Snowflake Startup Spotlight: Dust

Welcome to Snowflake's Startup Spotlight, where we ask startup founders about the problems they're solving, the products they're building and the lessons they've learned during their startup journey. In this edition, meet Dust's Co-Founder and CTO, Stanislas Polu, and learn how the company is building a collaborative AI platform to reduce the distance between your data and the decisions your teams are making.

I'm inspired by technical shifts that force us to question old assumptions. Large language models have made intelligence abundant, but abundance creates a new problem: How does a company use changing sources of intelligence without giving up control of what makes it unique? As AI spreads across individual tools and vendors, the company's context, agents and ways of working can become fragmented. I want to build the layer that lets organizations absorb constant progress while keeping those assets and choices their own.

Dust is the collaborative AI platform that rewires the way companies work and enables them to control the intelligence they use and own the data they generate.

Companies are gaining access to more intelligence, but they risk losing control over it. Much of workplace AI is still adopted by one person, assistant or model at a time. That creates three forms of dependency: reliance on a single model provider, loss of control over the data and traces generated by agents, and limited visibility into costs and behavior.

The traces matter. They capture how work was completed: the instructions, context, model and tool choices, outputs and human feedback. Over time, that record may become one of the company's most valuable AI assets, helping it evaluate workflows and develop systems tailored to how it actually operates. Companies shouldn't give that asset away by default.

My co-founder, Gabriel Hubert, and I had seen the growing pains of coordination while working at Stripe, as it grew from roughly 150 people into a much larger organization. Information gets lost, decisions become harder to trace and teams stop sharing the same context. When LLMs became capable enough for serious work, our hypothesis was that companies didn't need another model. They needed a shared layer between humans, company data, existing systems and the available models.

Dust is that layer. It lets teams work collaboratively with AI, switch among supported frontier and open source models, and govern the context, agents and traces their work generates. What one person builds can become a reusable company capability without becoming dependent on a single source of intelligence.

We're helping companies turn agentic traces into an asset they can understand, improve and retain.

The data required to do a job rarely lives in one place. A single question can require warehouse data from Snowflake, account context from a CRM and qualitative information from Slack, Notion or call notes. But the result is only part of the value. The trace around it shows which context, instructions, models, tools and human decisions produced the work.

Dust brings those elements into a shared company system. An AI operator can study what works to improve an agent or skill and make it reusable by colleagues. As the company changes models or providers, it keeps the accumulated context and learning around its workflows. In the future, that corpus could also help companies evaluate or train models for their specific work. The model may change, but the intelligence the organization builds around it should remain its own.

Snowflake is a connected source of structured business data that Dust agents can work with. Organizations can connect Snowflake to Dust so teams can query relevant warehouse data in natural language and combine it with context from other company systems.

That matters because useful business questions often cross system boundaries. Snowflake can remain the structured data foundation and source of truth, while Dust helps humans and agents use that data as part of broader work.

This is an important part of company-owned AI. A company should be able to keep Snowflake as its source of truth, apply the relevant permissions and control how that context is used by its agents, rather than rebuilding its data strategy around an AI vendor.

Snowflake makes structured data usable beyond the data team. Paired with Dust, a business user can ask a question in natural language, use relevant warehouse data alongside other company context and turn the result into a repeatable workflow rather than a one-off analysis.

This is the direction we care about: reducing the distance between a company's data and the decisions or actions it should inform, without weakening the company's control over that data. Access and control are not opposites. Data teams can encode trusted access and business logic, while more colleagues gain a practical way to use them.

The mistake is to treat AI as a race with one permanent winner. Models will keep changing in capability, cost, latency and specialization. Companies should not have to predict the winner or rebuild their operating layer whenever the frontier moves.

Dust is multimodel by design. Organizations can choose among supported frontier and open source models and update those choices as the landscape changes, without rebuilding their company context and workflows. This is what sovereignty means to us in practice: not doing everything locally or rejecting external providers, but staying in control when a provider, price, policy or model changes.

Companies will rent intelligence from many sources. The durable asset is what they build around it: how the organization works, what it knows, what it has decided and where human judgment matters. That layer should belong to the company, not to whichever model happens to be leading today.

The next shift is from individuals prompting assistants to companies coordinating digital teams of humans and agents. Our principle today is simple: Agents propose; humans dispose.

Before presenting a research paper early in my career, I worried that I didn't have enough authority to defend the work. My adviser reminded me that I had spent more time on that specific problem than almost anyone and should trust the depth of the work.

That advice has stayed with me as a founder. Deep focus can create its own doubt, because if a solution becomes obvious to you, you assume it must be obvious to everyone. However, often it's not the case. You should remain open to being wrong, but you should also respect the insight that comes from sustained attention to a real problem, then keep testing it with users.

Don't fall in love with a technology before understanding whether it can support a large and important market.

At TOTEMS, the company Gabriel and I founded before Dust, we were fascinated by what the technology could do. We did not ask early enough how large the business could become if everything worked. With Dust, we have tried to make a few consequential bets, test them quickly and avoid optimizing something that is still insignificant in the market.

That lesson shaped one of our early principles: "No product-market fit, no GPUs." We could have focused on training models, but it would've been a seductive distraction. We chose the product layer where customers had a concrete problem and where their usage could teach us what to build next.

AI will become less like a collection of applications that individuals open and more like a workforce that companies coordinate. Access to intelligence will not be scarce. The scarce resource will be the organizational ability to direct it, without becoming dependent on one provider's models, interface or roadmap.

I see Dust as the shared system where companies build and govern their own AI: their context, agents, skills, model choices and the ways humans and agents work together. Successful ways of working should become reusable across the organization and remain valuable when the underlying models change.

The goal is not for every company to own a foundation model. It's for every company to own its AI capital, including the soft and hard skills, the learnings, know-how, workflows, traces and operations it builds with AI. Dust is building for that future.

Learn more about Dust at https://dust.tt/. If you're a startup building on Snowflake, check out the Snowflake for Startups program for info on how Snowflake can support your goals.