T
Travis Thompson
Guest
Three competing philosophies for enterprise meaning, one seam where each of them gives way.
Key Takeaways:
TOC
Enterprise leaders evaluating AI agents converge, sooner or later, on a shared premise: an agent is only as reliable as the meaning it operates on. A table of numbers does not tell an agent what a customer is, what a hold status implies, or which records across five systems refer to the same entity. This premise has produced a rapidly maturing market for semantic layers: platforms variously described as ontologies, knowledge graphs, or context engines, depending on which part of the data stack a given vendor entered from.
Procurement teams evaluating this market tend to default to a familiar exercise: build a feature matrix, score each platform against it, and select the highest total. The matrix typically covers openness of the underlying standard, ease of adoption, integration with existing data estates, and cost of ownership. These are defensible criteria. They are also incomplete in a way that surfaces only during operation, well after a multi-year contract has been signed.
The choice enterprises face in this market is a dilemma. Calling it a competition only obscures that. Three distinct architectural philosophies compete for the same budget line, each defensible on its own terms.
Enterprise semantic layers can be grouped into three approaches, distinguished by where the semantic structure originates and who, or what, is responsible for maintaining it.
Each philosophy answers a legitimate organisational need, and none of them stands in for a named product.
Procurement frameworks in this category score platforms on criteria that a demonstration can settle: standard openness, deployment speed, lock-in exposure. A fourth criterion is missing from all of it: the distinction between a system that represents a rule and a system that enforces it once that rule crosses into a second system.
Consider a rule common to regulated industries: a payment above a defined threshold, once flagged with a compliance hold, requires sign-off from someone with legal authority before release. All three philosophies can represent this rule within their own boundary:
The rule holds until an agent hands the payment to a system outside that boundary, an increasingly ordinary pattern as organisations connect agents across departments and vendors.
The fact travels. The rule doesn't.
Grounding and governance get treated as one capability in most evaluation frameworks. They aren't.
Every platform surveyed here demonstrates strong grounding. None demonstrates governance that survives a boundary it didn't build: a seam no sales demonstration is built to visit.
Call this semantic theatre: a rule that looks enforced because it's visible, but only holds inside the one platform built to see it.
The absence of a governance criterion from standard evaluation frameworks follows from three structural conditions in the market, none of which requires attributing bad faith to any vendor.
A prospective buyer can observe a semantic layer answer a natural-language query in real time during a sales cycle. A cross-boundary governance lapse cannot be staged in the same setting, because it requires two independently built systems interacting under production conditions, typically months after a contract is signed. Procurement optimises for what can be observed in the room, and this criterion cannot be observed there.
Closing this gap requires formal constraint modeling: rules expressed in a form that can be checked mechanically rather than described in a policy document. Enterprise data organisations typically staff analysts, data engineers, and increasingly machine learning engineers. Few staff a role dedicated to authoring and validating formal business rules independent of any single vendor's platform. Buying decisions understandably do not account for a skill set the buying organisation has no internal capacity to evaluate.
A line item such as "supports automated semantic tagging: yes or no" is straightforward to score in a request for proposal. A criterion such as "enforces a constraint across two platforms neither was individually designed to reason across" resists that format. Few vendor sales engineering teams are equipped to answer it with specificity, and few buyer evaluation teams have experienced its absence with sufficient cost to have added it to the template.
Structural incentive keeps this criterion out of the room. Oversight has little to do with it.
The capability with the greatest bearing on organisational risk (governance that holds when an agent's work crosses into a second platform) has no seller among the three philosophies surveyed here. It is a capability an enterprise must construct independently, layered on top of whichever platform it selects, using formal methods that few data organisations are currently staffed to apply.
Choosing among the three philosophies is the tractable seventy per cent: every vendor has a commercial reason to help with that part. The remaining thirty per cent, governance that holds across a seam, has no commercial sponsor at all. Organisations commonly resolve that by leaving it unnamed, on the assumption the seam will hold because it always has.
That bet is invisible until the seam breaks.
That assumption holds up for some deployments. Constructing formal, portable governance carries a real cost, and for some agent deployments the exposure at stake may fall short of justifying that cost. The narrower point: that assumption should be a deliberate decision, with its risk stated to the relevant stakeholders, rather than inherited by default because no evaluation template asked the question.
The exposure created by this gap grows well past the point of purchase. Every additional agent an enterprise connects, and every additional platform boundary those agents cross, adds another seam where the same untested assumption has to hold. A single handoff between two systems is a manageable risk, even without formal testing behind it. A dozen handoffs, spread across procurement, finance, and customer operations, turn one small assumption into an enterprise-wide dependency that no single team owns or can see in full.
This is where an organization's ontology strategy gets tested in practice: in the months after rollout, once agents start handling more handoffs among themselves than with a human in the loop. The bet made at contract signing grows in scope with every agent added afterwards, and few organisations revisit it once the initial deployment is judged a success. A governance gap that was tolerable at one seam becomes a materially different risk once dozens of agents route work through it every day.
Short of committing to construct a formal governance layer (a substantial undertaking in its own right), enterprise buyers can extend their evaluation process with a small set of questions designed to surface this gap before a contract is signed rather than after deployment.
None of these questions has a comfortable answer today. Raising them before rollout, rather than after an incident, is the intended value of the exercise. A vendor able to respond to these questions with specificity has engaged with the seam directly. A vendor who answers with specifics has looked at the seam. A vendor who doesn't has told you something too.
No vendor-neutral standard has achieved broad adoption for closing this gap, and none is proposed here as a substitute. The formal methods that could express a rule once and enforce it everywhere an agent operates exist in research and in a handful of advanced deployments: the market simply hasn't made that discipline easy to acquire, staff, or evaluate.
Three legitimate philosophies now compete for the same budget. A governance gap sits beneath all three, and none of them currently offer a way to close it.
Key Takeaways:
- Enterprises evaluating semantic layers for AI agents are really choosing between three distinct architectural philosophies, not one category of product, and each carries a different characteristic blind spot.
- Standard procurement criteria such as standards-openness, ease of adoption, and lock-in exposure miss a fourth, more consequential dimension: whether a business rule survives when an agent hands work to a system outside the platform where the rule was defined.
- The gap is best understood as the difference between grounding telling an agent what something means and governance constraining what an agent may do about it everywhere it operates, a distinction none of the three philosophies fully resolve.
- The gap persists in procurement not because of vendor misrepresentation but because of three structural conditions: cross-boundary failures are hard to demonstrate in a sales cycle, the expertise needed to close the gap is scarce, and RFP scoring rewards checklist-friendly features over architectural completeness.
- Enterprise buyers can surface this gap before signing a contract by asking vendors to demonstrate rule enforcement across two integration points rather than one, and by identifying internally who owns the risk of a rule failing to survive a handoff.
TOC
- Three Architectural Philosophies for Attaching Meaning to Enterprise Data
- The Evaluation Criterion Missing from Standard Procurement Frameworks
- The Compliance Hold Left Behind at the Handoff
- The Line Between Grounding and Governance
- Three Reasons This Gap Survives Procurement
- The Cost of Buying Grounding Without Governance
- Questions to Raise Before the Contract Is Signed
- Conclusion: An Open Question for the Enterprise Ontology Community
Enterprise leaders evaluating AI agents converge, sooner or later, on a shared premise: an agent is only as reliable as the meaning it operates on. A table of numbers does not tell an agent what a customer is, what a hold status implies, or which records across five systems refer to the same entity. This premise has produced a rapidly maturing market for semantic layers: platforms variously described as ontologies, knowledge graphs, or context engines, depending on which part of the data stack a given vendor entered from.
Procurement teams evaluating this market tend to default to a familiar exercise: build a feature matrix, score each platform against it, and select the highest total. The matrix typically covers openness of the underlying standard, ease of adoption, integration with existing data estates, and cost of ownership. These are defensible criteria. They are also incomplete in a way that surfaces only during operation, well after a multi-year contract has been signed.
The choice enterprises face in this market is a dilemma. Calling it a competition only obscures that. Three distinct architectural philosophies compete for the same budget line, each defensible on its own terms.
Three Architectural Philosophies for Attaching Meaning to Enterprise Data
Enterprise semantic layers can be grouped into three approaches, distinguished by where the semantic structure originates and who, or what, is responsible for maintaining it.
Each philosophy answers a legitimate organisational need, and none of them stands in for a named product.
The Evaluation Criterion Missing from Standard Procurement Frameworks
Procurement frameworks in this category score platforms on criteria that a demonstration can settle: standard openness, deployment speed, lock-in exposure. A fourth criterion is missing from all of it: the distinction between a system that represents a rule and a system that enforces it once that rule crosses into a second system.
The Compliance Hold Left Behind at the Handoff
Consider a rule common to regulated industries: a payment above a defined threshold, once flagged with a compliance hold, requires sign-off from someone with legal authority before release. All three philosophies can represent this rule within their own boundary:
- An inferred-context system tags the payment entity as held.
- An evolved-contract system encodes the approval requirement into the business definition.
- A governed-action system models the release as an action type requiring authorisation.
The rule holds until an agent hands the payment to a system outside that boundary, an increasingly ordinary pattern as organisations connect agents across departments and vendors.
The fact travels. The rule doesn't.
The Line Between Grounding and Governance
Grounding and governance get treated as one capability in most evaluation frameworks. They aren't.
- Grounding: what a system tells an agent an entity means.
- Governance: what a system lets an agent do about it, everywhere the agent operates.
Every platform surveyed here demonstrates strong grounding. None demonstrates governance that survives a boundary it didn't build: a seam no sales demonstration is built to visit.
Call this semantic theatre: a rule that looks enforced because it's visible, but only holds inside the one platform built to see it.
Three Reasons This Gap Survives Procurement
The absence of a governance criterion from standard evaluation frameworks follows from three structural conditions in the market, none of which requires attributing bad faith to any vendor.
Demonstrability is asymmetric
A prospective buyer can observe a semantic layer answer a natural-language query in real time during a sales cycle. A cross-boundary governance lapse cannot be staged in the same setting, because it requires two independently built systems interacting under production conditions, typically months after a contract is signed. Procurement optimises for what can be observed in the room, and this criterion cannot be observed there.
The required expertise is scarce
Closing this gap requires formal constraint modeling: rules expressed in a form that can be checked mechanically rather than described in a policy document. Enterprise data organisations typically staff analysts, data engineers, and increasingly machine learning engineers. Few staff a role dedicated to authoring and validating formal business rules independent of any single vendor's platform. Buying decisions understandably do not account for a skill set the buying organisation has no internal capacity to evaluate.
Scoring rewards feature completeness over architectural completeness
A line item such as "supports automated semantic tagging: yes or no" is straightforward to score in a request for proposal. A criterion such as "enforces a constraint across two platforms neither was individually designed to reason across" resists that format. Few vendor sales engineering teams are equipped to answer it with specificity, and few buyer evaluation teams have experienced its absence with sufficient cost to have added it to the template.
Structural incentive keeps this criterion out of the room. Oversight has little to do with it.
The cost of buying grounding without governance
The capability with the greatest bearing on organisational risk (governance that holds when an agent's work crosses into a second platform) has no seller among the three philosophies surveyed here. It is a capability an enterprise must construct independently, layered on top of whichever platform it selects, using formal methods that few data organisations are currently staffed to apply.
Choosing among the three philosophies is the tractable seventy per cent: every vendor has a commercial reason to help with that part. The remaining thirty per cent, governance that holds across a seam, has no commercial sponsor at all. Organisations commonly resolve that by leaving it unnamed, on the assumption the seam will hold because it always has.
That bet is invisible until the seam breaks.
That assumption holds up for some deployments. Constructing formal, portable governance carries a real cost, and for some agent deployments the exposure at stake may fall short of justifying that cost. The narrower point: that assumption should be a deliberate decision, with its risk stated to the relevant stakeholders, rather than inherited by default because no evaluation template asked the question.
The exposure created by this gap grows well past the point of purchase. Every additional agent an enterprise connects, and every additional platform boundary those agents cross, adds another seam where the same untested assumption has to hold. A single handoff between two systems is a manageable risk, even without formal testing behind it. A dozen handoffs, spread across procurement, finance, and customer operations, turn one small assumption into an enterprise-wide dependency that no single team owns or can see in full.
This is where an organization's ontology strategy gets tested in practice: in the months after rollout, once agents start handling more handoffs among themselves than with a human in the loop. The bet made at contract signing grows in scope with every agent added afterwards, and few organisations revisit it once the initial deployment is judged a success. A governance gap that was tolerable at one seam becomes a materially different risk once dozens of agents route work through it every day.
Questions to Raise Before the Contract Is Signed
Short of committing to construct a formal governance layer (a substantial undertaking in its own right), enterprise buyers can extend their evaluation process with a small set of questions designed to surface this gap before a contract is signed rather than after deployment.
None of these questions has a comfortable answer today. Raising them before rollout, rather than after an incident, is the intended value of the exercise. A vendor able to respond to these questions with specificity has engaged with the seam directly. A vendor who answers with specifics has looked at the seam. A vendor who doesn't has told you something too.
Conclusion: An Open Question for the Enterprise Ontology Community
No vendor-neutral standard has achieved broad adoption for closing this gap, and none is proposed here as a substitute. The formal methods that could express a rule once and enforce it everywhere an agent operates exist in research and in a handful of advanced deployments: the market simply hasn't made that discipline easy to acquire, staff, or evaluate.
Three legitimate philosophies now compete for the same budget. A governance gap sits beneath all three, and none of them currently offer a way to close it.