Webinar
November 20, 2026
|
5 min read
Built by One Team, Used by Everyone: Taking Localization AI Company-Wide
Generic AI is now used across marketing, product, support, and legal, not just the localization team. In a Slator webinar, LILT and Vinted shared seven lessons for taking localization AI company-wide: start with trusted assets, pilot the right content, tier review by risk, and make the governed path the easiest.
LILT Team
AI translation is no longer something only the localization team uses. Marketing, product, support and legal teams are all generating and translating content on their own, often in public AI tools. The question for localization leaders has shifted from "should we use AI?" to "how do we make our governed AI the easiest option for everyone, without losing quality or control?"
In a recent Slator webinar, Kearney Hughes, Director of Customer Success at LILT, and Justina Dolgovičienė, Localization Manager at Vinted, where she leads AI projects in localization, shared how they approach that problem together. Vinted's program runs on four in-house localization project managers, around 60 freelance linguists and two regular vendors, and covers everything from customer support to marketing, product and legal content. As Justina put it, "it'd be difficult to find a localization workflow today where AI doesn't play any role."
Here is what they've learned about taking localization AI company-wide.
Seven lessons from the webinar
1. Generic AI gives you good moments, not consistency
The problem with generic machine translation or a public LLM isn't that it can't produce a decent translation. "It's that a good result in that one moment... doesn't necessarily become consistent across teams, languages and content," Kearney explained. Teams end up with different terms for the same product, brand voice starts slipping within weeks, and the business loses visibility into what is being translated and where sensitive data is going.
Justina shared a "funny, not funny" example. While generating French SEO content for Vinted's web pages, between rounds of prompt editing and QA, the output read "clothes for secondhand kids" instead of "secondhand clothes for kids." The word order was inherited straight from English, even with style guides and terminology in place. Her rule of thumb: "The less context is in the content itself, the more additional context you have to provide."
Part of the problem: not every AI model learns from your reviewers. A general-purpose LLM or generic machine translation engine doesn't absorb your team's corrections, so reviewers end up fixing the same mistakes again and again. LILT's models learn from reviewer edits in real time, so each correction improves the next translation. As the webinar framed it, review corrections become "learning that compounds instead of repeating."
2. The real blocker is fragmentation, not technology
Even with an advanced AI setup, what slows Vinted down most is "lack of operational consistency between different teams that require localized content." Every team has its own tools and ways of working. "It's workarounds on workarounds," Justina said, and that is hard to fold into one workflow, however good the AI is. She also warned against underestimating the cost of maintaining something you built in-house: building may feel easy, but tracking and maintaining it takes real time and money.
3. Start with assets you trust, then pick the right pilot
Because LILT's models learn from every reviewer edit, what you feed them matters.
Before launching a model that learns from your reviewers in real time, Kearney's advice is to start not just with what you have, but with what you know is good. Feeding a learning model content you aren't confident in can degrade its quality, and long-running programs often carry years of translation memory that needs a quality check first. Own your TMs, termbases and style guides, know how they're stored and exported, and make sure "you shouldn't have to start from scratch if your tools or partners change."
To prove the difference between generic AI and a model that learns from your corrections, Justina picks a first domain that is:
- Long form, so the content carries its own context
- Repetitive in structure, because "machines love structure"
- Accuracy-critical rather than creative
- High volume, so the model has plenty to learn from and the payoff is biggest
- Low financial and legal risk
Her example: FAQ entries. And how quickly do results show? With models that learn and adapt to your content, Kearney said, "you could see improvement in a matter of weeks," provided you set a baseline and review quality together.
4. Tier review by risk, content type and language
Not all content needs the same safeguards. Vinted sets risk tiers with stakeholders based on legal and financial risk, relevance to users (for example, traffic to a page) and the nature of the content. Product content can be edited at any time, but "once you send marketing communication, it's out there."
Ambiguity matters too. In Vinted's catalog, "accessories" can mean jewelry and scarves or add-ons for a mobile phone, and generic MT "has no way of knowing that." Unless the AI can be given that context, that content gets a human in the loop.
Before moving any content type to AI only, Kearney recommended a representative sample reviewed language by language, quality scores from qualified reviewers against an agreed standard, a clear view of what's at stake if there's an error, and a check-in afterwards to confirm quality still holds. Justina added a baseline, a clear definition of success, edit and QA data and, where possible, business impact from an A/B test.
5. Use agents that fix, not just flag
Kearney drew a sharp line between AI QA tools that point out issues and agents that act on the content. At LILT, agentic workflows make a first pass on the AI output and fix repeat issues, which "give our human expert reviewers more time to focus on nuance and judgment" on brand voice and the things that matter.
The question isn't whether to remove humans, she said, but "where do humans belong, and where is expert verification in that agentic workflow?" There's an upfront lift, cleaning up TM and termbases and investing reviewer time to fine-tune models, "perhaps it takes a few weeks, perhaps it takes a few months," but the improvement shows up in the analytics.
For regulated or legal content, neither speaker would start with AI only. Justina's approach: use AI translation as the first workflow step and monitor the edits. Even if a reviewer stays on, "it takes them like 10 minutes to go through the whole document instead of five hours."
6. Speak leadership's language when someone asks "can we just use AI?"
LQA scores rarely land outside localization, Justina found. Bring metrics stakeholders care about instead. Vinted ran an A/B test on marketing content, human versus AI, and monitored user engagement.
Kearney recommended a scorecard tied to the success criteria leadership already cares about. "It's not as simple as, would AI only be cheaper?" she said. The fuller picture is speed, quality cost, edit and review effort, and reach: what you give up if you can't connect with the right markets in the right brand language.
On cost savings, Justina was candid, quoting a Lithuanian saying: better to be poor but honest. Cost cuts sometimes do mean quality cuts. That can be fine if the risk is taken consciously, agreed with stakeholders in advance and monitored. The key, she said, is monitoring and communicating.
7. Make the governed path the easiest path
How much translation happens in public AI tools across Vinted? "The short answer is I don't know," Justina said, and she doesn't think it can be stopped. Instead, her team offers help: QA, prompt suggestions, and a solution that is better and more convenient than what teams already use. "If you have something that already works and is practically effortless, [people] will not try to build something of their own."
Kearney described the target experience: marketing, product and support should request translation where they already work, without needing to understand "your complex localization setup." "The whole ethos of LILT is that we want to meet people where they already are." Meanwhile the localization team defines approved terminology and workflows, decides where human review is needed, and monitors it. "The access becomes broader... but then the localization team's expertise and control doesn't disappear." She noted customers already calling their LILT models through MCP and from inside their AI assistants.
Will AI shrink the localization role?
Both speakers said no, if localization leaders take the lead. "It's okay to worry," Justina said, but the choice is to "seize the control and define how it is used in your company" or step back and let others do it. "If you take control early enough, I feel like you may also make your role way bigger than it was before."
Asked by an attendee whether enabling self-serve translation means optimizing yourself out of a job, Justina pointed to everything that remains: managing resources, monitoring data, making decisions and designing workflows. The role moves from "someone who sends the documents from one person to the other" to "a strategic partner." She expects new roles too, such as localization-focused data analysts and automation engineers, with quality needing more attention, not less.
Kearney agreed: with more content than ever, governance of brand voice and data "is more important than ever." And Justina offered a reminder about volume. People owned one pair of shoes 150 years ago and many more today. "The same thing will happen with content," heading toward hyper-localization and hyper-personalization.
Your company-wide localization AI checklist
- Audit your language assets. Know which TMs, termbases and style guides you trust, and make sure you own and can export them.
- Set a baseline and define success with stakeholders before you change anything.
- Pilot on the right content. Long form, structured, accuracy-critical, high volume and low risk, like FAQ entries.
- Tier review by risk, content type and language, not one rule for everything.
- Use agents that fix, not just flag, so expert reviewers spend time on nuance and judgment.
- Earn AI-only status with evidence: representative samples, quality scores, edit rates and scheduled re-checks.
- Report in business terms: engagement, speed, reach and effort, not only LQA scores.
- Meet teams where they work, with governed models inside the tools they already use.
- Keep integrations agnostic and avoid workarounds, so you can change models or partners without rebuilding.
Built by one team, used by everyone
The thread running through the conversation: AI widens who can access language, and localization teams set the rules for how it's done. As Justina said of the Vinted and LILT partnership, "It's teamwork."
LILT helps organizations make their content and data multilingual, faster, more accurately, securely and at scale, with models that learn from your reviewers, agents that fix issues before a human sees them, and integrations that bring governed translation into the tools teams already use.
Watch the full webinar on demand for the complete conversation, or talk to LILT about taking your localization AI company-wide.
Share this post
Find some time with LILT
Enterprise-grade content seamlessly translated with AI to help your business scale globally.
Book a MeetingShare this post