Laurent Giraid has spent his career at the intersection of high-scale data infrastructure and the practical deployment of artificial intelligence. As a technologist who witnessed the evolution of machine learning from academic curiosity to the backbone of modern enterprise, Giraid brings a unique perspective on why the current generation of AI agents often stumbles when faced with real-world complexity. He is deeply invested in the mechanics of natural language processing and the ethical guardrails required to keep automated systems from spiraling into misinformation. In this conversation, Giraid explores the critical role of semantic context, explaining how historical data usage acts as the essential map that prevents agents from hallucinating when navigating massive, multi-layered data environments.
The discussion centers on the persistent issue of high error rates—often exceeding 65%—when AI agents are dropped into sprawling data estates without proper guidance. We examine how DataHub’s new Context Intelligence layer aims to bridge this gap by mining years of validated SQL query logs to build a “living knowledge base” for these agents. The conversation moves through the technical transition from raw schema to “semantic anchors,” the strategic importance of keeping a platform-neutral context layer in an increasingly competitive market, and the vital role of human experts in resolving conflicting business logic. Giraid also highlights the practical success of this approach in massive Snowflake environments, such as the 10,000-table ecosystem managed by Miro.
When the data team at Miro first began testing AI agents against their Snowflake environment, they discovered a staggering failure rate, with the models getting the wrong answer more than 65% of the time. Why does simply providing an advanced model with access to a large data estate lead to such a dramatic breakdown in accuracy?
The core of the problem isn’t that the large language models lack the raw intelligence to write code; it’s that they are essentially flying blind without a map of the business’s internal logic. When an agent is dropped into a massive environment like Miro’s, which contains more than 10,000 individual tables, it faces a paralyzing number of choices without any semantic layer to guide its routing. The agent sees a sea of raw schema and has no way to distinguish which specific data assets actually correspond to nuanced, real-world business questions. This lack of context leads to what technologists call “hallucinated joins,” where the agent attempts to connect tables in ways that make no sense mathematically or logically. By relying solely on column names and raw schema, the AI is forced to guess instead of knowing, resulting in that frustrating 65% error rate that makes the system unusable for high-stakes enterprise decisions.
DataHub is now leveraging years of query history to solve this specific context gap. How exactly does mining historical SQL logs create a more reliable roadmap for AI agents than just reading the database schema?
A database schema tells you what exists, but query logs tell you what actually works and what has been validated by the humans who run the company every day. DataHub is essentially inverting the traditional “text-to-SQL” paradigm by turning years of analyst history into a retrievable knowledge base of what we call “golden queries.” This infrastructure isn’t just something built for a new release; it’s grounded in the same robust plumbing that has supported over 3,000 production deployments and managed lineage for more than 100 metadata sources. By extracting patterns from these high-quality, scheduled pipelines, the system creates structured text definitions called semantic anchors that the AI can reference. When an agent prepares to generate SQL, it first consults these anchors to see how successful analysts have joined these tables in the past, effectively standing on the shoulders of the 15,000 contributors who have hardened the project’s logic over the last 11 years.
One of the most interesting aspects of this new layer is the concept of “filtering for signal” from noisy warehouse logs. Could you describe the process of distilling millions of raw logs into these highly useful “golden queries”?
Warehouse query logs are notoriously messy, filled with experimental one-off queries, failed syntax, and redundant calls that would only confuse an AI agent if they were fed into it directly. To find the signal, the engine must isolate the “golden queries”—those high-quality analyst workflows and scheduled production pipelines that represent proven, hardened business logic. We move beyond simple metadata to understand the true intent behind the data flow by analyzing how information moves from operational systems like Postgres and MySQL into cloud warehouses like Snowflake or BigQuery. The engine extracts the underlying patterns and translates them into structured text definitions that act as a source of truth for the model. This ensures the agent doesn’t repeat the mistakes of a junior analyst’s first-day experiments, but instead follows the proven paths already cleared by the organization’s top data experts.
In the Miro case study, the team moved away from exposing raw schema and instead organized data into “well-defined data products.” How does this architectural shift change the experience for a user asking a question via a tool like Claude Chat?
This shift is fundamental because it places a “context layer” as a protective barrier between the user’s natural language request and the final SQL generation. At Miro, instead of letting the agent wander aimlessly through 10,000 Snowflake tables, the architecture routes the request through a DataHub MCP that maps user intent to specific, pre-validated data products. When a user asks a complex question via Claude Chat or Claude Cowork, the context layer pulls in entity relationships and business intent to identify the correct database entities before a single line of code is written. This prevents the agent from guessing and ensures that the final SQL generation is based on a narrow, accurate set of inputs rather than a confusing web of raw data. It transforms the experience from a high-stakes guessing game into a precise, governed workflow where the agent understands the “why” behind the data it is accessing.
You’ve mentioned the importance of human validation in this process to resolve conflicting definitions. Why is this “human-in-the-loop” element still so critical when we are trying to automate these data processes?
Even the most advanced AI cannot resolve a fundamental organizational conflict where two different departments calculate the same metric using completely different logic. DataHub surfaces these discrepancies, highlighting cases where definitions conflict and raising them for a domain expert to resolve manually. The Context Hub allows these experts to review AI-proposed context and even simulate the impact of changes before they are published to the “living knowledge base.” This step is vital for governance and regulatory compliance, as it ensures that the semantic index is not just a reflection of automated patterns, but a curated repository of truth. Without this human layer, you run the risk of simply automating chaos; with it, you create a system that grows more accurate and trustworthy every time a human expert clarifies a definition.
With giants like Microsoft, Oracle, and Pinecone all building their own contextual memory and semantic layers, how does a platform-neutral approach like DataHub’s stay competitive in what some are calling a “platform war” for context?
The competitive edge lies in the reality that most modern enterprises are not monolithic; they operate across a diverse patchwork of environments including Snowflake, Google BigQuery, and various on-premise databases. While many vendors are focused strictly on structured tables, the real value comes from the ability to integrate diverse metadata for both structured and unstructured objects, including documents and images. DataHub positions itself as the “connective tissue” that provisions context into existing endpoints like Microsoft Fabric IQ or Snowflake semantic views, rather than trying to replace them. In this burgeoning “platform war,” the winner isn’t necessarily the one who owns the data warehouse, but the one who controls the decision layer at runtime. By staying platform-neutral, you provide a unified intelligence layer that works regardless of where the data actually sits, which is exactly the kind of flexibility that a complex, global enterprise requires.
What is your forecast for the future of AI agents in the enterprise over the next three years?
Over the next three years, we will see a shift where AI agents are no longer viewed as standalone novelties but as the standard interface for all enterprise data, provided we master this context layer. We are moving away from passive data cataloging toward a world of continuously refreshed semantic intelligence where the bottleneck is no longer the model’s raw power, but the quality of the “memory” we provide it. I forecast a major consolidation in the market where context management becomes a core pillar of IT strategy, treated with the same level of importance as security or cloud storage. If organizations can successfully turn their years of silent query history into active, retrievable knowledge, we will see that 65% failure rate drop to near zero, unlocking a level of operational speed that makes our current manual data analysis look prehistoric.
