Verbit

Why 95 percent accuracy is not enough

In transcription, the difference between 95 percent accuracy and 99 percent accuracy is not a difference of four percent. It is the difference between unusable and usable. A transcript with one error in every twenty words means roughly one mistake per line, and anyone reading it has to listen to the recording to check it. In other words, they have to redo the work they paid for. In some settings this bar is not a matter of convenience but of law: a court record, an academic lecture made accessible to a deaf student, or a television broadcast that is required to carry captions. Verbit, founded in November 2016 by Tom Livne, Eric Shellef and Kobi Ben Tzvi, built its entire business model around exactly this gap: an automatic model produces a fast, cheap draft, and a human transcriber brings it up to a level the customer can actually use. It is also worth understanding why this is so hard. Human speech is full of disruptions that the eye does not see in written text: people talk over each other, swallow words, speak with accents, and use professional terms that do not appear in any general dictionary. It is precisely in the places where accuracy is critical, such as a courtroom or a university classroom, that all of these conditions occur together.

Human in the loop as architecture, not compromise

This approach is called human in the loop, and at Verbit it was not a transition stage on the way to full automation. It was the product itself. Behind the system stood a huge network of freelance transcribers (according to Hebrew Wikipedia, about 35,000 transcribers in about 120 countries) who worked on drafts the model produced. The organization needed to run a network of this size, route work to it by language and field, and meet fixed delivery times is an asset in its own right, and no less important than the algorithm. In May 2022 the company itself employed about 570 people, about 200 of them in Israel. Its main target markets were higher education, the legal system and media. These are exactly the places where a transcription error is a legal or regulatory problem, so the customer is willing to pay a higher price for certainty. There is also an employment model here worth noticing: thousands of people around the world working as freelancers on drafts that a machine produced. This is a relatively new kind of work, created because of artificial intelligence and not in spite of it. And it is exactly the kind of work that technological efficiency later shrinks.

Funding and acquisitions: the road to a $2 billion valuation

Verbit raised money quickly: about $11 million in its first stage, $23 million in January 2019, $31 million in January 2020, $60 million in November of that year, and $157 million in June 2021, the round that made it a unicorn. In November 2021 came its largest round, $250 million, at a valuation of about $2 billion. At the same time it expanded through acquisitions: in May 2021 it bought VITAC, one of the largest caption providers in North America, for about $50 million, followed by Automatic Sync Technologies the same year and the British company Take 1 in 2022. The logic was reasonable: bring in long-standing customers in media and law, and run their work volume on cheaper technology. Everything rested on the assumption that the gap between what the model could do and what the customer required would stay open long enough. In hindsight, this pace of acquisitions also increased the company's exposure. The more business it bought that was priced by transcription volume, the more it depended on the price per minute of audio, which was exactly the price that was about to fall.

What happened when the models improved

The gap did not stay open. The leap in the quality of automatic speech recognition reduced the amount of human correction needed, and transcription prices in the market fell as a result, especially in the entertainment industry, which had been one of the main sources of revenue. When the price of a service falls, both sides of the equation are hit at once: revenue per minute of audio shrinks, and at the same time it becomes harder to justify the cost of the human transcriber sitting in the middle. This is one of the most striking cases in which an AI company was hurt by the very progress of the technology it was built on. It was not hurt by a competitor that caught up with it, but by the fact that the problem it solved became easier. Companies that found themselves in this situation and survived usually did so by moving their value elsewhere: workflows, regulatory compliance, deep integration into the customer's systems, or a service that is not priced per minute of audio. The distinction is worth sharpening: Verbit did not lose to a better competitor. It lost because the capability it sold became available and cheap. This is a risk that is very hard to hedge against, because it does not depend at all on the company's own decisions.

The numbers of the decline

Big Tech 50, a publicly traded partnership that holds shares in private companies and reports their estimated value, lowered its estimate for Verbit from $2 billion at the 2021 peak to $512 million in a report published in 2024. In February 2024 founder Tom Livne stepped down as CEO, and Yair Amsterdam replaced him. Rounds of layoffs continued for years: about ten percent of the workforce in July 2022, thirty employees in July 2023, dozens of jobs in March 2024, twenty of them in Israel, and another twenty-two employees in September 2025. At that point the workforce stood at about five hundred people worldwide, including contractors. The reason the company gave for the latest round is worth quoting exactly, because it sums up the whole story: "the integration of artificial intelligence into workflows, as well as organizational adjustments." In other words, the company itself explains that it needs fewer people because the technology has improved. It is important to remember that valuations of private companies are not a market price but an estimate by whoever publishes them, so they should be read as an indication of direction and order of magnitude, not as an exact number.

What did not disappear

It is important not to conclude from this that the need itself has disappeared, because it has not. Making recorded content accessible to people with hearing loss is a legal requirement in many places around the world, and the legal system requires accurate transcription for the court record. The question that remains open is not whether the service is needed, but who provides it and at what price, and how much of the human work is really necessary when the automatic draft improves from year to year. The lesson of this case goes far beyond transcription. It applies to any company that built a product around a gap in what models could do at a given moment: that gap is an asset that wears away on its own, at a pace the company does not control. Verbit continues to operate as of 2026, but the industry discussion about it is less about growth and more about how a business model built on a temporary weakness of the technology adapts to a world in which that weakness has closed. For readers outside the industry there is a very useful conclusion here: when a product is presented as powered by artificial intelligence, it is worth asking which part of it is really automatic and which part is done by people behind the scenes. The answer is changing quickly, and the price changes with it.