Many organizations struggle to deploy AI systems in production due to messy, unstandardized data with inconsistent naming, null values, and varying formats.
Traditional approaches like standardizing data into a single system (Snowflake, Databricks), Master Data Management (MDM), or manually built semantic layers have not solved the problem.
Data domains change frequently, making it impossible to keep manually defined semantic layers up-to-date.
McKinsey reported that Fortune 500 companies lose an average of $250 million due to poor data quality.
Manually adding definitions to semantic layers is insufficient because business concepts (e.g., "new customer," "marketing spend") have many nuances and edge cases that cannot be exhaustively predefined.
Knowledge graphs, while useful for relationships, struggle to capture the dynamic, implicit "at risk" definitions or to integrate with large, constantly changing operational data stores like Snowflake tables.
The fundamental problem is that AI, particularly vanilla LLMs, does not inherently understand a business's specific language, tribal knowledge, or unique definitions (e.g., "GM" meaning gross margin vs. general manager).
The proposed solution is an "agentic semantic layer," which functions like a newly hired, smart analyst who learns over time.
This AI system continuously improves and learns from user interactions, corrections, and steering, becoming an experienced analyst for the company's specific domain.
PromQL aims to build such an AI by using a foundational LLM to generate "PromQL plans," a deterministic domain-specific language for data retrieval, computation, and semantics.
The LLM generates the plan, but the execution and answer generation are handled by a deterministic runtime on a distributed query engine, preventing hallucinations and ensuring reliability by not feeding data back to the LLM for final answers.
In a demonstration, PromQL initially failed to find top customers by revenue because it assumed "succeeded" status, but the actual data used "paid" and "pending," which it then learned to correct.
PromQL can handle complex, multi-system queries, such as identifying unique organizations from email domains, finding the third-highest revenue org, analyzing support tickets for sentiment, and issuing credits via API.
The system explains its thought process, and users can edit its "brain" (the generated plan) to course-correct, leading the AI to learn and adapt to specific business logic.
PromQL learns from user interactions, suggesting metadata improvements and updating its semantic layer (e.g., understanding cryptic table names like "zorp" and "plug," or that a "budget" is in cents).
Every instance of the semantic layer is version-controlled, allowing for rollbacks and continuous improvement.
This learning process allows the AI to infer meanings, find relationships across tables, and adapt to company-specific definitions like the start of a financial quarter.
An agentic semantic layer enables an AI to evolve from knowing nothing on "Day Zero" (e.g., what an enterprise customer means, how to match customer IDs, financial quarter starts) to becoming 100% accurate on complex tasks by "Day 30."
It learns business terms, maps relationships across systems, and discovers calculation variants.
This approach significantly reduces the time and effort traditionally spent on data readiness, allowing immediate AI deployment and self-improvement.
Customer testimonials from a Fortune 500 food chain and a high-growth fintech company reported 100% reliable AI on their hardest questions after evaluating many other vendors.