A<\/strong><\/h2>\n\n\n\n

The following Q&A were addressed after the end of the session as a complement of information:<\/p>\n\n\n\n

Q:<\/strong> Is this a subscription tool? For a small agency or teamwork, PM,  translator, proofreader. Do we need different subscriptions for different users in the process?<\/p>\n\n\n\n

A:<\/strong> LanguageCheck.ai operates on a pay-per-use model. However, you can purchase discounted packages of words, with savings of up to 30%.
Here you can find all the details https:\/\/languagecheck.ai\/#pricing
<\/a>There is a Pro version that gives you the opportunity to use some of the filters shown during the demo. Even in this case you can find the information on our website https:\/\/languagecheck.ai\/#pro<\/a><\/p>\n\n\n\n

A Team version is going to be released in June. Right now, multiple users can access the same account and run checks at the same time.<\/p>\n\n\n\n

Q:<\/strong> How does the tool react to punctuation which may differ from one source language to another target language (e.g., em-dashes)<\/p>\n\n\n\n

A:<\/strong> LanguageCheck.ai takes into account the specifics of each language and flags only the relevant issues for the target language set. As an instance, if you have a source segment in English with an em-dash and in the Italian translation, instead of the em-dash you have a semi-column or a comma (depending on the case), the tool won\u2019t flag anything.<\/p>\n\n\n\n

Q:<\/strong> Can you tell us more about the origins of LanguageCheck.ai and how Aqrate, an Italian company, came to build a globally-oriented translation QA platform?<\/p>\n\n\n\n

A:<\/strong> At the end of 2022, we began exploring the use of AI for quality assurance, as QA had become an important area of focus for Aqrate. We were investing a significant amount of time in ensuring translation quality, but this effort was not generating the expected financial return<\/p>\n\n\n\n

Throughout 2023, we developed the first version of our tool, running extensive tests and making several adjustments along the way. It took almost a year to create a version of LanguageCheck that could be applied to real-world use cases, such as translation memory cleanup.<\/p>\n\n\n\n

This initial experience was valuable because it allowed us to collect feedback from the vendors responsible for the final TM cleanup, who used our reports as part of their work.<\/p>\n\n\n\n

Based on those results, we decided to invest further in the development of a standalone translation QA tool, with a specific focus on bilingual formats. At the same time, we began exploring possible integrations with CAT tools and TMS platforms.<\/p>\n\n\n\n

Q:<\/strong> Can you explain how LanguageCheck.ai can make the lives of freelancers or LSPs easier or more lucrative \u2013 especially when working with high-volume or high-stakes content like legal, pharmaceutical, or technical texts?<\/p>\n\n\n\n

A:<\/strong> Our users are already seeing measurable results: on average, they are saving 70% in time and costs, exactly as expected. This allows them to increase productivity, improve margins, and potentially handle up to three times more work with the same resources.<\/p>\n\n\n\n

LanguageCheck makes this possible by identifying the translations that require revision and providing a detailed explanation of the errors detected. Reviewers can therefore focus only on the content that needs attention, rather than checking the entire translation manually.<\/p>\n\n\n\n

The process is especially effective when LanguageCheck is used after machine translation to reduce post-editing effort on large projects. When a glossary is attached, LanguageCheck can also verify terminology consistency, making it particularly useful for legal, pharmaceutical, and other technical content.<\/p>\n\n\n\n

Q:<\/strong> What can an average user expect in terms of costs when utilizing your tool?
 (Pricing model, subscription vs. pay-per-use, cost of support, any add-on services, etc.)<\/p>\n\n\n\n

A:<\/strong> LanguageCheck.ai costs only a fraction of a traditional review process, while helping you reduce review costs by an average of 70%. Instead of checking the entire translation manually, reviewers can focus only on the segments that require attention.
On our website you can find further information on pricing model and subscriptions https:\/\/languagecheck.ai\/#pricing<\/a><\/p>\n\n\n\n

By the way, at the moment we do not charge costs for our support services.<\/p>\n","protected":false},"excerpt":{"rendered":"

We’ve all been there. A big project lands in your queue, the machine translation or legacy TM has done its thing, and now you’re staring at thousands of segments wondering how on earth you’re going to review them all without losing your mind -or your margins. Marco Baglioni and Greta Cavaliere from LanguageCheck.ai sat down with over 200 attendees to talk honestly about AI in translation review: what works, what doesn’t, and why the human in the loop is still very much the star of the show.<\/p>\n","protected":false},"author":10,"featured_media":1654,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_genesis_hide_title":false,"_genesis_hide_breadcrumbs":false,"_genesis_hide_singular_image":false,"_genesis_hide_footer_widgets":false,"_genesis_custom_body_class":"","_genesis_custom_post_class":"","_genesis_layout":"","jetpack_post_was_ever_published":false,"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[11],"tags":[13,39],"class_list":{"0":"post-1653","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","7":"category-ata-tektalks","8":"tag-ata-tektalks","9":"tag-languagecheck-ai","10":"entry"},"jetpack_featured_media_url":"https:\/\/www.ata-divisions.org\/LTD\/wp-content\/uploads\/2026\/06\/mohamed_hassan-business-4576778-scaled.jpg","jetpack_sharing_enabled":true,"jetpack_shortlink":"https:\/\/wp.me\/pQoPc-qF","_links":{"self":[{"href":"https:\/\/www.ata-divisions.org\/LTD\/wp-json\/wp\/v2\/posts\/1653","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ata-divisions.org\/LTD\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ata-divisions.org\/LTD\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ata-divisions.org\/LTD\/wp-json\/wp\/v2\/users\/10"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ata-divisions.org\/LTD\/wp-json\/wp\/v2\/comments?post=1653"}],"version-history":[{"count":1,"href":"https:\/\/www.ata-divisions.org\/LTD\/wp-json\/wp\/v2\/posts\/1653\/revisions"}],"predecessor-version":[{"id":1655,"href":"https:\/\/www.ata-divisions.org\/LTD\/wp-json\/wp\/v2\/posts\/1653\/revisions\/1655"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ata-divisions.org\/LTD\/wp-json\/wp\/v2\/media\/1654"}],"wp:attachment":[{"href":"https:\/\/www.ata-divisions.org\/LTD\/wp-json\/wp\/v2\/media?parent=1653"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ata-divisions.org\/LTD\/wp-json\/wp\/v2\/categories?post=1653"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ata-divisions.org\/LTD\/wp-json\/wp\/v2\/tags?post=1653"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}