Lilt

Lilt: Platform And Customers

Lilt: Translation With A Human In The Loop

• Self-learning translation — each human correction improves the model, so quality compounds inside your own content.
• Bring your own LLM or use Lilt's context-aware models, per pipeline.
• ISO 17100 as well as SOC 2 Type II and ISO 9001 — the translation-services standard, not just a security badge.
• Positioned to replace your TMS and language vendors with one platform rather than sitting beside them.
• A separate offering for AI developers: closed benchmarks, oracle judgements and forward-deployed engineers.

How Does Adaptive Machine Translation Learn?

Lilt describes smarter, flexible models with self-learning that improves accuracy with every human interaction, and pairs agentic workflows with expert verification.

That loop is the product, and it is worth understanding why it beats the conventional arrangement.

The standard enterprise setup is a generic engine followed by a human post-editing pass. The engine translates, a linguist fixes it, and the corrections go into the delivered file and nowhere else. Next month the engine makes the same mistakes, because nothing told it otherwise.

An adaptive system feeds the correction back. When a reviewer changes your product name from a translated form to the trademarked one, the model learns that for every subsequent occurrence.

Compounded across a large content estate, that is the difference between paying for the same corrections forever and paying for them once. It also explains why the platform wants your whole volume rather than a project — the learning is proportional to what it sees.

A generic engine plus post-editing pays for the same corrections forever. Adaptive translation pays for them once.

Bring Your Own Model, Or Theirs

Alongside its own context-aware models, Lilt states you can use your own LLM.

That option matters more in translation than in most categories, for a reason specific to the domain.

Some organisations have already built or fine-tuned models on their own terminology, particularly in regulated fields where the vocabulary is fixed and the stakes are high — pharmaceutical labelling, legal filings, financial disclosure. Being forced onto a vendor's model discards that work.

It also hedges a commercial risk. Translation volume is large and recurring, which makes model pricing a material line item, and a platform that can route to a cheaper or better model protects you from one provider's decisions.

The trade is that self-learning presumably works best on the models Lilt controls. Worth asking directly whether the adaptive loop applies equally to a customer-supplied model, because if it does not, the two features are alternatives rather than a combination.

Why Localisation Is Not Just Translation

The use cases listed run well beyond text: website localisation, regional marketing, product launches, brand campaigns, technical documentation, helpdesk and support, professional services, clinical trials, eLearning and regulatory compliance.

That spread is the actual shape of the problem, and it is why localisation projects are consistently underestimated.

Translating a sentence is the visible part. What surrounds it is everything else: strings inside software that break layout when German makes them forty percent longer, date and currency formats, images containing text, legal disclaimers that differ by market, and a helpdesk that now needs answers in seven languages.

Clinical trials and regulatory compliance are the most demanding items on that list, because a mistranslation there is not an embarrassment but a submission failure or a safety issue.

Which is why the certifications matter more here than the feature list, and why the honest framing of this category is that content is the easy half — the same reason a dubbing tool like Dubbing AI solves a narrower problem than a localisation platform does.

• Ask whether self-learning applies to a customer-supplied model.
• Scope the non-text work — layout, formats, images, legal — before pricing.
• Confirm which ISO 17100 scope the certification covers.
• Pilot on a content type you can measure, not your hardest one.
• Keep expert review in the loop where a mistranslation carries risk.

Enterprise Requirements Behind The Words

Lilt publishes three certifications: SOC 2 Type II, ISO 9001 and ISO 17100.

The third is the one worth noticing, because it is specific to this industry and most AI translation vendors do not hold it.

SOC 2 Type II covers security controls and ISO 9001 covers quality management generally — both are common and both are meaningful. ISO 17100 is the international standard for translation services, and it specifies requirements on translator competence, revision by a second person, and process documentation.

That last requirement is the interesting one in an AI context, because ISO 17100 was written around human translators and revision. A platform holding it while running agentic workflows has had to reconcile the two, which is a stronger claim than a self-declared quality process.

The question for procurement is scope. Certifications cover defined services and defined locations, and a certificate held for one workflow does not automatically cover another — so the right ask is which services the ISO 17100 registration actually names. Learning a language is an entirely separate problem, served by tools like Language Reactor and ELSA Speak rather than by a localisation platform.

Replacing A TMS And A Vendor Roster

The commercial pitch is explicit: replace fragmented tools, your translation management system and your language service vendors with one multilingual platform built on agentic AI.

That is an unusually direct thing to say, and it is worth taking at its word rather than as a figure of speech.

Displacing a TMS is a systems project. It sits between your content repositories, your developers, your reviewers and your vendors, and every one of those integrations has to be rebuilt. Displacing vendors is a commercial and relationship change on top.

The upside is real if it works, because that fragmentation is where localisation cost and delay actually live — files moving between systems, versions diverging, nobody able to say what is translated.

The realistic read is that this is a programme with a business case, not a subscription. Which is fine, and it means the evaluation should be run like a platform migration: one content type first, measured against your current cost and turnaround, before anything is decommissioned.

Localising A Product Without Breaking It

Localisation fails at the seams — the string that overflows, the date that reads as the wrong month, the disclaimer that is not legally valid in market.

This sequence puts the expensive discoveries first, while they are still cheap.

1. Inventory everything that needs translating, including strings and images.
2. Test layout with the longest target language before translating anything.
3. Build the glossary and do-not-translate list up front.
4. Pilot on one content type with a measurable baseline.
5. Keep expert review where a mistranslation carries legal or safety risk.
6. Measure whether corrections stop recurring — that is the adaptive claim.