Verified Data is the Missing Piece of Agentic Infrastructure With Gary Kotovets (Chief Data & Analytics Officer at Dun & Bradstreet)

Dun and Bradstreet maintains a commercial graph covering more than 650 million businesses, updated daily and intraday, collecting over 11,000 data elements per business across more than 200 markets globally. Gary Kotovets, Chief Data and Analytics Officer, argues that data quality is the foundational problem in agentic AI, drawing a direct line from his Bloomberg experience with millisecond-sensitive financial data to his view that as AI workflows become more autonomous with fewer humans in the loop, trust in agent decisions depends entirely on the quality of the underlying data. He treats bad data and hallucination as mutually reinforcing dangers, and frames D&B's verified datasets as grounded truth that removes the hallucination problem at the data layer while leaving model-serving as a separate challenge to manage.

D&B performs approximately 100 billion data quality checks monthly across all data refreshes in its supply chain, continuously monitoring and modifying quality rules as source data changes to maintain accuracy, timeliness, and consistency. Data staleness is described as a particularly difficult problem because it is hard to catch before it affects agent output and propagates downstream, and D&B addresses this through a program requiring upstream supply chain changes to be communicated through operational processes. The company runs a continuous pipeline to test and validate open source and frontier language models against its datasets and agentic task execution, treating model selection as context-dependent rather than categorical.

D&B started its AI journey around 2023 with a centralized platform for rapid development and deployment. Its risk platform includes a KYC agent and a credit check agent, each supported by sub-agents performing entity matching, firmographic data retrieval, beneficial ownership checks, corporate linkage, CEO and officer verification, physical location checks, supplier checks, and geopolitical and climate risk zone checks. The logic of existing risk analytics software was converted directly into agent prompts. D&B also built a generic tools library with rigid rules that all internal agent developers must call, but agents have been observed breaking those rules over time, which led to the implementation of monitoring capabilities to verify agents follow their designed steps.

D&B has deployed small language models across at least seven use cases, achieving a 95 to 97 percent reduction in token costs while also improving quality and accuracy because the models are more focused than large general models. A company summaries product using SLMs and LLMs generates deeper insights beyond self-reported industry codes, with one example being the identification that a landscaping company also performed tree pruning, a distinction that materially changes workers compensation insurance premium calculations. D&B also creates pre-calculated deterministic outputs such as credit scores and verified corporate hierarchies that agents can consume without performing their own compute, with Kotovets citing 50 to 70 to 80 percent savings in token costs for agents that use these verified sources instead of web scraping.

D&B has made its data available via MCP to serve as a context layer for agents embedded in customer CRMs, procurement applications, and data environments, and released an agent-to-agent protocol for agentic workflows. A business verification match agent was made available on Google Marketplace. Kotovets identifies a key unsolved challenge in multi-agent setups as verifying the agents themselves and confirming they belong to the business they claim to represent, particularly across different companies, describing this as a digital handshake problem that current infrastructure does not adequately solve.

Every model and use case workflow at D&B is reviewed by the legal team, the chief risk officer's team, and technology and product groups before deployment under a governance framework covering legal, compliance, and ethical dimensions. Kotovets warns that poorly governed environments force every new agent through approval chains involving ten people and five technology teams, making scale impossible, and that enterprises rushing to buy tools before addressing governance and data quality risk serious trouble. His recommended sequence is governance framework first, getting data in order second, and acquiring the right tooling stack third.

Kotovets predicts that as foundation models commoditize, the data and context layer will become the most valuable component of the enterprise AI stack, and that verification and validation of agent workflows will become increasingly critical with few providers able to deliver the capability. Looking three years out, he expects most enterprise decisions will involve agents and that it will become genuinely difficult to determine whether the counterparty in a B2B interaction is a person or an agent, adding the caveat that the right controls and monitoring must be in place to prevent agents from making unchecked wrong decisions. He calls small language models the most overhyped idea in enterprise AI despite deploying them successfully, argues fine-tuning is preferable to retrieval-augmented generation, and compares the overall AI landscape to the dot-com era in terms of how it is likely to evolve.