D&B Rebuilt a 642M Business Database Because AI Agents Couldn't Use It

Built for patience, not for agents

Dun & Bradstreet has maintained business records since before the telephone existed. Its Commercial Graph spans 642 million companies, their subsidiaries, credit risk profiles, and corporate hierarchies. For nearly two centuries, the system served humans well. Credit analysts waited patiently for queries to return. Risk managers worked through fuzzy entity matches manually. Sales professionals read reports and made judgment calls. The database assumed patience, literacy, and ambiguity tolerance.

The agent shift broke the old graph

That assumption collapsed when D&B's customers started wiring AI agents directly into procurement, credit, and supply chain workflows. An agent does not wait five minutes for a batch job. It does not squint at two similar company names and guess whether they share a parent entity. It cannot tolerate a hierarchy that resolves differently depending on who asks. When the consumer is code running a loop, every fuzzy edge in the graph becomes a potential failure point.

So D&B rebuilt it. The company took a data infrastructure that had served nearly 200,000 customers globally and redesigned it for machine consumers. D&B made corporate relationships explicit. It rendered risk signals machine-readable. The new architecture favors real-time decision support over slow report generation. In short, D&B replaced a system built for human interpretation with one built for automated execution.

What this means for builders

If you are shipping an app with AI features this year, you will face the same friction D&B just spent 180 years accumulating. Most backends grew up around human request-response cycles. A user clicks a button, waits a moment, and reads the result. Agents do not read. They ingest, decide, and act, often thousands of times per hour. They compound errors when a relationship is unclear. They burn tokens re-querying data that should have streamed to them automatically.

The model gets the glory, but the backend sets the ceiling. You can spend a week tuning prompts and swapping between Claude, GPT, and Qwen, but if your database treats the agent like a human with a dashboard, the result will stall. Agents need backends that push updates in real time, preserve state across long runs, and expose structured relationships that code can traverse without guessing. That is the exact gap Botflow addresses by running on Convex, a backend purpose-built for AI agents.

D&B's rewrite is a signal. Other data giants will follow, because legacy backends cannot serve the new consumer. For indie hackers and small teams, this is actually good news. You do not carry 180 years of technical debt. You can choose an AI-native backend today and skip the rewrite. When you pick your stack, ask one question. Was this database built for a person refreshing a page, or for an agent that never sleeps?