Dun & Bradstreet brings business identity into AI workflows

Dun & Bradstreet supplies business identity and commercial information that an AI assistant cannot reliably invent. Its D&B.AI offering connects that data to supported AI workflows, including through Model Context Protocol. The practical opportunity is to resolve the correct business entity, retrieve authorised facts and preserve their provenance before a person or system uses them in a decision.

01 / ProductA business-data provider inside the AI workflow

D&B.AI places Dun & Bradstreet’s commercial information within AI-assisted work. The underlying company remains Dun & Bradstreet, with dnb.com as its corporate domain; dnb.ai is a product and developer destination. Clearlake completed its acquisition in August 2025, so the company should not be assessed as though its former public-market listing were still current.

The data proposition centres on identifying businesses and connecting relevant facts about them. The nine-digit D-U-N-S identifier gives a workflow a stable reference after matching. That is particularly useful when names are abbreviated, reused across subsidiaries or presented differently in several systems. An identifier helps organise evidence, but a proposed match still needs validation before downstream records are enriched.

The available tools separate tasks such as matching, data retrieval, payment information, ownership and provenance. An assistant can call a relevant tool instead of constructing an answer entirely from model memory. Access remains constrained by the tool and account entitlement. The existence of a tool in a catalogue does not mean it is included in a free preview.

02 / AudienceA fit for entity-specific commercial research

Teams researching customers, suppliers or counterparties are a natural audience. Their recurring problem is often not a lack of prose: it is uncertainty about which legal or operating entity a record describes and which facts apply to it. D&B can supply external context, while internal systems retain the organisation’s own relationship history and operational decisions.

This is also useful for developers building an assistant over business records. A clear entity reference can connect the user’s request to a bounded retrieval action. The developer still needs to decide when the assistant should ask for clarification, which fields it can display and how the result is logged. A broad request for everything about a company is a poor starting specification.

Quantexa is a useful comparison for entity resolution and connected decision intelligence. Moody’s is relevant when the research requirement centres on risk and financial information. Compare the specific entities, fields, freshness and permitted uses required by the workflow; neither a familiar brand nor a large dataset establishes that a particular question is covered.

03 / WorkflowA proposed supplier-context workflow

Start with a small set of supplier records containing names, addresses and the country in which the relevant entity operates. Include deliberate ambiguities, such as two subsidiaries with similar names. This is a proposed evaluation, not a test performed for this article. The goal is to observe whether the assistant can identify the right entity and explain the information it retrieves.

Use the matching capability to propose an entity, then examine the identifying details before accepting the D-U-N-S number. A parent company and its local subsidiary may both be legitimate businesses while being wrong for the record being enriched. Keep the original input and the accepted match together so another reviewer can understand why the link was made.

After the entity is confirmed, request only the data needed for the task. Separate firmographic context from payment, risk or ownership information, because each answers a different question. The tool catalogue helps define that boundary. Do not let a convenient natural-language interface conceal which service returned each field or whether the account had access to it.

Preserve provenance beside the generated narrative. Ask the assistant to distinguish a retrieved fact from its interpretation, and retain the relevant source context for later review. Where a field is absent or unavailable, return that state clearly. Filling a gap with a plausible industry assumption would defeat the purpose of connecting an authoritative data source.

For a developer pilot, follow the MCP setup using the credentials and supported client configuration supplied through the access process. Keep credentials outside prompts and ordinary output logs. Verify a narrow retrieval before introducing the integration into a broader assistant that can write to operational systems.

Finish with a human review of the proposed enrichment. An appropriate first result is a draft update linked to the accepted entity, rather than an unattended change to every similar-looking account. Measure incorrect matches, missing fields, repeated queries and time spent resolving ambiguity. These observations reveal whether the bottleneck is entity identification, data coverage or the surrounding application design.

04 / PricingTwo previews with materially different limits

The preview limits page describes the MCP preview as free for 30 days, with 100 domestic records and 100 rest-of-world records. Domestic scope is determined by the country supplied during signup. Consumption is based on unique D-U-N-S records under management: repeated enrichment of the same record does not create another unique record for that allowance.

The prompt-interface preview is different: 25 transactions over 90 days. Do not merge that allowance with the MCP route or assume the longer duration applies to developer access. Getting started and access options describe the routes. This research did not submit an application, receive credentials or establish that any particular organisation would be admitted.

Production pricing and rights need a scoped commercial agreement. Confirm the relevant data products, countries, update expectations, permitted uses and integration route. The catalogue restricts some capabilities: SBRI access requires the relevant existing customer relationship and is outside the preview, while other tools point to additional access arrangements. A preview credential is not evidence of unrestricted production access.

05 / DistinctionsIdentity and provenance make the integration useful

The strongest distinction is the connection between a natural-language request and a recognised business entity. Without that step, a model may combine information about similarly named firms or apply parent-level context to the wrong subsidiary. D&B provides a structured basis for retrieval. The application must still expose uncertainty and allow an incorrect match to be corrected.

Another distinction is the ability to bring the information into the user’s existing AI environment through MCP. That can reduce the need to manually copy a company name into a separate research interface. It also moves responsibility into the integration: the assistant needs a clear tool description, appropriate permissions and a way to retain source context when it produces a summary.

Dun & Bradstreet’s September 2026 announcement extends AI-assisted commercial analysis through its platform and MCP access. This is evidence of an active product direction. Vendor-reported improvements are not a substitute for measuring the organisation’s own research process, and this article makes no tested speed or decision-quality claim.

06 / QuestionsResolve rights, freshness and incomplete matches

The first open question is fit between the data contract and the intended application. An internal analyst view, a customer-facing response and a stored enrichment table may create different requirements. Ask the supplier to confirm the intended use explicitly. The public preview pages alone do not establish every right to retain, redistribute or expose returned information.

The second question is freshness at the field level. A recently retrieved response can contain underlying observations from different periods. Ask how the relevant data is updated and how the interface communicates missing, stale or unavailable information. The answer should be assessed against the business process rather than a blanket expectation that every field is real-time.

The third question is what happens when matching remains ambiguous. Build a stop-and-review path and allow the original input to remain unresolved. Also test the end of a preview allowance, an unavailable tool and an authentication failure. A useful integration reports those conditions clearly instead of producing an unsupported answer from its own memory.

07 / DecisionStart with a correct entity and a narrow question

Dun & Bradstreet is a strong candidate when an AI workflow needs dependable business context tied to identifiable entities. Begin with one repeated research task, confirm the matching and licensing boundaries, and keep the evidence visible. Wider automation becomes more credible only after the team can explain both the returned facts and the gaps.