Dust and the Quiet Rise of Custom Enterprise AI Agents

The consumer story around AI in 2026 is mostly about one assistant getting more capable — ChatGPT, Claude, Gemini, all trying to do everything for everyone. Inside companies, the more interesting trend is close to the opposite: building lots of small, narrow agents, each wired tightly into one specific workflow.

What does Dust actually let a company build?

Dust is a platform for assembling custom AI agents that connect to a company's actual tools and data — Slack, Notion, GitHub, Salesforce, internal databases — rather than living as a generic chatbot with no real context. A company might build a support-ticket triage agent, a sales-call summarizer, and an internal-docs Q&A agent, each scoped narrowly to one job.

• Connectors into existing company software so agents work with real, current data instead of what a user happens to paste in.
• Narrow scope per agent — one agent handles one workflow well rather than one agent trying to handle everything adequately.
• Permissions and access controls scoped to what each agent actually needs to see, which matters a lot to security teams.

Why is the narrow-agent approach winning over one big assistant?

A few practical reasons keep showing up when companies explain this choice.

• Grounding: an agent scoped to one workflow with direct data access gives better, more specific answers than a general assistant guessing from vague context.
• Trust and governance: it's far easier to audit and restrict what a narrow agent can see and do than to fully trust one all-powerful assistant with access to everything.
• Faster iteration: teams can build, test, and fix a small agent for one job much faster than trying to get a single general assistant to handle every edge case well.
• Lower blast radius: if a narrow agent makes a mistake, the damage is contained to its one job rather than spilling across every function it touches.

Dust isn't alone here

This same "build your own narrow agents" pattern shows up across a cluster of enterprise platforms — Glean, Lindy, Relevance AI, and the agent-building features big vendors like Microsoft and Salesforce are adding to their own suites. It's less a single product trend and more a shift in how companies are choosing to deploy AI at all.

Does this mean general AI assistants are losing inside companies?

Not losing, exactly — most companies still use ChatGPT Enterprise, Copilot, or Claude for general knowledge work like drafting and summarizing. But for the specific, repeatable workflows that make up the bulk of day-to-day operations, the pattern increasingly is: general assistant for ad hoc thinking, narrow custom agent for the recurring job that needs real data and real accountability.

Bottom line

The quiet trend inside real companies isn't one AI to rule them all — it's dozens of small, well-scoped agents doing specific jobs well. Dust and its peers are betting that's the more durable way enterprise AI actually gets used.