Enterprise AI at Scale: Inside the Dun & Bradstreet + LangChain Playbook

Dun & Bradstreet Data Scientists Talk Enterprise AI Progress, Pitfalls and Outlook with LangChain CEO

Artificial intelligence capabilities are advancing rapidly, but many enterprises still struggle to deploy them effectively in real‑world environments.

Generative AI (Gen AI) experiments have produced powerful prototypes over the past few years. But production is different: when faced with real data, real users, and real governance requirements, those prototypes struggle to perform. Systems behave inconsistently, outputs are hard to explain, and costs rise in a non-linear fashion as complexity and scale increase. 

Dun & Bradstreet can help. As a leading global provider of business data and analytics, D&B empowers clients' AI transformation journeys by pairing trusted, high-quality data with cutting-edge technologies — including from innovative partners like LangChain, whose tooling helps organizations build powerful, production-ready AI solutions at scale.

We hosted a webinar with LangChain CEO and co‑founder Harrison Chase, D&B AI leaders Ilya Meyzin (SVP of Data Science), Pierre Deville (VP of Data Science), and Thibault Dory (Senior Director of Data Science) to discuss the challenges facing enterprise AI adoption and explore solutions. From recent trends to promising developments, the conversation shed light on what it takes to make AI work for businesses at scale. Read on for the highlights.

Why AI Pilots Struggle to Scale

Early generative AI (Gen AI) tools were optimized for language fluency. They behaved like answer boxes: text in, text out — great for drafting, summarization, and other simple tasks.

But enterprise environments introduce constraints that answer boxes aren't built to handle. Business‑critical information lives across APIs, databases, and permissioned platforms, where access controls, data latency, traceability and explainability matter. Without structured retrieval and validation, models are left to infer or guess — a risk that compounds quickly at scale.

Just as importantly, answer boxes offer little control over how results are produced. They compress complex tasks into a single, opaque step, making it difficult to enforce predictable behavior, explain how an answer was reached, or trace and debug errors. In production environments, where consistency, auditability, and governance are requirements rather than nice‑to‑haves, that lack of structure becomes a liability.

In short, early AI pilots stall because the systems and processes around them were never designed for real‑world enterprise use. As Chase put it: “It's easy to get a prototype. It's hard to get something that works well enough to put into production.”

So how do you optimize your organizations' systems to support enterprise AI at scale? That's where the concepts of model intelligence and context come into play.

Understanding Model Intelligence and Context

At scale, AI model intelligence alone doesn’t determine production outcomes. In an enterprise environment, how that model is instructed, constrained, connected, and evaluated — that is, context — becomes the deciding factor.

Model intelligence determines what is possible. It defines:

• Complexity of tasks the system can attempt

• Depth of reasoning and synthesis it can perform

• Quality and coherence of language output

• Range of actions the system can plan or infer

Context determines what is practically achievable in production. Examples include:

• Clear instructions that define intent and boundaries

• Access to the right data at the right time

• Transparent tooling and execution logic

• Guardrails that enforce predictable, repeatable behavior

• Together, these elements determine how well an AI system performs.

“[A] big reason why [Large Language Models] mess up is that they might have the wrong context,” Chase said. “'I’d actually argue that this happens more often than not.”

He warns businesses to pay close attention to context engineering, saying “that's one of the biggest things that that I'd absolutely be paying attention to and thinking about, which is how can I engineer this context to get to the LLM in the right shape and format."

Bridging the Gap with AI Agents

Early enterprise AI attempts often failed because models were deployed on complex problems without (1) sufficient enterprise context and structure, and (2) robust reasoning ability to decompose tasks into solvable components and form a solution plan. Organizations compensated with extra human oversight, repeated prompting, and/or downstream validation. 

Agentic systems are a major step forward: (1) they can be grounded in rich enterprise context and structure (tools, data, permissions, constraints); (2) they represent a significant advance in system-level reasoning: they break problems down and plan tasks, select tools, retrieve and validate data, adapt execution based on intermediate results, and maintain traceability throughout. In other words, agents approach problem solving much like humans do!

"With the breadth and complexity of the data that we're managing at D&B, we realized very early on that agents were a big game-changer for us,” Deville said.

"Very early on, we had considered building multiple applications to talk to very specific data types,” Meyzin added. “Instead, [LangChain’s] framework, especially LangGraph, allowed us to build one application — which later became ChatD&B that we rolled out to our customers — that allowed us to talk to pretty much all of D&B data that that our customers use within one centralized agentic assistant. And that was quite an impressive breakthrough.”

In practice, this translates into a design choice about where to apply agentic reasoning versus deterministic execution. The most effective enterprise AI systems combine both modes: agents for judgment, coordination, or synthesis, and deterministic workflows where consistency and controls are critical. 

ChatD&B — our conversational Gen AI assistant, built using LangChain — follows the same model to deliver data from Dun & Bradstreet’s extensive knowledge graph in seconds. Its orchestrator agent breaks down complex requests into simple steps, routes each step to purpose-built deterministic tools (APIs, retrieval systems, validation layers), and then assembles the final response. The result is answers grounded in timely, trusted Dun & Bradstreet business data and analytics — including firmographics, risk signals, ESG attributes, creditworthiness, and more — along with transparent execution trail showing exactly which data and tools were used to generate the answers.  

Chase commented, "That's why [D&B] was among the first to put this kind of application into production. If you look under the hood of ChatD&B, it's way more complex than [calling an LLM]. There are all these pieces that bring some determinism in the right places but still bring flexibility. I think that balance is hard to get right.”

As a result, clients and thousands of Dun & Bradstreet team members can use ChatD&B to generate detailed, question-specific briefs in seconds, surfacing the breadth, depth, and diversity of D&B's trusted business data that traditional methods would take hours to complete. 

Don't Force Agents Where Simpler Patterns Win

A recurring theme in the webinar was judgment: use agents where they create clear value, and don’t treat every problem like an agent problem. 

Chase and the D&B team stressed that agentic design is powerful, but not all-powerful. "A lot of the hype is that everything should be an agent. Agents can do incredible things, but they're not needed for every use case,” said Dory. 

For well-bounded tasks — where execution is linear, business rules are explicit, and the sequence of steps is known in advance — deterministic workflows are usually the better choice. They tend to be faster, cheaper, and easier to govern. 

Agentic systems, on the other hand, add the most value when the path is non-linear: ambiguity is high, intermediate decisions are required, tools must be coordinated dynamically, or multi-step synthesis is needed.

Meyzin added, “in many cases, an agentic system is an overkill and a simple pipeline that uses LLMs for a subset of steps would do the trick. 

The goal isn't maximum agentic autonomy — it's maximum business value through the right architecture and controls.

MCP and A2A: Making Enterprise Agentic Capabilities Portable

As enterprise AI matured from RAG to agents, the next challenge became interoperability: how to make enterprise tools and data usable across numerous applications and third-party platforms.

Chase described MCP in practical terms: “MCP is a way for agents that you don't control to call custom tools that you build.” This is especially relevant for enterprises integrating with third-party agents. When organizations cannot modify the underlying agent, MCP provides a standard interface to connect internal tools safely and consistently.”

Dory provided a concrete example: “Through MCP, with configuration only…you can connect D&B's MCP server and suddenly your CRM agent has access to all the required dimensions in D&B data, with minimal required integration.” 

A2A follows a similar interoperability logic for multi-agent collaboration, though it’s still earlier in maturity. The core takeaway: adopt open protocols where they reduce integration friction and expand reuse, not as an end in themselves.

Takeaways for Enterprise Leaders

ChatD&B's rapid path from concept to production points to a practical playbook for enterprise AI:

Start from a clear business problem. As Chase emphasized, clarity on the problem is the biggest predictor of progress.

Choose the simplest architecture that meets business requirements: agentic, deterministic or hybrid. Deterministic workflows should handle linear, rules-driven steps; agentic orchestration should handle ambiguity, tool selection, and synthesis.

Iterate quickly using real-workflows and production-like data. “Move from prototype to pilot fast, evaluate against actual user behavior, and improve through short feedback loops,” said Deville. 

Treat context engineering as a first-class discipline. Reliability depends on the quality of instructions, retrieval, tool outputs, and formatting passed to the model. Better context usually beats prompt cleverness. 

Design for production from day one. Observability, auditability, access controls, fault tolerance, and cost governance are not phase-two concerns; they are prerequisites for scale. 

Apply interoperability standards pragmatically. Adopt MCP/A2A where they reduce integration friction and increase reuse across customer and third-party applications.

Meyzin closed the conversation: "Looking at the AI landscape, we're not just witnessing a technological shift, but a truly strategic business shift. I think the message here is the time to act is now, and that means building with intention, investing in the right infrastructure and talent, and staying very focused on what truly drives value.”

Learn more about D&B.AI: https://www.dnb.com.hk/solution/hot-topics/generative-ai