LILT AI

01Who they are, right now

An enterprise translation platform, in the awkward middle of becoming something more.

LILT, headquartered in the Bay Area with roughly 540 people on payroll, sells an AI platform that does one unfashionable thing extremely well: it translates the operating documents of large companies and government agencies into more than a hundred languages, and learns from every correction a human reviewer makes. The translation industry has spent the better part of a decade arguing about whether neural networks would replace translators. LILT, with the diplomatic restraint of a company that actually employs translators, has been arguing something subtler. The model and the human are the product. Neither one ships without the other.

That stance has aged surprisingly well. While the rest of the market lurched between "pure machine" and "pure human" pitches, LILT quietly built a platform - Assist, Verify, Connect, Translate, Converse, Train - that treats translation like an engineering discipline. Pipelines. Quality gates. Feedback loops. The kind of thing a procurement officer at Intel can put a price on.

02The problem they saw

Machine translation got good. Translation workflows did not.

Anyone who has worked inside a global company knows the loop. Marketing writes English. A vendor portal swallows it. Weeks pass. Strings come back in a spreadsheet that nobody can quite reconcile with the CMS. Someone re-uploads the file, and a typo introduced in transit ships to forty markets. The actual translation is often fine. The system around it is held together with goodwill and frequent flyer miles.

Underneath that operational mess sits a stranger problem. Off-the-shelf machine translation does not know your company. It does not know that your product is "the Platform" and not "the platform," that you call customers "members," that your legal team refuses the word "guarantee." Every Tuesday, somebody fixes the same mistake the model made last Tuesday. The model never finds out.

03The founders' bet

Two Google Translate researchers walked out with a contrarian idea about feedback.

Spence Green and John DeNero met at Google, where both worked on the systems that power Google Translate. Green was finishing a Stanford PhD on interactive machine translation. DeNero, by then teaching computer science at UC Berkeley, had spent years on statistical translation models. In 2015 they left to start LILT. The pitch was strange enough that polite venture capitalists asked them to repeat it.

Their bet: the next decade of translation would not be won by whoever built the largest model. It would be won by whoever built the tightest feedback loop between the model and the experts who corrected it. Adaptive machine translation - where the system updates from human edits in something close to real time - was, in 2015, a research curiosity. LILT made it the product. The model improves by living next to its users instead of being shipped to them every quarter.

The thesis quietly attracted serious money. Sequoia Capital led the seed, returned for the Series A, the Series B, and the Series C. Redpoint and Intel Capital joined in. By 2022, the company had stacked four rounds totaling more than ninety-five million dollars. That is not a vanity number. That is what it costs to run enterprise sales cycles and a research lab at the same time.

04The product

Seven things that, when you squint, are one thing: an enterprise translation operating system.

The platform is best understood as a content supply chain rather than a translator-in-a-box. Source content flows in through connectors. Adaptive models do the first pass. Reviewers - sometimes in-house, sometimes from LILT's managed network - polish the output, and every keystroke teaches the model something about how this particular customer talks. The fixed pipeline pieces have names.

LILT was also, somewhat unfussily, one of the first translation companies to ship a Model Context Protocol server and an Agent-to-Agent card - hooks that let other AI agents plug LILT directly into their workflows. If you have read this sentence and shrugged, you are not the target audience. If you have read it and immediately thought of a use case, you are exactly the target audience.

05The proof

Customers, capital, and a chart that makes the trendline visible.

LILT's public customer roster reads like a tour of categories where mistranslation is expensive. Intel, where firmware documentation is a regulated artifact. Emerson, where industrial manuals have to be precise in any language. Juniper Networks. Orca Security. Quietly, government contracts in defense and intelligence, where the technology has a slightly different name and a much shorter description.

The pattern is the dull kind of pattern that founders quietly love. Each round is roughly two to three times the last, the same lead investor keeps writing checks, and the customer logos keep accumulating around use cases that touch revenue, not just marketing. That is the difference between translation as a cost center and translation as a growth lever - which is, conveniently, the difference LILT spends most of its sales calls explaining.

06The mission

Stated plainly, then stress-tested.

LILT's official mission is to "make the world's information available to everyone, no matter the language they speak." It is the kind of sentence that gets engraved on lobby walls. It is also, importantly, the operating principle of the product. Every connector, every adaptive model, every reviewer workflow is in service of a single output: more of the world's content, available natively, in more languages, with fewer humans copy-pasting into spreadsheets at midnight.

The vision is more ambitious, and a little harder to say with a straight face: every digital product and service, available in every language on the internet. The honest version is that LILT does not need to ship that vision alone. It needs to ship the rails. Other people's content rides on them.

07Why it matters tomorrow

Agents are about to need translators of their own.

Here is the awkward thing about agentic AI. The agents are very good at English. Many of them are passable at the largest commercial languages. Almost all of them get worse, often dramatically, in everything below the top fifteen. As agents take on more real work - filing tickets, drafting policy, talking to customers - the gap between "English-language agent" and "useful global agent" widens into a chasm. LILT is one of a small number of companies positioned squarely inside that chasm, with infrastructure that other agents can call.

This is the part the company does not need to oversell. The trendlines do the work. More multilingual content. More regulated multilingual content. More agents needing reliable language layers underneath them. A platform whose business case improves every time someone, somewhere, files a customer support ticket in a language the model has never seen.