Dust is the AI agent platform Tandem has been deploying for internal document search since autumn 2025. My review fits in three sentences. It does one hard thing very well, plugging artificial intelligence into a company's real documentation without inventing answers. It costs less than most people assume. And its self-service tier carries a limit nobody mentions, a maximum of three connectors.
This article covers what it does, what it costs in October 2026, what I took away from client work, and the three points that decide whether it suits your company.
Dust, what the platform actually does
Dust is a software platform that connects a company's documentation, tools and teams so AI agents can run on top of them. The vendor sums up its promise in one line on its homepage, collaboration between people and agents in a shared workspace. Every agent gets instructions, a data perimeter and tools. It answers by citing its sources.
The company was founded in 2022 by Gabriel Hubert and Stanislas Polu, both Stripe alumni, the latter having also spent three years at OpenAI. It raised 40 million dollars in May 2026, around 34 million euros, in a round co-led by Sequoia and Abstract with Snowflake and Datadog taking part. Sifted covered the round, which brings total funding past 60 million dollars. The vendor claims more than 3,000 customer organisations and over 300,000 agents deployed.
The house vocabulary deserves a note. Dust talks about AI Operators, the people closest to the work who build and run their own team's agents. The vendor states plainly that transformation will come neither from outside consultants nor from centralised innovation units. For an agency like Tandem, that claim lands head on. It is also largely right, and I come back to it below.
Why Tandem deploys Dust for internal search
Document search is the use case that unlocks the most value for the least effort. An employee spends a sizeable share of the week looking for information that already exists somewhere. I covered this use case and its impact in our report on AI use cases in business. The platform we deploy for it is Dust.
- Document spaces connect to Google Drive, Notion, Confluence and SharePoint.
- The CRM connects too, whether that is Salesforce, HubSpot or Attio.
- Permissions are set per space and per team, which avoids exposing HR files to the whole company.
- Every answer cites the source document and the date it was last updated.
The result we observe on our deployments fits in one figure, search time divided by ten on average. A few minutes become a few dozen seconds. Onboarding speeds up for new joiners, and repeat questions between teams drop. The payback maths is simple. An employee who saves one hour a month has already paid for the licence.
What Dust costs in October 2026
The grid is published in euros excluding tax on the vendor's pricing page, checked in October 2026. The Business tier covers teams of up to one hundred people. Beyond that, an Enterprise contract takes over.
| Tier | Price per seat per month | Credits included | Who it fits |
|---|---|---|---|
| Free | 0 euro | 500 credits for life | Trying it out alone |
| Pro | 24 euros yearly, 30 euros monthly | 8,000 per month | A team getting started |
| Max | 120 euros yearly, 150 euros monthly | 40,000 per month | Heavy daily usage |
| Enterprise | On quote | Pooled across the workspace | Beyond one hundred people |
One useful benchmark. In October 2025 I quoted around 29 euros per user per month in my newsletter for this kind of solution. The list price has come down since, and the structure has changed. Re-checking a product price before quoting it stays the basic reflex, even when the figure came from you.
Credits, the point that worries users
Dust now bills two things, the seat and the consumption. A credit is charged on every message, depending on the model used and on the tools the agent calls. A simple question put to an efficient model costs little. A deep research task chaining several steps and several tools costs far more. Credits reset at the start of each period and never roll over.
That shift is the most frequent complaint in published reviews. On G2 the platform scores 4.6 out of 5 across 87 reviews, checked in October 2026, and several users say they fear the move to usage-based billing. The recurring praise goes to how fast an agent can be built and to getting several models without having to pick one. The same reviews report one month to implement and a four-month payback.
A different model per task, the argument that holds
The tier gives access to more than twenty frontier models in the same interface, from GPT to Claude through Gemini, Mistral and DeepSeek. The agent that drafts and the one that reviews a contract do not need the same engine. I was already making that case in my LinkedIn post on the pyramid of AI tools, where each tool answers one precise need rather than all of them.
The practical benefit goes beyond comfort. Switching an agent to another model takes seconds, with no need to rewrite instructions or reconnect data. When a provider raises prices or lets latency slip, you move within the day. That reversibility is worth a lot, and comparison articles rarely highlight it.
Frames, the feature that surprised me most
An agent usually replies with a wall of text. Frames turn that reply into an interactive interface the team can share and edit. On a supplier quote comparison, the difference is striking. The agent normalises units, aligns line items and flags significant gaps. The output lands as a workable table rather than prose.
The figure I keep from that demonstration is two minutes fifty-two, against half a day by hand. I documented it in my newsletter in November 2025. It is also the feature users most often name as their favourite in published reviews, which is unusual for a display feature.
Is Dust really a sovereign platform
That is the argument everyone leads with, and it deserves a close look. The vendor publishes its subprocessor register, which is already a rare mark of transparency. Twenty subprocessors are listed, seven of them model providers. Only one of those, Mistral AI, processes data solely inside the European Union.
The good news sits elsewhere. European data residency is already included in the self-service tier. Many platforms reserve it for their Enterprise contract. The vendor is SOC 2 Type II certified, states it never trains models on your data and declares a zero-retention policy with its providers. For an organisation that wants to go further, the only complete answer remains running models locally, with the cost and constraints that implies.
How many connectors, really
Here is the most concrete point in this article, and the least documented elsewhere. Three pages from the same vendor, read on the same day in October 2026, give three different counts. The homepage advertises more than seventy connectors. The pricing page mentions more than twenty data sources. The technical documentation lists nine native connections.
The nine native connections cover Confluence, Google Drive, GitHub, Notion, Slack, Snowflake, Zendesk, Intercom and SharePoint. Everything else goes through an interconnection standard called MCP, which plugs in almost any tool that exposes a compatible server. The Business tier allows five remote ones, the Enterprise tier opens them up without a cap.
What to remember is the line in the vendor's own comparison table. On the self-service tier, active connectors are capped at three. A company wanting to plug in its drive, its documentation space, its CRM and its messaging already exceeds that on day one. That single point pushes many projects towards an Enterprise quote, and no French comparison article read in October 2026 mentions it.
What Dust's co-founder told me in 2024
I hosted Gabriel Hubert, co-founder and chief executive of Dust, in an interview filmed in September 2024. Two years on, what he described then looks a lot like the AI Operators doctrine the company puts forward today.
- Gains are measured task by task. He quoted an order of magnitude of 5 to 10 % across a full working week, with huge variation between tasks.
- The cost of experimenting has collapsed. A large IT programme used to demand years and an integrator. Testing an assistant takes two hours and a short training session.
- Business experts build it themselves. The people who understand a process best are best placed to decide the shape of the tool, and they now have the means to make it.
On that last point, my field experience adds a caveat. Business teams do build, and they build fast. But they build well when someone has first set the frame, picked the use cases worth the effort and trained the first internal champions. Without that step, agents multiply without ever reaching production. That is exactly the mechanism I describe in my analysis of companies that see no return on their AI investment.
The governance that goes with that frame is detailed in our article on AI governance in business, and the full diagnosis of missing returns in our report on the 80 % of companies with no ROI.
Buying Dust or building your own agents
That is the first question every executive asks. I devoted a LinkedIn post to the build versus buy trade-off. That still holds. In most cases off-the-shelf solutions win. Specialist vendors iterate far faster than an internal team, and the technical debt of a bespoke build is expensive.
The useful boundary runs like this. Dust answers very well when the need is conversational, document-driven and shared across teams. An orchestrator such as n8n answers better when the need is an event-triggered flow writing into several systems with nobody at the screen. The two coexist happily, and the distinction between assistant and automation is detailed in our comparison of workflows, assistants and agents.
Should you choose Dust for your company's AI agents
Yes, in three situations. If your first need is finding information scattered across your tools, Dust is the shortest path to a visible result. If you want to avoid locking yourself to a single model provider, the platform lets you switch in seconds. And if you care about data staying in Europe, European residency is available from the self-service tier.
No, in two others. If your need is an automated flow running with nobody at the screen, an orchestrator will do better and cost less. And if you expect a 24-euro subscription to settle your company's AI question, you will be disappointed. The licence is the smallest line in the budget. Scoping, tidying the documentation and training the teams make up all the rest.
Tandem supports more than 50 companies on this kind of deployment, from choosing the platform to training the teams. If you want to scope yours, our Copilot offer starts from use cases before it talks about tools. And to understand why document search works so well, our article on context and retrieval-augmented generation explains the underlying mechanism.
