Ten Years of Scar Tissue

Every enterprise that has tried to build a knowledge graph has a decade like this on file somewhere: a semantic layer that made sense in year one, three re-orgs and two acquisitions later held together by exceptions nobody remembers writing.

The industry calls this “ontology scar tissue” — semantic drift, the 80/20 integration trap, maintenance debt, and a chronic inability to get stakeholders to agree on what a word means. These are treated as engineering folklore, war stories traded at conferences. They are not folklore. They are the empirical record of what happens when an organization tries to build a decision-grade semantic layer without a discipline for maintaining it — and the record now runs through consultancy research, peer-reviewed systems literature, and one company’s two-decade commercial specimen, not just anecdote.

The four scars, reread

Semantic drift is not entropy. It is the absence of a CTX discipline — one of the five Substrate Fitness Criteria Luminity uses to test agentic readiness. McKinsey’s own research on the data-driven enterprise states the condition plainly: in most organizations, data “often has no true owner.” A schema without an owner of record will drift, by construction, the moment two departments touch it independently.

The 80/20 trap — where the last fifth of the integration work eats four-fifths of the budget — is what happens when coordination-grade patching is used to solve an alignment-grade problem. BCG’s research on enterprise data architecture found integration and people costs, including third-party system-integrator and consulting spend, on pace to double within five years, driven by data complexity rather than ambition. The last 20% does not compress. It requires someone with standing to make a judgment call and be accountable for it.

Maintenance debt is Governance Displacement in its native habitat: the mapping layer still exists, the documentation still exists, but neither reaches the surface where decisions actually get made. Bain’s own architecture guidance for agentic AI makes the same point from the vendor-advisory side: without a unified schema and data-contract layer, every new agent deployment compounds what Bain calls integration debt. Nearly a decade earlier, UC Berkeley’s RISELab argued the underlying cause directly — that context and metadata have to be a first-class system service, not something reconstructed after the fact. Most enterprises are still treating it as an afterthought.

Lack of consensus is a symptom, not a root cause. It is what a centralized, single-schema design produces the moment an organization grows past the size where one team can hold the whole model in its head. Decentralization does not create the disagreement — it exposes disagreement a rigid schema was suppressing by fiat.

None of this is new, and none of it is specific to AI. MIT’s Michael Stonebraker, whose data-integration research spans several decades, puts a number on the baseline condition: roughly 15% of data in a typical enterprise repository is missing or wrong, and scalable integration has remained the field’s central unsolved problem despite what he calls “the long history of the problem in both academia and industry.” Four different scars, and across the consulting side and the research literature alike, accountability repeatedly emerges as the common structural failure: nobody was accountable, in production, for keeping the semantic layer true to the business it described.

The specimen: what a decade — or three — forward-deployed actually costs

Two institutions have spent longer than a decade proving how expensive that accountability actually is — one academic, one commercial.

Stanford’s Center for Biomedical Informatics Research has run the field’s leading ontology-authoring project, Protégé, since the early 1990s. Three decades of the best-resourced, most methodologically rigorous ontology tooling effort in academia, and the field still has not made ontology maintenance cheap or automatic. That is the first piece of evidence that scar tissue is a property of the work itself, not a symptom of underinvestment.

Palantir has spent close to two decades industrializing the same accountability on the commercial side. The practice Shyam Sankar named forward-deployed engineering sends software engineers — not consultants, not solutions architects — to sit inside a customer’s operations and map the customer’s actual data into a working Ontology, Palantir’s term for the governed semantic layer its Foundry platform runs on. The distinguishing rule practitioners cite: if the FDE hasn’t shipped something in the last two weeks, they’re not doing FDE — they’re doing sales engineering in a different jacket.

This is expensive, slow to scale, and was mocked for most of its history as “just consultants” or “not a real software company.” Sankar’s own account of why Palantir kept doing it anyway is direct: the company refused the alternative of building software and then, in his words, “throwing our software over the wall for consultants to implement.” His broader argument is that most enterprise software fails not at the model or the platform but at what he calls the second 80% — the unglamorous work of making a system true to how a specific business actually runs — and that industry has a long history of outsourcing that work to people with no stake in the outcome.

Karp’s framing of the same discipline, stated more recently and more combatively, is that the ontology and the deployment layer — not the model — are where enterprise AI value actually accrues, a position he has been repeating on financial television essentially weekly this year. He has described the alternative, where enterprises rent intelligence from a model provider without owning the semantic layer underneath it, as a kind of “commodity cognition” — his term for the belief that once every model is roughly as capable as every other, the durable advantage moves to whoever owns the layer that makes the model specific to your business. Whatever one makes of Karp’s delivery, the underlying claim tracks the four scars precisely, and now tracks McKinsey, BCG, Bain, Berkeley, MIT, and Stanford as well: an ontology nobody is accountable for building and maintaining will not become an asset just because a more capable model sits on top of it.

The warning inside the playbook

None of this means the FDE model is a template every enterprise should copy directly. It means the opposite of what the current market is doing with it. “Forward deployed engineer” has become the fastest-growing job title in AI, and independent accounts of the space are already flagging the gap between the label and the discipline it originally named. A professional-services team renamed FDE without the ownership, the outcome accountability, or the standing to say no to a bad integration is not adopting Palantir’s playbook. It is wearing its jacket. Every enterprise now hiring “FDEs” at scale is making a bet about which of those two things they’re actually buying — and most of the industry has not yet had to find out the hard way. Even the analyst layer nominally positioned to arbitrate this is compromised: most of what circulates as “Gartner says” about ontology and knowledge-graph tooling is vendors citing their own Magic Quadrant or Cool Vendor placement, not an independent Gartner position on the discipline itself.

That gap — between the FDE label and the FDE discipline — is where Post 2 of this series starts, applied to the modern tools claiming to solve the same four scars without any of the accountability that made the original discipline work.