# Catch the Butterfly: PuppyGraph Raises $10M Series A to Build the Data Engine for AI Agents > Today, we’re announcing PuppyGraph’s $10M Series A, led by Particle Future with participation from defy.vc, bringing our total funding to $15M. - Canonical URL: https://www.puppygraph.com/blog/series-a - Published: 2026-11-10 - Author: Weimo Liu Sometimes the most important discoveries begin with something small: a connection nobody thought to look for. A delayed shipment from a supplier. An unexpected change in a product component. A pattern buried across thousands of enterprise tables. On its own, each detail may seem insignificant. But when connected, these details can reveal a chain of events that fundamentally changes how a business understands a problem. This is the Butterfly Effect: how small events can set much larger consequences in motion. In enterprise data, the challenge is that no one knows in advance which small detail matters. Finding it means exploring many possibilities and testing each one against the data. There is a fundamental difference between building an AI copilot that answers questions about data we already know exists and giving an AI agent a business problem, then letting it investigate thousands of enterprise tables to uncover the connections that matter. We believe this represents the next evolution of enterprise intelligence: people describe the business problem, and AI agents take on the investigation—discovering hidden relationships, tracing dependencies, and surfacing insights that might otherwise go unnoticed. Today, we’re announcing PuppyGraph’s $10M Series A, led by Particle Future with participation from defy.vc, bringing our total funding to $15M. We’re building the data engine that makes this possible: enabling AI agents to explore enterprise data as one connected graph, without moving or duplicating the underlying data. ## From AI Copilots to Agents That Investigate Look at how AI shows up in business intelligence today. A person who knows what data exists wants a dashboard, and an AI assistant builds it. Or someone asks a precisely framed question in natural language, such as "What was revenue by region last quarter?", and Text-to-SQL turns it into a query. Both are useful, but they are still copilots. The person does the analyst's job: knowing the data, deciding what to look at, and framing the question. AI makes each step faster. Agents can now do much more. They can plan an analysis, execute queries, read the results, and decide what to investigate next. Where an analyst might have time to chase one or two leads, an agent can generate many hypotheses, test them against the data, and discard those that the evidence does not support. That makes a different kind of business intelligence possible. People describe the outcome they care about, and the agent works out how to investigate it. Nobody has to direct every query or know in advance which tables contain the answer. Consider a sales leader asking: "Why are we missing our sales target?" A useful investigation might begin by separating fewer deals from smaller deals and deals slipping into the next quarter. Suppose the agent discovers that slipped deals account for most of the gap. It then looks for what those deals have in common, testing the obvious explanations first. The data rules them out one by one: the slipped deals are not concentrated in a region or a sales team, they don't follow a pricing change, and they are not the deals where a competitor was in play. What many of them do share is that they involve products with long delivery delays. The agent follows those products to their components, then follows the components to their suppliers. One supplier has fallen behind on a part used in several affected products. That is a lead worth testing. The agent compares delivery dates and deal histories, checks similar products that do not use the same component, and looks for exceptions. It returns a likely explanation, the evidence supporting it, and the questions that remain unresolved. The likely explanation sits several steps away from the original sales data: from deals to products, from products to components, and from components to a supplier. A sales leader would rarely think to start there. The agent did not need to know where to start. It tested possibilities, dropped the ones the data did not support, and let each result determine the next query until the evidence pointed somewhere. This is the shift we are building for: **the agent as the analyst, not the analyst's copilot.** ## Why Enterprise Agents Need More Than Text-to-SQL SQL is extremely effective for filtering, aggregations, and complex analysis. Text-to-SQL makes these capabilities accessible through natural language, especially when the relevant tables, relationships, and business definitions are already well understood. The harder problem begins when the agent must discover those definitions and relationships across thousands of tables. A company may have data spread across databases, warehouses, and data lakes, with thousands of tables and few people who understand how everything connects. Knowing SQL syntax does not tell an agent which connections are meaningful in a particular business. Two columns can share the same name and data type without representing the same thing. A student identifier and an employee identifier might both be integers. Joining them could produce a valid SQL query and a convincing-looking result, even though no business relationship supports that connection. An execution error is visible. A successful query built on the wrong relationship can be much harder to catch. In an investigation, an agent may use that result to determine its next question, carrying the mistake through several more steps. As the number of tables and possible relationships grows, providing more schema descriptions to an agent does not necessarily solve the problem. More context consumes more of its available context window, and the agent still has to apply those descriptions correctly to every query. Much of this knowledge previously lived with the people writing the queries. They knew which customer identifier to use, what a transaction represented, and which joins were valid. An enterprise agent needs access to this business knowledge in a form it can discover and use throughout an investigation. ## PuppyGraph: The Data Engine for AI Agents We believe the answer is to give agents a comprehensive knowledge graph built over enterprise data, with the ability to query that graph directly. The graph's ontology defines the business entities—customers, deals, products, components, suppliers—and the relationships between them. The underlying records remain in the enterprise's existing data systems. This gives an agent a structured map of the business to explore before retrieving the data it needs. It can discover that a deal involves a product, a product contains components, and a supplier provides a component. These relationships define valid paths for investigation. An ontology can also serve as reference material to help an agent write SQL. With PuppyGraph, however, the agent can execute graph queries directly through the relationships defined in the model. The ontology becomes part of the query interface, rather than just additional context in a prompt. If the agent references an entity or relationship that the ontology does not define, PuppyGraph rejects the reference and returns structured feedback. The agent can inspect the model again and correct its query. It might discover a valid path through intermediate entities, or determine that the relationship it needs has not yet been modeled. The team defines and approves the business relationships; it does not need to anticipate every question someone might ask about them. Once those relationships exist, agents can combine them into new investigations. This does not guarantee a correct conclusion. An agent can still have a poor hypothesis or conduct an incomplete analysis. It does address a specific source of silent errors: an agent inventing a relationship outside the approved graph model and treating the resulting connection as evidence. Establishing that model requires business knowledge. PuppyGraph's AI-assisted schema builder explores existing tables and proposes entities and relationships, reducing the manual work required to create the ontology. The data team reviews and approves the model, including the business meaning that table structures alone cannot establish. PuppyGraph solves two fundamental technical problems: making enterprise knowledge graphs practical to establish at very large scale, and providing a distributed graph query engine fast enough to support an agent's repeated investigations. Our engine executes multi-hop graph queries directly over supported databases, warehouses, and data lakes, without requiring teams to load their data into a dedicated graph database or maintain a separate graph copy and its loading pipelines. ## What Is Working in Production PuppyGraph already enables agents to query across 3,000 enterprise tables as one unified graph. The practical requirement is to let agents discover relevant relationships and retrieve evidence as their investigations develop, without putting every table definition into a single prompt. As Leon Goldberg, CTO of Sola Security, puts it: "We run AI agents across thousands of tables and billions of edges a day." PuppyGraph 1.0 brings together the capabilities required for these workloads: - **AI-assisted graph schema generation:** Proposes business entities and relationships from existing enterprise tables, helping teams establish an ontology with less manual effort. - **Ontology validation:** Rejects undefined entity and relationship references and returns structured feedback that agents can use to correct their queries. - **Distributed graph execution:** Runs multi-hop traversals directly over supported databases, warehouses, and lakes. - **Enterprise security:** Enforces configured row-level permissions during graph traversals, including intermediate steps. Role-based access controls and user impersonation tie data access to user identity. - **Agent integration:** Supports MCP, allowing compatible AI tools and PuppyGraph's built-in agent to access enterprise data through graph queries. Performance matters because an agent explores by testing many hypotheses, and every test takes queries. A query that takes minutes can interrupt the entire process when the next analytical step depends on its result, and it limits how many possibilities the agent can afford to explore. PuppyGraph supports subsecond graph queries for supported workloads. In a separate benchmark, it completed a 10-hop query across billions of edges in 2.26 seconds on a four-machine cluster. Performance depends on the workload, but our engineering goal is consistent: make deep relationship queries fast enough to support interactive investigations. Our customers include AMD, Palo Alto Networks, Datadog, Netskope, Coinbase, Trend Micro, Baker Hughes, Sola Security, HighStreet Insurance Partners, and Prevalent AI. PuppyGraph's annual recurring revenue grew 7x over the past year. We also see the opportunity expanding within customers. AMD began with IT workloads, and that success has opened the door to connecting information security and digital marketing data. Other enterprises have expanded from observability into fraud detection, and from initial agentic workloads into additional business domains. As a relationship model connects more systems and departments, it becomes useful for questions that previously stopped at organizational boundaries. ## Keep the Platform You Already Have PuppyGraph is designed to work alongside an enterprise's existing data platform, not replace it. An agent can use both its existing SQL query engine and PuppyGraph. It can issue filters and aggregations through SQL, then use graph queries when the investigation calls for following relationships across entities. Applications can choose the appropriate engine for each question. PuppyGraph brings graph compute to data in supported systems including Apache Iceberg, Databricks, Snowflake, ClickHouse, DuckDB, PostgreSQL, Microsoft Fabric, and Cloudera. Teams can add relationship queries without moving data into a dedicated graph database or maintaining another copy. Agents can access PuppyGraph through MCP using compatible AI tools or PuppyGraph's built-in agent. The data platform team can introduce this capability within its existing architecture. Security remains essential throughout the investigation. PuppyGraph 1.0 enforces configured row-level permissions during graph traversals, including intermediate steps. Role-based controls and user impersonation ensure that an agent's ability to follow relationships remains bounded by what its user is authorized to see. This combination matters: agents need broad access to discover relationships across enterprise data, but that access must remain governed by the enterprise's existing security requirements. ## Where We Are Going We believe enterprise analytics will increasingly begin with a person describing a business problem and an agent conducting the investigation. The result should not simply be another dashboard or a generated SQL statement. It should be an explanation supported by evidence that a person can examine, question, and act on. A sales leader could investigate supply-chain context. A security investigator could trace relationships behind an exposure. An operations team could follow a disruption across systems. Each investigation would work within the user's permissions while reaching beyond the data someone had already assembled for a dashboard. That future depends on the data platform. Models bring general knowledge and planning capabilities. Enterprises need to supply their own entities, relationships, business context, and access rules, together with an engine that can retrieve evidence at the scale of their data. Our Series A will fund engineering and enterprise go-to-market expansion. Our priorities are clear: reduce the work required to establish the graph model, make relationship queries practical at enterprise scale, and help more teams put AI agents into production. We are grateful to our customers for trusting us with these workloads, to our team for building the engine behind them, and to Particle Future and defy.vc for supporting the next stage. ## See an Investigation on Your Data Platform If you are building enterprise agents and have run into the limits of discovering relationships, following business rules, or executing deep queries across your data, we would like to show you what we have built. [Book a demo](https://www.puppygraph.com/book-demo) to see an agent investigate a business problem with PuppyGraph and discuss how it could work with your existing data platform.