Twin Labs is trying to make workplace automation look less like a workflow builder and more like a person using a browser. The Paris-based startup is developing a product for repetitive business tasks that lets an AI assistant interact with software interfaces directly.
The goal is practical: reduce the time people spend repeating the same steps across internal tools. That could include onboarding employees, reordering items, downloading financial reports from several SaaS products or reaching out to potential prospects.
Why Twin Labs is focused on the screen
Many automation tools depend on APIs and carefully designed multi-step processes. Twin Labs is taking a different route. Its system is described as closer to a web browser: it can load pages, click buttons and enter text.
That matters because many workplace tasks already happen inside software interfaces. A human does not always think in terms of APIs. They open a service, find the right field, check the right box and move to the next system.
Twin Labs wants its product to follow that same pattern. Instead of asking teams to translate a process into a rigid automation flow, the company is working toward an assistant that can learn from screen recordings and natural language descriptions.
The example in the source is employee onboarding. A company may need to add a new hire to payroll, send a Slack invitation, create a Google Workspace account and invite the person to set up an account with the healthcare insurance provider. None of those steps is necessarily difficult, but the sequence has to be completed correctly.
Multimodal models changed the approach
Before focusing on multimodal models, Twin Labs tried building autonomous agents with traditional LLMs. According to co-founder and CEO Hugo Mercier, that path did not work reliably enough.
Overall, the conclusion is that LLMs are completely unreliable. This means that LLMs are making the wrong decisions
The startup then shifted to models with vision capabilities, including GPT-4 with Vision (GPT-4V). The reason is simple: a model that can interpret a software interface may be better suited to browser-based work than one that only handles text.
Mercier said GPT-4V had been trained on many software interfaces and underlying code bases, which opened new possibilities for automation. In his words, when the model sees an interface, it can understand what a button is meant to do.
That does not make the product finished. The source makes clear that Twin Labs is still building toward its broader vision. But the technical bet is clear: if an AI agent can see and understand the interface, it may be able to perform routine steps without relying on formal integrations for every service.
The first product will be more controlled
Twin Labs does not plan to begin with a fully open task-creation system. The startup first expects to ship a product with a library of pre-trained tasks. That controlled approach is meant to help ensure the tasks work properly before customers build their own automations.
After that, the company expects to open the platform so clients can create their own tasks. That would move the product closer to the larger idea: an AI assistant trained on how teams already perform repetitive work.
The distinction is important. A pre-trained library can limit the range of what the system has to handle. Customer-created tasks would be broader, because different companies may use different tools, sequences and internal procedures.
The source also points to cost as a challenge. Completing a task currently costs quite a bit of money, though API and infrastructure costs are described as rapidly going down in the AI space. For a product focused on repetitive work, the economics matter because the value depends on whether automation is cheaper and more convenient than manual execution.
Funding and the road ahead
Twin Labs was co-founded by Hugo Mercier and Joao Justi. The two co-founders spent the last six months building a prototype.
The company also raised $3 million in pre-seed funding. Backers include Betaworks, Motier Ventures, Factorial (Matthew Hartman & Clem Delangue) and many angel investors, including Florian Douetteau (Dataiku), Thomas Wolf (Hugging Face), Charles Gorintin (Alan), Mehdi Ghissassi (DeepMind), Romain Huet (OpenAI), Irwan Bello (OpenAI), Romuald Elie (DeepMind), Yan-David Erlich (Weights & Biases), Olivier Pomel (Datadog), Rodolphe Saadé (CMA CGM), Thibaud Elziere (Hexa), Quentin Nickmans (Hexa), Philippe Corrot (Mirakl) and Rand Hindi (Snips, Zama).
The company is entering a space where many people still think of AI products as chatbots. Twin Labs is pointing to another kind of interface: an AI agent that acts inside the tools people already use.
That shift is the central idea behind the startup. The product is not simply about making internal processes faster. It is about whether AI automation can move from answering prompts to carrying out routine software work on a screen.
If Twin Labs can make that reliable, the payoff would be in the small, recurring tasks that fill the workday. Those tasks are often easy to describe, but they require care, order and consistency. The company’s challenge is to prove that an AI agent can deliver that consistency while operating through ordinary web interfaces.