14 of the 30 Most Valuable Vertical AI Companies Now Train Their Own Models

Nearly half of the top-valued vertical AI companies have told the public they build models of their own, and healthcare is furthest along, with 7 of its 9 companies doing so, CB Insights research found.

Most start from an existing model and train it further on a single job, such as spotting drug names in a doctor’s dictation. The aim, CB Insights said, is fewer errors or lower costs on the work customers repeat most. For a buyer renewing a vertical AI contract, that makes the vendor’s own model a second option beside an outside one.

What did the count measure?

CB Insights ranked the 30 companies by their latest priced-round valuation and scored each on five signs of research capability, from model training and published tests to hiring for research. Seven show at least four, four of them in healthcare; Harvey, Hippocratic AI and Abridge show all five.

The scoring counts only public activity, so it shows how heavily a company invests in research, CB Insights said, not how good its models are.

How far has Harvey gone?

Harvey published its own Legal Agent Bench in May, then in August released Tenet, a research preview built by training Moonshot AI’s Kimi K3 further on those kinds of tasks. Harvey says Tenet completes almost twice as many of the benchmark’s reserved test tasks as the base Kimi K3 and places second on the benchmark overall. CB Insights adds that running Tenet costs less than a quarter of what leading models cost. All three results rest on Harvey’s own benchmark.

Cost matters because AI providers bill by the token. According to investor Lightspeed, monthly usage on Harvey’s platform rose from 1 trillion tokens in January to 14.6 trillion in June, before Tenet existed.

Why do legal and finance companies test more than they train?

Eight of the 11 legal and finance companies in the group publish their own benchmarks, yet only two of them, Harvey and Feedzai, have described training a model in the past year, CB Insights found. A good test, CB Insights argues, is what lets a company pick among outside models and show whether training its own would pay off. Rogo, which sells an AI finance analyst, drew the 928 questions in its benchmark from finance practitioners.

What do the labs and platforms do next?

The specialists still depend on companies that compete with them. OpenAI launched legal and finance products in September, and its legal launch counted Harvey and Legora, the second most valuable legal company in the group, as API customers. In February, Epic, the largest US health records provider, made its own documentation tool broadly available, in competition with scribes such as Abridge and Ambience.

CB Insights expects more legal and finance companies to start training.

A buyer can ask a vertical vendor which model does the work, who trained it, and whether anyone outside the vendor has tested it.