The Quiet Defection From Microsoft’s AI Ecosystem
Enterprise AI was supposed to be Microsoft’s game. With Copilot baked into Office 365 and a distribution network covering hundreds of millions of corporate seats, the company had every structural advantage. But a growing number of mid-size and enterprise companies are quietly migrating their AI workflows to Dust, a Paris-based startup that has built its platform around one idea Microsoft has struggled to execute: making AI genuinely useful for how specific companies actually work, not how a generic productivity suite assumes they work.
Dust is not a chatbot wrapper.
The platform lets companies build custom AI agents connected directly to their internal knowledge bases – Notion docs, Slack threads, Salesforce records, GitHub repositories, internal wikis – and deploy those agents across teams. The result is an AI layer that answers questions grounded in a company’s actual data, not Microsoft’s training corpus. That distinction is where Copilot keeps losing the argument.
Where Microsoft Copilot Falls Short for Enterprise Teams
Copilot’s core problem is not capability – it is context. Microsoft’s tool is trained to work across generic Office workflows, which means it performs reasonably well on tasks like drafting emails or summarizing meetings. But when an enterprise team needs an AI agent that understands their internal escalation process, knows their product documentation by version, or can answer a support engineer’s question using three different internal tools at once, Copilot hits a wall. The integration is surface-level, not structural.
Dust is built to close that gap. The platform’s architecture treats each company’s knowledge infrastructure as the primary training layer. When a company connects Dust to their internal tools, the agents that get built on top can actually traverse that data with specificity. A sales team’s agent knows the difference between a churned account and an at-risk one because it is reading from the same CRM notes the sales team writes. A customer support agent can pull from product changelogs in real time because the connection to internal documentation is live, not a one-time sync.
This is also where Dust’s pricing strategy starts making sense to finance teams. Microsoft Copilot is licensed per user, at a rate that stacks on top of existing Microsoft 365 costs. Dust operates on a team-level model with flexible agent deployment, which means companies are paying for actual utility rather than seats on a platform most employees will open twice and forget. For operations leads evaluating AI spend, that math is increasingly hard to ignore.
The Competitive Pressure Building Around Copilot
Microsoft is not standing still. The company has invested heavily in expanding Copilot’s connector ecosystem and has been rolling out more sophisticated retrieval features tied to SharePoint and Microsoft Graph. But the speed of that development has not kept pace with what enterprise teams are demanding right now, and startups like Dust are filling that gap while Microsoft is still building the scaffolding.
Dust co-founder Gabriel Hubert has been direct in positioning the product against Microsoft’s approach, arguing publicly that enterprise AI should not be a feature inside an office suite but a configurable layer that sits across all of a company’s tools. That framing has resonated with engineering-led companies and operations-heavy organizations that see their internal knowledge as a competitive asset – not something to be summarized by a general-purpose assistant. The companies most likely to move to Dust are the same ones that already have sophisticated internal tooling and are frustrated that Copilot cannot connect to any of it meaningfully.
The startup space around enterprise AI connectors is getting crowded – tools like agentic coding platforms are also pushing deeper into workflow automation – but Dust’s focus on non-developer teams gives it a specific angle that pure developer tools cannot replicate. A head of support or a revenue operations manager is not going to configure a coding agent to answer internal questions. Dust is built so they do not have to.
What Dust Gets Right That Larger Players Keep Getting Wrong
The pattern Dust is exploiting is familiar from other enterprise software categories: a large platform builds a feature that is good enough for 80 percent of users, and a focused startup builds a product that is genuinely excellent for the 20 percent who need more. The difference here is that the 20 percent Dust is targeting – operations, support, sales, and knowledge-intensive teams – are the exact teams companies spend the most money trying to make more efficient. That is not a niche. That is the core of enterprise productivity spend.
Dust’s agent builder requires no code to configure, which means deployment does not sit in a ticketing queue waiting for an engineering team. Managers can build, test, and update agents without technical support, which means the platform can adapt when processes change – and in most companies, processes change constantly. That operational flexibility is something Microsoft Copilot, tied as it is to Microsoft’s own deployment and update cycles, structurally cannot match at the same speed.
The deeper issue for Microsoft is that Copilot’s value proposition assumes companies want AI inside Microsoft’s ecosystem. Dust assumes the opposite – that companies want AI that works where their data already lives, regardless of which vendor it belongs to. As enterprise data continues to fragment across Slack, Notion, Linear, Salesforce, and dozens of other tools that have nothing to do with Microsoft, that assumption is looking less like a startup pitch and more like an accurate description of how work actually happens.
