AI Engineer World's Fair 2026
Your Moat Is Your Data Model
Loading slides…
Read the talk
Your Moat Is Your Data Model
The Gates Foundation’s Strategic Intelligence Platform turns operational knowledge into a graph that agents can traverse, while keeping reporting rules and organizational context explicit.
From a talk by Mike Phipps
Before you start: Familiarity with database relationships and AI tool calling will help; the article explains the graph structures as they appear.
What should an AI platform team own?
If tools such as Claude Code make applications faster to build, what remains defensible once those applications reach production? Shipping introduces monitoring, upkeep and downstream dependencies. The team must decide how much of that stack it wants to maintain—and whether users need another chat application when they already work in Claude or ChatGPT.
For Mike Phipps and his team, the durable asset is knowledge of internal processes: the tacit understanding required to interpret organizational data correctly. A better model does not automatically know what an internal field means, how a reporting convention works or which relationships matter to a business question. Encoding that knowledge gives the team something useful across successive models and interfaces.
That work became the Strategic Intelligence Platform, or SIP. Phipps reports that SIP entered enterprise production in the preceding month for a population of about 4,000 people across the Gates Foundation. The platform makes operational knowledge available to AI without making a proprietary chat interface the center of the product.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The organization the graph must represent
The foundation’s work spans child mortality, nutrition, agriculture and education, with more than 25 years of accumulated operational data. Its interventions include creating market incentives, encouraging innovation and connecting public and private organizations. Extracting useful evidence from those investments requires understanding the structure around the data, not merely searching the documents it produced.
Phipps’s 2023 snapshot describes more than 2,000 grants, many above $5 million, targeting more than 100 countries. The scope also includes grantees, alumni, employees, multiple program strategies and activity across almost all U.S. states. The financial scale is substantial: the 2023 annual report records $7.749 billion in total charitable support, distinct from $6.827 billion in direct grantee support. The broader total should not be read as grant payments alone.
Funding flows through divisions including Global Development, Global Health, Gender Equality and the United States Program, abbreviated USP. Those divisions, their strategies and their investments are part of the structure an agent must understand when answering questions about the foundation’s work.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From siloed records to shared meaning
SIP structures operational data for agentic retrieval. Its knowledge graph is designed with an agent as the consumer. Traditionally separate systems of record supply structured and unstructured information to a data lakehouse, bringing internal enterprise records and programmatic outputs of investments under one roof. A curation layer processes that material before SIP exposes it through agentic chat and workflows. The result is a semantic graph that agents can reason across, rather than a collection of isolated source systems.
Consolidating storage does not settle the meaning of the records. Data owners must explain field definitions, dataset structure and valid joins. They also supply knowledge of limitations, systematic quirks, safeguards, security trimming and reporting conventions. A technically plausible answer can still fail if it uses a definition that differs from the organization’s established reports. Correctness includes the organization’s rules for interpreting its data.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Prepare the data before making it easy to retrieve
The curation pipeline handles several distinct responsibilities:
- Preprocessing: Filter and deduplicate datasets, preserve meaningful document ordering and resolve inconsistencies across documents upstream.
- Extraction: Extract structured fields, divide unstructured documents into semantic chunks and convert information in figures into text that retrieval can reach.
- Graph enrichment: Use tags to establish connections and generated metadata to populate graph properties.
These steps determine what information is available to the agent and how that information connects to the rest of the organization.
Governance becomes more consequential when AI makes information easier to find. Data that was technically accessible but difficult to discover can become immediately available through a natural-language question. Phipps therefore treats PII masking, sensitive-data classification and per-user entitlements as necessary parts of the pipeline. Existing accessibility is not sufficient evidence that a dataset is safe to expose through a much more effective retrieval interface.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Start with annual reviews, then follow the funding
The graph begins with a recognizable organizational process: annual strategy reviews. Phipps describes more than 80 strategy teams whose annual reviews determine yearly budgets. Review meetings bring unstructured documents into the model while providing structured links to other systems of record. The displayed model is conceptual: a single type on the diagram can represent many instantiated nodes, with one-to-one or one-to-many relationships in the actual graph.
The first hierarchy is funding. Phipps describes a five-level additive directed acyclic graph, or DAG, in which all levels matter together. Roll-ups must preserve the funding path rather than treat a single level as the complete answer. An in-path shortcut connects that path, while the relationship he calls “Funds to BAU” is where funding-team budgets are stored.
Funding teams have portfolios that support investments, and multiple teams can fund the same investment. This makes funding a many-to-many relationship, not a single owner field on an investment record. Investments are the foundation’s product; internal funds express priorities among different kinds of investments. The model can retain transaction-level detail or represent annual aggregations, depending on the question being answered.
From an investment, the graph can extend to different kinds of organizations and to observable outputs such as published reports and products. Connecting those outputs to the investment also connects them to its organizational context. Phipps identifies this as an area with substantial room still to populate, rather than a finished inventory of everything the investments have produced.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Management is a different relationship
The management hierarchy answers a different question about the same investment. Here, each organizational level has meaning independently; the essential distinction is how attribution works, rather than whether it follows the funding hierarchy’s topology. Contains relationships connect teams through the hierarchy. A team can manage an investment directly, while an ancestor team also receives indirect attribution for investments managed by its descendants.
SIP precomputes that indirect attribution. After creating the containment and direct-management relationships, it derives a management roll-up edge. This lets retrieval use a relationship whose organizational meaning has already been established instead of rediscovering the roll-up rule for every question.
For a small teaching example, suppose team-parent contains team-child, and team-child directly manages investment-a. With illustrative Cypher labels and relationship names, the derivation can be expressed as:
cypher
MATCH (ancestor:Team)-[:CONTAINS*1..]->(direct:Team)
-[:MANAGES_INV]->(investment:Investment)
WITH DISTINCT ancestor, investment
MERGE (ancestor)-[:ROLLUP_MANAGES_INV]->(investment);
For those records, the query would add a roll-up from team-parent to investment-a, preserving the original containment and direct-management relationships. DISTINCT collapses repeated ancestor–investment pairs before MERGE creates or matches the derived edge.
Both lenses remain available on the same investment:
| Lens | Relationship being represented | Roll-up concern |
|---|---|---|
| Funding | Teams contribute through funding paths | Preserve additive contributions |
| Management | Teams manage directly or through descendants | Attribute descendant investments to ancestors |
A question about a team’s investments can therefore require clarification: does the user mean what the team funds, what it directly manages or what falls under its management hierarchy? Keeping the relationships separate makes those interpretations representable.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
People connect the organizational views
People provide a third hierarchy. Organizational charts, reporting relationships and team membership connect individuals to the operating structure. The same person can also appear as an investment owner, a meeting attendee or a director. Those roles add context that would be lost if personnel records remained isolated in an HR system.
The cross-system graph depends on finding shared entities in those formerly separate sources. A person, team or other organizational entity becomes a point where related records can be stitched together. Once those connections exist, an agent can traverse the organizational structure instead of treating each source as a separate body of facts.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Documents become discoverable in context
At the time of the talk, SIP had ingested one document source. Its Neo4j model connects meetings to documents, then documents to semantic sections or chunks. Full-text indexes support searching that content, while the relationships keep each result attached to its meeting and organizational context. An agent can also retrieve information through graph traversal alone.
The complete model brings four source systems into one graph, exposed to agents as a semantic layer through the Model Context Protocol, or MCP. Agents can discover structure and reason across it at query time. The modeling process also benefits developers: drawing the connections quickly exposes missing datasets and gaps in the team’s own understanding of the organization.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Deliver the graph where users already work
SIP reaches users through MCP connections to the chat products they already use, including Claude and ChatGPT. The team chose not to own the chat interface and, for some general-chat use cases, not to own the agent interaction either. Its investment stays concentrated in the organizational knowledge those interfaces can access.
Phipps says the team forked an off-the-shelf Neo4j MCP server and modified it substantially. Changes included schema updates and mechanisms for passing state back to SIP, such as conversation IDs and message numbers. The current official Neo4j MCP server is a related implementation, not a specification of SIP’s historical fork or its private tool contract.
General chat is one entry point; constrained workflows are another, still under development in the talk. Phipps describes access through Cowork or Claude Chat, with MCP Apps bringing workflow interfaces into the chat experience. Sandbox-based agents would run the workflows against the same knowledge-graph backend. This keeps the shared data model useful while allowing a more controlled experience than an open-ended conversation.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Evaluate against the live graph
Evaluations reveal where a data model is incomplete or ambiguous. Users may ask questions that admit several interpretations, or receive answers that do not conform to established reporting standards. SIP’s team works with data owners to turn those standards into targeted evaluation questions.
The reference answer cannot simply be a frozen value because the structured data keeps changing. The evaluation procedure therefore preserves a reference query for each question:
- Organize questions by complexity. Separate simpler requests from those requiring more involved reasoning.
- Maintain a reference graph query. Encode the intended interpretation of each question in a query.
- Run it against the live graph. Retrieve the expected result when the evaluation runs.
- Compare the agent’s answer. Judge the response against that current result rather than an obsolete snapshot.
This separates changes in the underlying records from failures to interpret or retrieve them correctly.
The evaluation pipeline uses an LLM as a judge and tracks pass@1 alongside stability: whether repeated asks of the same question produce consistent answers. Failures feed back into the data model, domain rules and schema descriptions. Evaluation is therefore a way to improve the knowledge supplied to the agent, not only a way to score its final response.
Phipps describes the system’s pass@1 and stability as very strong, without giving numerical results in the spoken account. He says remaining misses tend to involve ambiguity: an answer can be valid under one interpretation yet differ from what the user intended. That distinction makes reporting definitions and domain rules part of the continuing engineering work, even when retrieval returns factually defensible results.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Expand coverage without losing the shared context
The next phase starts by filling gaps in records from systems already represented in SIP, then expanding the primary graph to additional enterprise datasets. Teams also want to connect their own data to the enterprise model through a federated graph experience; Phipps describes the approach as work in progress. Further agentic experiences build on that expanding foundation. The open engineering problem is how to make more of the organization’s knowledge available while preserving the shared meanings that make cross-system answers useful.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
The official introduction to interactive interfaces delivered by MCP tools inside chat clients.
Further reading
The foundation's 2023 spending breakdown by division and strategy, with definitions of its financial totals.
An MCP server providing graph-schema discovery and Cypher query tools for compatible AI clients.
Instructions for creating and querying full-text indexes over node and relationship properties.
Full transcript
Loading transcript…