The durable advantage in enterprise AI is an organization's data model and tacit knowledge about how its questions should be answered.
2
The Gates Foundation's Strategic Intelligence Platform connects structured and unstructured data in a Neo4j knowledge graph that agents access through MCP.
3
Retrieval evaluations expose gaps and ambiguities in the data model, which the team fixes through updated domain rules, schema descriptions, and graph queries.
Summary
Mike Phipps argues that models, chat interfaces, and agent frameworks are becoming easier to replace, while an organization's understanding of its own processes remains harder to copy. At the Gates Foundation, his team built the Strategic Intelligence Platform to model 25 years of grantmaking, internal operations, people, organizations, funding paths, management structures, and documents. The platform brings siloed systems into a shared data lakehouse, processes structured and unstructured content, and exposes the resulting Neo4j knowledge graph to agents through MCP. The graph supports different views of the same work, including funding and management hierarchies. Documents are split into semantic sections and connected to the organizational entities they describe. Phipps also explains how data owners contribute reporting conventions and other tacit knowledge, then help create evaluation questions. Live graph queries are compared with expected answers, while stability and judge-based evaluation reveal places where the model or domain rules need revision.
An organization's process knowledge is harder to replace than its AI stack
Phipps starts with the question of what remains defensible when models, coding tools, and chat products change quickly. Production systems bring monitoring, upkeep, dependencies, and questions about where users should access the system. His team decided that its advantage was understanding internal processes and the tacit knowledge required to run them. That knowledge is being modeled in the Strategic Intelligence Platform, which rolled out for enterprise use across the Gates Foundation. Phipps says new model releases do not threaten this part of the system because the modeled process knowledge belongs to the organization.
The platform turns a large, long-running operation into agent-retrievable data
The Gates Foundation has more than 25 years of work across areas such as child mortality, nutrition, agriculture, and education. In one year, its operating picture included more than 2,000 grants, 4,000 employees, activity across more than 100 countries, and more than 7 billion dollars in annual disbursement. Phipps presents this scale to explain why extracting insight is difficult. The platform brings internal systems of record and programmatic data into a data lakehouse, then applies a curation layer before exposing a cross-system semantic graph to an agentic interface.
Data owners supply the rules that make answers match organizational practice
Phipps says engagement with data owners is required when modeling complex systems of record. The team needs their help to understand field meanings, joins, limitations, safeguards, security trimming, and reporting conventions. Correctness includes answering a question in the way the organization has answered it before. The curation pipeline handles structured and unstructured inputs through preprocessing, filtering, deduplication, extraction, semantic chunking, tagging, and added metadata. Governance is part of the same work because AI makes previously accessible data easier to expose. The team masks personally identifiable information and reviews sensitive classifications and user entitlements.
Multiple hierarchies let the graph answer through different organizational lenses
The graph models more than one hierarchy. In the funding hierarchy, all levels of an additive directed acyclic graph matter, and paths connect funding teams to portfolios, investments, and budgets. Multiple funding teams can support one investment, and the model can represent transactions or annual aggregates. A separate management hierarchy captures direct and indirect management. The team creates derived edges for rollups, such as a lower team remaining attributable to a higher-level manager. This allows one investment to be viewed through both a funding lens and a management lens.
People and shared entities connect systems that previously stood apart
People, organizations, and org charts are modeled as part of the same graph. People can own investments, attend meetings, direct work, and report through a team structure. These records historically lived in an HR source system, but they provide context for the wider organizational picture. The team stitches siloed systems together through common entities. Once those entities are connected, an agent can traverse relationships across the sources and reason about the organization as one structure instead of treating each source system as an isolated dataset.
Documents become useful to retrieval when their sections connect to graph entities
The platform combines structured records with unstructured documents. Meetings provide an entry point for documents, which are split into semantic sections or chunks and connected back to the organizational structure. Full-text indexes support search and retrieval, while the agent can also use graph retrieval directly. Phipps describes this document layer as an area with more work ahead, since the team had ingested only one document source at the time of the talk. The connection between document content and structured entities gives the agent organizational context for material that would otherwise remain isolated text.
MCP keeps the durable graph behind replaceable user experiences
The graph is exposed through one semantic layer and an MCP connection to agents. Phipps says the chat interface, user interface, and general agent interaction are not the durable parts, so the team serves the platform through tools users already have, including Claude and ChatGPT. The team forked Neo4j's off-the-shelf MCP server, modified its schema, and added state such as conversation IDs and message numbers. It is also building more constrained workflow experiences through Claude chat, MCP apps, and sandbox-based agents that use the same graph-backed platform.
Evaluation turns retrieval failures into changes to the data model
Phipps treats evaluation as part of data modeling. Questions from data owners are organized into complexity tiers and matched to reporting standards. Because structured data changes, each evaluation question has a graph query that runs against the live graph at evaluation time. The result is compared with the agent's answer. The team measures pass-at-one and stability, including whether repeated versions of the same question produce the same answer. Misses often expose ambiguity about user intent rather than a plainly wrong answer. The feedback loop updates the data model, domain rules, and schema descriptions.
"What was not defensible was the chat interface, was the UI, and even in some cases the general chat cases, the agent interaction."15:45
Who should watch
You are building an internal AI system and need to decide which part of the stack your team should own.
Your organization has siloed operational systems, complex reporting rules, and data owners whose knowledge is not written down.
You are evaluating graph or agent retrieval and want failures to improve the underlying model instead of only tuning prompts.
