Dust.tt

Dust.tt is a French enterprise AI company co-founded by Gabriel Hubert and Stanislas Polu. Its platform connects language models to company knowledge and tools, allowing teams to build and share custom assistants. By 2026, Dust described its direction as “multiplayer AI”: people and agents working in persistent shared workspaces, with common context and visibility into ongoing work.[2][3][4]

That shared-workspace product followed earlier experiments in developer tooling and browser assistance. The founders’ careers before Dust provide context for its combination of model engineering and product development.[2][1][4]

Hubert and Polu previously founded Totems, which Stripe acquired in 2015. Both worked at Stripe before taking different paths: Polu spent three years at OpenAI working on language models’ reasoning capabilities, while Hubert became head of product at Alan. They subsequently reunited to create Dust.[2]

Polu recalled that seeing GPT-4 internally in September 2022 helped motivate Dust’s focus on making model capabilities useful through products. Dust initially built tools for developers designing and deploying LLM applications. Its interface let them inspect each model interaction across multiple examples of a multistep workflow. Polu’s rationale was practical: developing against one example risked overfitting, while running several examples produced an interleaved console stream that was difficult to understand.[1][2]

Dust also experimented with XP1, a browser extension available by early 2023. Polu described it as deliberately useful within the capabilities of the smaller model it used: extracting information from a LinkedIn page or summarizing a newspaper article, rather than depending on access to GPT-4.[1]

By June 2023, Dust had focused on centralizing and indexing internal company data. Connectors continuously fetched information from Notion, Slack, GitHub and Google Drive; semantic search retrieved relevant material, which Dust supplied as context to a language model. The result was an answer assembled from company knowledge, rather than a list of search results. Dust was working with design partners on how to package that platform.[2]

By June 2024, companies could create several task-specific assistants and use them through Dust’s web interface or directly in Slack. Examples included answering support questions from a knowledge base and past interactions, drafting job descriptions from earlier examples, and writing SQL queries using company database schemas. Assistants could use models from OpenAI, Anthropic, Mistral or Google.[3]

In November 2024, Polu emphasized simple workflows described in English instructions, with tools for semantic search, structured queries and web search. Dust used an assistant to turn a week of Slack incident and shipped-work messages into a meeting table, and other assistants to generate financial graphs. Combining those components into a single weekly-meeting agent remained an objective; Polu explicitly said Dust had not yet explored meta-agents.[1]

Dust chose a horizontal product that people across a company could adapt to their own operational work. Polu acknowledged that this made selling the product harder than selling a clearly defined vertical solution. He credited designer Ed with making the interface and language approachable, including using “instructions” rather than “system prompt.” Dust remained open source, though Polu said open source was not its go-to-market strategy.[1]

Dust maintained most integrations itself and used APIs for most internal systems. Polu argued that owning connections end to end helped preserve information structure—for example, distinguishing a Notion database containing text from one containing quantitative data. Dust bought Temporal infrastructure to coordinate updates from Slack, Notion and GitHub and trigger workflows, while retaining its own unified interface to model providers.[1]

Polu distinguished Dust from a model research lab in the November 2024 interview: “We are 100% deliberately a product company. We don't do research. We don't train models. We don't even run GPUs.” His earlier reasoning research belonged to OpenAI; Dust’s stated focus at that point was extending what existing models could accomplish through the product.[1][2]

Dust’s 2026 announcement described persistent workspaces organized around teams, initiatives or workflows. Its context layer combined semantic search with MCP connections for queries and actions across tools. Dust also described suggestions for improving agent instructions and skills based on conversation patterns, alongside granular permissions, cost and usage monitoring, audit trails and agent analytics.[4]

Hubert described using Dust to coordinate the Series B announcement with Polu and the marketing, data and operations teams. Agents synthesized information from Slack, Snowflake, HubSpot, Notion, Gmail and Google Drive, helped prepare launch materials and flagged blockers. Humans refined the messaging, coordinated approvals and made final decisions while the workstreams advanced in parallel.[4]

Dust raised a $5.5 million (€5 million) seed round led by Sequoia, reported in June 2023, followed by a $16 million Series A led by Sequoia Capital, reported in June 2024. On May 18, 2026, Dust announced a $40 million Series B with Abstract, Sequoia, Snowflake Ventures and Datadog.[2][3][4]

In June 2024, TechCrunch reported $1 million in annual recurring revenue. Qonto estimated monthly usage by 75% of its 1,600-person team; the report gave 80% weekly usage at Alan and 86 custom assistants created at Pennylane. In the November interview, Polu reported a 25-person Dust team, a highest customer deployment of 88% daily active employees, and approximately 60–70% weekly activation among its then-current enterprise customers.[3][1]

Dust’s 2026 Series B announcement reported use by more than 3,000 organizations globally and more than 300,000 deployed agents. Those totals were not defined as paying customers or active agents. The company also reported more than 300 agents across 11 departments at Persona and a company-wide AI strategy at Doctolib reaching 3,000 employees; the latter described strategy reach rather than measured active usage.[4]