Dun & Bradstreet has spent over 180 years constructing a complete business database. Its Business Graph, overlaying 642 million companies and their relationships, company hierarchies and threat profiles, was designed for folks. Credit score analysts, threat managers and gross sales professionals who may watch for question outcomes and work by way of ambiguous entity matches. AI brokers can’t do any of these issues.
When D&B's clients began pushing brokers into credit score, procurement and provide chain workflows, the Business Graph that had reliably served almost 200,000 clients globally turned an issue. The methods constructed to serve human analysts have been the improper structure for machines. So D&B rebuilt.
"We want to consider brokers as our new shopper class, evolving from our customary credit score analysts or gross sales and advertising professionals, et cetera, to additionally now catering to those clients' brokers," Gary Kotovets, Chief Information and Analytics Officer at Dun & Bradstreet, advised VentureBeat.
What broke when brokers began querying
The Business Graph was not a single database. It was a set of separate methods constructed for various use instances and completely different markets, held collectively by customized integrations. Human analysts navigated that fragmentation by way of SQL queries or pre-built interfaces. Brokers couldn’t.
The size of the underlying knowledge compounded the issue. The database had almost doubled in 5 years, increasing from greater than 300 million to greater than 642 million enterprise data, with 11,000 fields per file, in response to D&B. The agency now runs roughly 100 billion knowledge high quality checks per thirty days as data transfer by way of its methods. Querying that on the sub-second latency brokers require, towards a fragmented structure, was not workable.
The relationships the graph tracked have been additionally the improper form. Legacy methods recorded static connections between entities. A CEO was linked to an organization. That was the road. Brokers engaged on credit score assessments or third-party threat want dynamic relationships: when that CEO leaves for a brand new firm, which group does their observe file observe? When a subsidiary modifications possession, how does that propagate throughout a company hierarchy? These questions required customized analyst work earlier than. Brokers can’t watch for customized analyst work.
The broader drawback will not be distinctive to D&B. Kotovets stated he has spoken with a whole lot of CDOs and CIOs over the previous six months and persistently heard the identical constraint: they might not construct what they needed in AI as a result of their knowledge foundations weren’t standardized, normalized or agent-queryable. D&B had that basis, constructed over a long time to serve human analysts. It nonetheless needed to rebuild for brokers.
What they really constructed
The rebuild began with consolidation. D&B migrated its fragmented databases to cloud infrastructure, redesigned the underlying schema and constructed an information material layer that normalizes data throughout markets whereas preserving regional compliance necessities. The result’s a unified data graph that tracks billions of relationships throughout 642 million firms, constantly up to date and enriched by AI-driven knowledge processing.
On high of that graph, D&B constructed a structured entry layer for brokers. Uncooked SQL entry at agent question volumes and latency necessities was not the reply. As a substitute, D&B created a set of instruments and expertise obtainable by way of MCP that package deal knowledge with context and route brokers to the correct data for particular queries. A match and entity decision engine sits behind each question, confirming that when an agent asks about an organization, the reply resolves to a verified, particular entity reasonably than a reputation match.
D&B solved agent identification from each instructions
Rebuilding the graph and including MCP entry solved the info retrieval drawback. It didn’t resolve the identification drawback. Brokers usually are not people, and the authentication mannequin constructed for human customers didn’t prolong to machines.
D&B constructed a brand new registration mannequin for brokers. They have to map to a verified IP deal with and register a person entry key, handled as an authenticated identification in the identical pipeline as a human person.
"We even have an idea of Know Your Agent, much like know your buyer, that does these further verifications," Kotovets stated.
That handles the inbound drawback: figuring out which firm an agent belongs to and what knowledge it’s entitled to question. However D&B additionally constructed for the outbound drawback: what occurs when a buyer's personal multi-agent workflow loses observe of which firm it’s analyzing.
In a workflow that chains a credit score examine agent, a KYC agent and a third-party threat agent, every queries D&B at a special step. With no mechanism to substantiate they’re all referencing the identical entity, a workflow can full whereas working on divergent data.
"They’ve to return again to our verification agent to make sure that they're nonetheless speaking to one another about the identical entity," Kotovets stated. "It's virtually like a digital handshake, in a way."
D&B's enterprise verification agent may be embedded into any workflow as a persistent reference level and is obtainable on Google's A2A protocol no matter which orchestration instrument a buyer makes use of.
4 issues enterprises should get proper earlier than deploying AI brokers
The rebuild uncovered necessities that transcend D&B's personal stack.
Information foundations come earlier than agent infrastructure. The CDOs and CIOs Kotovets spoke with over the previous six months persistently hit the identical wall: they can’t construct what they need in AI till their knowledge is clear, normalized and consolidated. D&B had that basis already. Most enterprises don’t, and they’re going to really feel it.
Design for dynamic relationships, not static ones. Enterprise knowledge methods sometimes file point-in-time connections: an individual belongs to an organization, an asset belongs to a subsidiary. Brokers engaged on credit score, threat or provide chain choices have to purpose throughout relationships that shift over time. If the underlying knowledge solely captures the static line, the agent will too.
Construct entity consistency checks into multi-agent workflows. When a number of brokers contact the identical entity at completely different steps, there is no such thing as a assure they’re all referencing the identical file by the point the workflow completes. That hole must be engineered for explicitly. Entity verification is a workflow design requirement, not an non-obligatory guardrail.
Embed lineage from the beginning, not as an afterthought. Each agent-produced reply ought to carry a traceable path again to its supply. In credit score, threat and provide chain choices, the price of an error is concrete. Lineage must be in-built earlier than scaling, not added after issues floor.
"You might at all times click on and see the place it got here from, and validate all of it the way in which again to the unique supply," Kotovets stated. "That's been the important thing for us in unlocking quite a lot of different capabilities, as a result of now we have that stage of certainty within the issues that we've finished."
