There are two types of AI companies; those that launched before ChatGPT, and those that launched after.
AI translation startup DeepL, which was founded in 2017, is in the former category — but it’s happily riding the wave of the AI boom, founder and CEO Jarek Kutyłowski tells Sifted on the latest episode of the Sifted podcast.
Cologne-based DeepL, which developed out of another company — multilingual dictionary Linguee — has been steadily developing and refining its AI language translation technology for the past seven years. It now counts more than 100k businesses and governments around the world as customers, including Zendesk, Nikkei, Coursera and Deutsche Bahn — along with millions of free users.
So far, 2024 has been a busy one. In January, DeepL expanded to the US — now its third largest market. In May, it raised $300m at a $2bn valuation, in a round led by Index Ventures. ICONIQ Growth, Teachers Venture Growth, IVP and Atomico also invested.
Its team is now approaching 1,000 people globally — but, despite growing its sales and marketing teams to draw in new customers, still more than 60% of the company is focused on R&D. “The DNA of the company is still very tech focused,” Kutyłowski told Sifted on the podcast.
Here are the (lightly edited and condensed) highlights from our conversation. Listen to it in full here.
How has the AI bubble that’s swept up in the past few years affected what DeepL is doing?
It's really interesting, both from the perspective of a researcher and somebody who is involved in the tech, that language has become one of the first frontiers of AI. We had large hopes for robotics, for self-driving cars — automated aspects that we considered easy for AI to have a large influence on. Whereas actually language — something that is so deeply human — has become the topic that everybody's talking about, and where AI has been able to bring in very concrete advantages to humanity.
We've grown up as a company in an environment with very strong competition from the beginning; language translation was not a field that was totally new when we started in 2017. And so the current situation, with even more companies venturing into that area — and with all foundational models capable in some ways to translate — hasn't totally changed our landscape. But we also feel that competitive pressure that is out there. There are now many more people all around the world, many more very smart people who are working on AI in general. And that means that we as a company — but also everybody who wants to participate here — needs to be even faster, even better and grow more.
We need to be even faster, even better
I do think that, essentially, that's something that's good for our users and for our customers. This field is making big leaps, and it will lead to better results. And we as a company have always been very, very focused on research, but doing so for a purpose — in order to change the way that people work with language. And I think, from that perspective, this is really an exciting time.
How does your model work? And how involved are humans in training it?
Translation is very interesting as a field, because on the one hand, you want to be very accurate in how you're trying to map how you're conveying a message in one language to another. And, if you're in a technical area, or if you're in a legal area, it's very important to get all the facts right. But it's also very much about how you're conveying this message, how you're talking; we as humans are very peculiar about how we understand language, how we feel about language, how we receive language. If it's off, then we do not believe the message. And therefore for us, it was always very important to be able to navigate that balance between having a very accurate model for those business situations where accuracy is incredibly important, but also create language that is easy to understand, that people want to read, that people want to listen to.
In order to train AI we still need a lot of human input because — specifically with language — it is humans who can master it the best. When it comes to some of the scientific questions in this world, computers became better at answering them some time ago; computers can add numbers and subtract them much, much better than we can as humans.
When it comes to language, I think humans still hold a pretty large edge over anything that that AI can produce
When it comes to language, and all of its nuances, and all of its intricacies and how it can be interpreted, I think we still hold a pretty large edge over anything that that AI can produce. And therefore it is very important for us to get a lot of this human feedback into the training and into assessing the quality of models.
When it comes to training, it is a lot about telling the AI how we want the translations to look. Models have been trained on the vast amounts of text created and translated on the internet. There's obviously great translations among those. But there's also not so good translations out there. Telling the models that we want to have those high quality translations, that we want them to translate in a specific way, is a lot of what our human input is concerned about. And, because we're working a lot with businesses and enterprises, it's also about validating whether the quality is where we want it to be, because you don't want to have two of your emails translated correctly and the third one fail catastrophically. You want the quality to be consistent so that you can reliably use it for your business needs.
How do your customers tend to use DeepL? And who are your customers?
I think the most important types of industries that we're working with are those that are very document heavy and very text focused; be it professional services companies, legal firms or technical firms, which need to exchange a lot of documentation and a lot of important data where once again, the quality aspect comes in. We're working with a lot of financial institutions and governments too; everywhere where communication and specifically textual communication is incredibly important.
The range of use cases is very broad, starting with emails and casual daily communication. Imagine there's a company which has an office in one country, and an office in a different country, and these offices need to work with each other all the time. There are also much more specific use cases, be it translating legal contracts coming from a different country, translating technical documentation so that your customers can understand it in another country. Plenty of our customers are translating all their customer support inquiries.
In countries where the overall English proficiency is higher DeepL is used more as an efficiency gain — and to master the language a little bit better. Just think about that email that you're sending to your customer, you can write it very well, maybe in a foreign language, but then with DeepL, you're gonna get it even better. And this is maybe gonna give you a few additional conversion points on that interaction or on that deal that you might be doing with this customer.
What’s next on the roadmap?
There are quite a few paths that we're looking at. Making DeepL even better for those professionals that need the last bits of quality is incredibly important. We want to make sure that DeepL can be as usable for everybody who wants to craft high quality texts. And there's still a lot to do, specifically when it comes to the translator and understanding the context you're working in. Quite often the translation is going to look quite different depending on who you're talking to, and what is your past history.
The translation of spoken language is definitely an important one too — and something that we are focusing a lot of our research on.
At the same time, as a research-driven company, we are also looking at how AI is going to develop — and what DeepL’s place in it will be. How much of that is going to be translation? How much of that is going to be DeepL Write [the AI-powered writing companion]? And how much do we want to invest into further areas?
What’s your edge? How do you keep ahead in such a competitive area?
First and foremost, I think it's a lot of fun to be in that tightly competitive space. What we can do very, very well is to apply our research-first approach on a very particular use case in a very particular problem: translation for businesses and for enterprises — where people really need high quality translation most. And that is the focus that we need to have in order to create a slightly smaller playground for us, where we can obsess about quality, where we can obsess about security and make sure that our customers get the best possible product which is then differentiated from the other big players on the market. From that perspective, I think carving out something that is special for you is the most important thing.
It sounds quite tough out there to hire AI talent. How do you convince AI talent to join your company rather than one of the others?
We are one of the major AI companies in Europe — and I think that helps a lot. The fact that we have a strong brand, specifically in Europe and Asia, where people know what DeepL is, and what it can do as a product, helps a lot too. It's much easier to convince anybody who's been using the product already about its usability and how it's going to change the world and what the impact is going to be.
I think Europe has a lot of technical talent. The field of AI is still pretty new, compared to maybe other fields that you would be hiring for. And the pace of change is pretty huge. So I would always try to find smart people who are really passionate about learning, rather than somebody who has a tonne of experience in this particular field — and that we can find a lot in Europe.
Update: DeepL's number of paying and free customers was updated on July 2 when the company supplied additional information.
