The problem

Most enterprise AI failures don’t come from weak models. They come from the fact that “customer,” “inventory,” and “capacity” mean three different things across a company’s ERP, CRM, MES, and WMS systems, and no one has ever forced those definitions to agree. A model can forecast a stockout perfectly and still be useless if nobody can say who’s allowed to approve the replenishment order, whether the data it used was current, or what happens the next ten times a human overrides its recommendation.

This is a problem any enterprise recognizes once you name it: dashboards and copilots multiply, but decisions still route through email and tribal knowledge because the systems don’t share a business vocabulary the AI can act inside of. I built this project to work out why that gap exists and what closes it.

The approach

This is a research and methodology project, not shipped software yet — I’ve been building out a framework called Lean Dynamic Ontology (extending Palantir’s Foundry ontology model with a lean-manufacturing delivery discipline) that gives a name and a checklist to the gap between “we have a model” and “the business actually acts on it.”

The core move is a four-level hierarchy: a data model describes what’s in your systems, a knowledge graph adds relationships between facts, an executable ontology adds actions and state (a machine isn’t just “a row in a table,” it’s “running / derated / stopped,” with rules about who may act on it), and Lean Dynamic Ontology adds the requirement that every one of those actions produce traceable, verified value — otherwise you’ve automated waste faster, not fixed anything.

The most concrete, reusable piece is a diagnostic checklist: seven forms of “semantic waste,” mapped directly from lean manufacturing’s seven wastes onto AI system design — modeling concepts nobody owns, cataloging relationships no agent ever calls, every project re-defining “customer” from scratch, decisions that stall because permissions conflict, facts used past their valid time window, predictions with no write-back path into the source system, and expert overrides that vanish instead of feeding the next model update. Instead of “our model needs more data,” this gives you a specific failure mode to point at. I also worked through a full concrete example — a steel mill replenishment decision — mapping out the objects, events, state, decision logic, governed actions, and evidence trail a real “minimum sufficient ontology slice” needs, plus a 12-week path for shipping one such slice end to end rather than trying to model an entire enterprise up front.

What I learned

The two most useful ideas came from applying the lean lens rigorously rather than treating it as a metaphor. First, “action discontinuity” — a model recommends something but there’s no write-back, approval, or rollback path wired to it — turned out to be the most common and most invisible failure: teams ship a dashboard, call it done, and never build the plumbing that turns a recommendation into a governed action. Second, “loss of field experience” — when an expert overrides a recommendation and that override isn’t captured — means organizations pay to relearn the same lesson repeatedly instead of it ever reaching the model. Neither of those shows up if you’re only asking “is the model accurate.” You only see them once you ask “does this loop actually close.”

Where this could go

The natural next step is turning the checklist and the ontology-slice template into something usable — a lightweight audit tool that scores an AI initiative against the seven waste types before more budget goes into it, and a repeatable intake process (value objective, decision moment, the ten required elements of a slice) for scoping new AI work honestly instead of starting from “let’s build a dashboard.”

At team or org scale this is the same problem enterprises pay consultancies to diagnose: why does agentic AI stall past the pilot stage. The seven-waste audit and the 12-week minimum-slice path are structured enough to run as an actual engagement format — a way to walk into a stalled AI initiative, name which of the seven waste types is actually blocking it, and scope the smallest slice of shared business meaning that proves the loop can close before committing to modeling the whole enterprise.

Takeaway

Foundation models are commoditizing fast; the thing that doesn’t commoditize is an organization’s accumulated, agreed-upon definition of its own business — and that’s the actual battleground for who wins with agentic AI.