H Company has established itself in a key segment of enterprise AI. Portrait of this French start-up which is advancing beyond the radar.
It is one of the most promising French AI laboratories. Founded in 2023, H Company specializes in AI for computer use: models capable of controlling software interfaces as a human user would. Less publicized than Mistral AI, the start-up has developed a family of models, Holo3, which outperforms the American behemoths (OpenAI, Anthropic) and Chinese (Alibaba, Moonshot…) in computer use. A discreet French success story.
Agents designed for business automation
Where the wave of consumer personal agents, illustrated by the OpenClaw phenomenon, has demonstrated the market’s appetite for AI capable of acting autonomously, H Company is positioning itself on the next level: the corporate personal agent. The start-up’s approach is based on computer use, the ability of an AI agent to perceive any graphical interface and interact with it. A rare competitive advantage, H Company masters the agent stack from end to end: the start-up develops both the engine model (VLM) and the agent that orchestrates it (scaffold).
Concretely, H Company’s models and agents target business processes with high IT friction. The typical use case? Workflows of 30 to 40 steps involving multiple enterprise software. “Of these 30 steps, 28 could go through an API. The remaining two are impossible to automate, because the system is too old or because it is an external tool beyond your control. This complexity would waste months with traditional approaches. With our agent, we record the journey and that’s it”, summarizes Gautier Cloix, former CEO of Palantir appointed CEO of H Company in June 2025, after the departure of Charles Kantor (ex-CEO and co-founder).
The most telling use case comes from the medical sector. The start-up devotes 10% of its deployments to pro bono missions, including Saint John Hospital in Bangalore, India. “The nurses devote 60% of their time to computer entry. Our agent takes care of this entry: they only intervene to validate. The rest of the time is given back to the patient,” explains the CEO. The result is a waiting time in the emergency room halved. On the customer side, the target profile remains the major account. “Three quarters of our base are large legacy companies, CAC 40 in France,” explains Gautier Cloix. The rest is divided between ETI and scale-up.
Cutting-edge models, rivaling OpenAI and Anthropic
And the results follow. In early April, H Company launched Holo3, the third generation of its family of models. With a score of 78.85% on OSWorld-Verified, the reference benchmark for computer use agents, Holo3 rose to the top of the world ranking, ahead of significantly heavier models, such as Claude Opus 4.6 from Anthropic or GPT-5.4 from OpenAI. “With our 3B model, we cover 90% of the tasks. The agent calls the heavier versions of Holo3 only when it is necessary. And 3B is 100 times cheaper for inference,” explains Gautier Cloix.
How does a start-up of 75 people manage to compete with the most funded laboratories on the planet? The CEO first points out a structural advantage: specialization. “The computer use teams at the American giants are roughly the same size as ours. There are not 300 people dedicated to that at Google,” he explains. Where the giants drive increasingly massive general models, H Company focuses all its R&D on a single objective. The other differentiator: the desktop. “We have invested massively in the desktop, because that’s still where the world works. Critical banking and insurance applications run on the desktop,” explains Gautier Cloix. A segment neglected by the competition, which focuses on the web, easier to simulate in training: “My contacts at Anthropic and at Google themselves tell me that they have not built up this arsenal of desktop environments.”
Tomorrow, robotics?
In a more distant vision, H Company’s efforts on computer use could serve the start-up in use cases much more anchored in the physical world… And why not in robotics. Vision models (VLM) trained to understand graphical interfaces share an architecture very close to those used in robotics: perceive a visual environment, interpret a situation, decide on an action. The pipeline that allows an agent to navigate a SAP is not fundamentally different from that which would guide a robot in a warehouse, nor is the training stack. In other words, the champions of computer use have all the weapons in hand to succeed in VLM robotics models. And H Company is no exception.
Asked about this possible extension, Gautier Cloix remains cautious without excluding the option. “It’s an active reflection internally. We already collaborate occasionally with Boston Dynamics and interact with several French players in the sector,” confides the CEO, who however prefers to speak of complementarity rather than pivot. And added: “Our priority remains unchanged: generating tangible results for large organizations. If robotics becomes a key lever to achieve this, we will be able to integrate it.”
