Numbers from AI systems I actually run.

Production example: Alfred, a governed market graph

The first receipt is the graph my agent layer reasons over today: 46 knowledge cards, 154 nodes, 170 edges. The cards expand into typed nodes, and the nodes into relationships with sources and dates. Deliberately small, because the node count was never the hard part. Every node earned its place through a source I could check, every role change carries a date, and one node changing state updates every answer that touches it. That's the whole argument of the ontology essay, and the same pattern the Palantir Foundry ontology industrializes: you define the vocabulary, then you draw the graph, then the model gets a world to reason over.

The second receipt is a failure I keep public. For eleven days in March, two of my systems disagreed about where a key contact worked, and my agent prepared outreach on the stale answer. Nothing crashed. The graph exists because of that disagreement, and the governance rules exist so it cannot happen quietly again: every message is human-sent, every answer names its nodes, and an outreach ceiling is enforced in code. The full operating story is in the Alfred case study.

This site, measured like a system

The third receipt is the site you are reading. Before the ontology sprint began on 7 October 2026, Google served this domain roughly two impressions a day, all for my own name. As of 20 October 2026 the trailing 28 days show 190 impressions across 13 non-brand queries, with the first non-brand click landing on 13 October, six days into the sprint, on the Alfred page at position 5. One click is not a victory lap; it's a receipt. The ontology essay carries 49 impressions against zero clicks, which is exactly why its search listing was rewritten on 20 October. The numbers get published with the flaws attached, or they are marketing. None of this asks for trust: the essays are public, the pages are live, and the raw daily series behind the sprint receipt is published as CSV (28 days, updated monthly). The portfolio transactions and the graph contents stay private for client and prospect reasons, and I will walk any serious buyer through them on a call instead.

The market is not waiting for anyone. MarketsandMarkets sizes the knowledge graph market at $1.9 billion in 2026, heading toward $9.88 billion by 2032, and an IDC study from May 2026, vendor-sponsored and therefore read with that discount, reports production graph deployments cutting hallucination rates by an average of 44 percent. Gartner predicts 40 percent of enterprise applications will feature task-specific AI agents by the end of 2026, up from under 5 percent in 2025. Agents at that scale do not run on vibes; they run on shared vocabulary.

The frameworks, free to reuse

Three of the frameworks behind these numbers are written up so your team can run them without me:

• The 20-nouns starting point: write down the 20 nouns your business runs on, define 5 relations between them, and require every answer to name its nodes. No citation, no answer.
• The evidence rule: a fact without a source does not enter the graph, and a role change without a date is a rumor. Corrections are logged, not smeared over.
• The five governance questions to ask before any agent touches customers: what may it never do, what are its hard caps, which learned rules can you veto, what drift tripwire reverts it, and who owns it at 2am with receipts.

Grounding is the enterprise word for it. NIST's AI Risk Management Framework asks for exactly this under explainability and traceability: the structure has to exist first, and the model can only walk a path someone built.

If you cite these numbers, link this page or the essay that makes the case. That is the whole license.

Building agents that need to know what anything is?

I help enterprise teams design the semantic layer under their AI: ontologies, knowledge graphs, and grounded agents that can show their work. The receipts above are the operating proof, and the essays are the working notes.