Every data platform in the enterprise market is currently claiming agent-readiness. The announcements are engineering-credible. The investments are real. The roadmaps are detailed. And without a precise standard against which to evaluate them, every one of those claims is unfalsifiable — which means every one of them, however generous or however aggressive, is equally valid.
That is not a useful state for enterprise architects making platform bets they cannot easily unwind.
The insight-to-decision shift that Gap 3 named is the dividing line. Data infrastructure built for insight was optimized to answer human questions: what is happening, what does it mean, what should we consider? Decision-grade substrate must do something fundamentally different. It must be consumable by autonomous actors that cannot ask clarifying questions, cannot wait for a schema briefing, and cannot rely on a human to assemble the meaning of what they retrieve downstream.
That difference is architectural. It is not a matter of latency, throughput, or model integration. It is a matter of what the data layer was built to do at its foundation — and what that optimization made possible or foreclosed.
The Falsifiability Problem
The core problem with agent-readiness claims is not that they are false. It is that they are not falsifiable without a standard. A platform that enables agents to query structured data, retrieve documents, and execute multi-step workflows is agent-capable in a meaningful engineering sense. The question the Substrate Fitness Criteria answer is not whether agents can operate on the platform. It is whether the platform’s data layer was built for the kind of operation that agentic AI requires at production scale — autonomous actors making consequential decisions without human intermediation.
That distinction does not show up in architecture diagrams. It shows up in how the platform handles the five conditions under which insight-grade infrastructure breaks down when autonomous agents replace human analysts as the primary consumer of the data layer.
Deloitte’s 2025 survey: nearly half of organizations cite searchability of data as an active challenge to their AI automation strategy — with 47% citing reusability. The problem is not that data is unavailable. It is that the data layer was not built to be consumed by actors other than the human analysts the platform was designed to serve.
The Substrate Fitness Criteria
Most enterprise data platforms have built what this series calls a cognitive substrate — infrastructure optimized for human interpretation, where the terminal output is a dashboard, a report, or a query result, and a human remains in the decision loop. The Substrate Fitness Criteria do not evaluate how well a platform performs as a cognitive substrate. They evaluate whether it can perform as something categorically different.
What the Criteria Establish
The Substrate Fitness Criteria are five architectural requirements that define decision-grade data substrate. A platform that meets all five has a data layer built for autonomous actors making consequential decisions. A platform that meets some has made structural progress. A platform that meets none — regardless of its agent deployment capabilities — is operating on insight-grade infrastructure with an agent layer positioned on top.
The criteria are not a feature checklist. They are architectural tests. Each one traces directly to the insight/decision distinction Gap 3 established. Each one is falsifiable. Together they form the evaluative framework that Posts 2 and 3 apply to what the market has actually built.
Criterion 1 — Discoverability by Autonomous Actors
Insight infrastructure is queryable. An analyst formulates a question, issues a query, and retrieves a result. The analyst knows what they are looking for. They can ask clarifying questions, refine their approach, and interpret partial results. The infrastructure does not need to be self-describing because the human querying it brings the schema knowledge.
Decision-grade substrate must be traversable — navigable by an agent that arrives without pre-loaded schema knowledge, cannot ask what tables exist, and cannot rely on a data dictionary maintained separately from the data itself. The substrate must surface what is available, what it means, and what it is connected to, in a form an autonomous actor can consume without external scaffolding assembled per use case.
The Deloitte finding that 48% of organizations cite searchability as an active challenge names this gap directly. Searchability in the insight context means analysts can find what they need. Searchability in the decision-grade context means agents can discover what is relevant without a human mediating the lookup. Those are not the same requirement and they are not resolved by the same architecture.
The Test
A substrate is discoverable by autonomous actors when its semantic structure is intrinsic — when the meaning, scope, and relationships of data elements are encoded in the substrate itself, not assembled by the application layer or the human analyst consuming it.
Criterion 2 — Contextual Richness at the Point of Consumption
Insight infrastructure delivers data values. A dashboard displays a number. A report surfaces a trend. The human reviewing it brings the interpretive context: what this number means in the current operational moment, how confident to be in it, what recent events might have affected it. That assembly is the analyst’s job — and insight infrastructure was built assuming the analyst would do it.
Decision-grade substrate must deliver contextual richness at the point of consumption — not as a downstream assembly step but as context that is atomically retrievable and decision-bound at the moment of consumption. An agent making a consequential decision needs to know not just what a value is, but when it was last updated, what confidence level applies, what operational state it reflects, and what caveats affect its interpretation. If any of those elements require a human to assemble — or require the agent to retrieve them separately from the data value itself — the substrate is insight-grade regardless of how capable the agent consuming it is.
This is the criterion that most clearly reveals why the ELT paradigm — despite its genuine architectural improvements over classical ETL — does not resolve the substrate challenge. ELT retains more of the original signal. Schema-on-read provides more flexibility. Neither change encodes contextual richness intrinsically. The assembly still happens downstream. In insight infrastructure, a human analyst does that assembly. In agentic systems, it falls to the agent — or it does not happen at all.
Criterion 3 — Action-Orientation
Insight infrastructure surfaces what is true. Its outputs answer questions: what is the state of the system, what changed, what is the trend. A human takes those answers and decides what to do. The infrastructure’s job ends at the delivery of the insight.
Decision-grade substrate exposes what is actionable. Operational state must be first-class and transactionally bound to data access — not stored separately and retrieved downstream. An agent consuming data from a decision-grade substrate encounters, in the same operation, what is true, what is permitted, and what the consequences of action are. The substrate does not merely answer questions — it enables decisions.
Insight infrastructure tells the agent what is true. Decision-grade substrate tells the agent what is true, what is permitted, and what happens next. The distance between those two deliverables is an architectural gap — not a feature gap.
This is the criterion that most clearly distinguishes genuine substrate from sophisticated insight infrastructure — and the criterion where the gap between cognitive and agentic substrate is architectural rather than additive. An agent that retrieves accurate, contextually rich data and then cannot determine what it is authorized to do with that data is operating on insight-grade infrastructure. Operational state that lives outside the data access transaction cannot be made transactionally bound through tooling or orchestration above the data layer. This is not a missing feature. It is a design orientation that would have to be reversed at the foundation.
Criterion 4 — Permission-Native Architecture
Authorization is the criterion where insight infrastructure most visibly breaks down under agentic load. In insight architectures, authorization is typically enforced at the application layer: the dashboard shows only what the user is permitted to see, the API returns only what the calling application is authorized to retrieve. The data layer itself carries no permission state. The controls are external.
Decision-grade substrate requires permission-native architecture — authorization baked into the substrate layer itself, not enforced by the application layer above it. When an autonomous agent traverses a substrate, it must encounter permission boundaries intrinsically: data elements it is not authorized to access must be inaccessible at the substrate level, not filtered by a downstream control that the agent must trust to operate correctly.
The Alignment Gate Connection
This criterion connects directly to the Harness Layer’s alignment-gate function. The harness can only enforce what the substrate makes enforceable. A permission-native substrate enables the harness to function as a genuine governance mechanism. A substrate relying on application-layer controls forces the harness to compensate for a structural gap it was not designed to fill. Permission-native architecture is not only a substrate requirement — it is a precondition for alignment-grade governance.
Criterion 5 — Provenance and Auditability by Design
Insight infrastructure is auditable in the retrospective sense. An analyst can trace back to the query that produced a result, the pipeline that delivered the data, the source that originated it. That audit trail exists to support human review — it is activated after a decision is made, by a human who wants to understand how the result was produced.
Decision-grade substrate requires provenance and auditability by design — structural, not retrospective. When an autonomous agent makes a consequential decision, the substrate must record what the agent consumed, what state that data reflected at the precise moment of consumption, why the agent was authorized to consume it, and what the chain of custody was from source to decision. Post-hoc reconstruction is not sufficient. The audit trail must be continuous, automated, and intrinsic to the substrate — not assembled after the fact from logs and pipeline metadata.
This criterion is what makes agentic AI governable at enterprise scale. Without provenance by design, the organization cannot answer the question that regulators, risk committees, and auditors will eventually ask: on what data, in what state, with what authorization, did this agent make this decision? A substrate that cannot answer that question structurally forces the governance function upstream — where it is more expensive, less reliable, and insufficient for high-velocity agentic systems.
Insight Infrastructure vs. Decision-Grade Substrate
Built for Human Analysts
Queryable by humans who bring schema knowledge. Contextual assembly happens downstream — the analyst interprets. Auditable on demand when a human initiates review. Authorization enforced at the application layer above the data.
Excellent for its intended purpose. The wrong load-bearing architecture for autonomous decision-making.
Insight-GradeBuilt for Autonomous Actors
Traversable by agents without pre-loaded schema knowledge. Contextual richness intrinsic to the data — meaning is encoded, not assembled. Provenance continuous by design. Authorization permission-native at the substrate layer.
Actionable: exposes not just what is true but what is permitted and what is possible.
Substrate-GradeThe Substrate Fitness Criteria establish the standard that makes the central question of this series answerable. The question is not whether data platforms can support agent deployment. Most can. The question is whether the data layer beneath the agent was built for the load that autonomous decision-making places on it — or whether it is being asked to bear a load it was never designed to carry.
