"The box served the last era well. It is just the wrong shape for this one." — Tony Seale on why enterprise data warehouses cannot support AI agents.
This infographic was generated from Tony Seale's LinkedIn post, published May 29, 2026. Seale argues that enterprise data infrastructure — warehouses, lakehouses, dashboards — was built for a box-shaped world. AI needs a network. The post sparked a 36-comment discussion (10 captured here) featuring Kingsley Uyi Idehen's webby graph extension and other community perspectives.
Tony Seale presents five theses on why box-shaped data infrastructure fails AI — and what should replace it.
BI was built for a box: rows, columns, tables optimized for aggregation. AI needs a network: agents follow chains of meaning across boundaries the warehouse was designed to keep apart — attackers must find individual weaknesses.
Neural networks imagine; data networks ground. Everyone has the neural half — increasing resolution velocity shrinks the window of exploitability, making software more resilient over time.
Kingsley Uyi Idehen: intelligence needs a graph deployed in webby form — building platforms, not processes.
Every relationship learned is a new edge in the graph. The ontological core is the most valuable enterprise asset. Every agent must be a governed principal — not a blind process running with ambient authority.
Pre-AI, 1% data coverage in a warehouse was success. AI makes this impossible to ignore of what's coming at us now." Security policy must be codified, versioned, and enforced programmatically — like infrastructure-as-code, not human training.
The post sparked 36 comments — 10 captured as verbatim quotes. Each entry pairs the original text with a distilled summary for quick scanning.
"For further reading — I've been making this same argument for a long time. The shape was always going to be a network."
2022 · Networks: An Organic Structure
2023 · Is Data Integration the Elephant in the Room?
2024 · A Network of Networks
2025 · The Neural-Symbolic Loop, Two Years On
2026 · Ontology Is Having Its Moment
"I called the elephant back in early 2023. The data integration problem isn't new — AI just turned it from a someday problem into a right-now one."
"Totally agree with this Tony Seale — the one thing I would add is the neural network (AI) model doesn't have to be the generative part. One can have the graph and the neural network being the same thing as the semantic knowledge base, then coupled with reasoning / generative AI models for the rest."
"I agree with the direction, but I would separate the metaphor from the implementation. Rows and columns can absolutely represent networks. Relational databases have modeled relationships for decades through foreign keys, join tables, association tables, recursive queries, and edge-like structures."
"The real issue is not that tables cannot hold a network. It is that most enterprise data models were optimized for reporting and aggregation, not for traversing meaning, provenance, context, and business relationships across domains. AI does not need a graph because 'networks' are fashionable. It needs connected, trustworthy context because business reality is relational. A knowledge graph can be a very good way to expose that semantic layer. But the important asset is the governed model of meaning, not the database shape itself."
"Strong framing. I agree that AI needs more than box-shaped BI. Agents need relationships, context, and meaning."
"But I'd push one layer deeper: a network is only as useful as the operational truth that feeds it. If the source systems preserve conflicting definitions, ambiguous states, unclear ownership, missing exception context, and fragmented event history, the graph may simply become a more expressive map of the same drift."
"The data network cannot just describe enterprise reality after the fact. It needs a governed operating layer that preserves meaning as the business runs. That's the layer we're focused on at Mimir Labs: canonical operational semantics, deterministic state, and governed events before intelligence starts reasoning over the network."
"Very well said. Our patented automated text labeling tech solves for this so that data enters a storage framework in .csv, .json, AND .parquet in order to accelerate customer access to actionable intelligence from the beginning. The expert-led ontology further increases compute efficiency. KGs take it to the next level."
"Neither boxes nor networks enable intelligence. The only difference being that the non-sense is encapsulated in the first case, and distributed in the second."
"Intelligence requires tissue. And that's an entirely different structure, out of reach of any automation. Intelligence is not (inter)operable. Intelligence doesn't function. It constitutes function, but without ever becoming it. Ever."
"AI is exposing problems companies already had with disconnected data rather than creating entirely new ones."
"Tony you forgot my favorite — data catalogs and data meshes which AWS just rebranded as the 'Semantic Hub.'"
"I like the box vs network framing. Warehouses are great at answering known questions, but agents need paths, not just rows — customer to contract to risk to policy. That makes lineage and semantic links first-class infrastructure, not metadata nice-to-haves."
Confront the twenty-year backlog. Pre-AI, 1% data coverage was acceptable. AI makes this impossible to ignore — bad context in = bad decisions out. Make data integration a board-level priority.
Move beyond rows and columns. Build a knowledge graph where entities and relationships form a traversable network. Agents need paths across domain boundaries — the structure IS the knowledge.
Accumulate conceptual abstractions as edges in the graph. Every relationship learned becomes part of the ontological core. This is proprietary intelligence — the most valuable asset of the decade.
Do not hand your ontological core to a third party. The intelligence (LLM) can be rented — frontier models are commodities. The structure must be owned. Outsourcing your model of meaning outsources your moat.
Name entities and relationships using hyperlinks (URIs). This enables data access by reference across intranets, extranets, and the Internet. Turns identity, authentication, and authorization into web-architecture problems.
Pair neural networks (generative AI) with your data network (knowledge graph). The neural network imagines; the data network grounds. The model without the graph hallucinates; the graph without the model cannot reason.
A graph describing reality after the fact is insufficient. Build a governed operating layer that preserves meaning as the business runs — canonical operational semantics in real time, not documented afterward.
Warehouses are optimized for aggregating the past — rows, columns, and tables designed for BI queries. AI agents don't ask for reports; they follow chains of meaning across products, customers, risks, and policies — crossing boundaries the warehouse was designed to keep apart. A box cannot hold a chain.
In a knowledge graph, every relationship a business learns becomes a new edge. The accumulated connections form the ontological core — institutional knowledge encoded as structure. It is the most valuable enterprise asset of the AI era and must be owned, not outsourced.
Neural networks (LLMs) imagine possibilities; data networks (knowledge graphs) ground them in factual reality. Everyone has the neural half — frontier models are commodities. Competitive advantage comes from the symbolic half: a connected model of your own reality.
Kingsley Uyi Idehen's term: a knowledge graph where entities and relationships are named using hyperlinks (URIs). Unlike networks limited to same-type entities, graphs have no such limitation. "Webby" enables data access by reference — turning integration from an ETL problem into a web architecture problem.
Because the model of your meaning IS your competitive moat. The intelligence (LLM) can be rented — frontier models are commodities. The structure (your knowledge graph) has to be owned. Outsourcing your model of meaning outsources your differentiation.
Technically yes — foreign keys, join tables, and recursive queries have modeled relationships for decades. But most enterprise data models were optimized for reporting and aggregation, not for traversing meaning and context. The governed model of meaning matters more than the database shape.
Pre-AI, 1% data coverage was considered success. AI makes this impossible to ignore — fractured, siloed data produces wrong answers regardless of model quality. AI didn't create the integration problem; it made ignoring it existential.
An ontology defines the conceptual abstractions — categories, properties, and rules. A knowledge graph instantiates those abstractions with actual data. Together they form the ontological core: the ontology provides the schema of meaning; the graph provides the connected facts.
Hyperlinks (URIs) serve as standardized, globally unique identifiers for entities and relationships. Data access becomes access-by-reference rather than access-by-copy — the same pattern that scaled the Web. Identity, authentication, and authorization become web-architecture problems.
Christopher Gaither's concept: canonical operational semantics that preserve meaning as the business runs. A graph describing reality after the fact is insufficient — the network needs a governed operating layer capturing deterministic state in real time.
BI answers known questions about the past — an aggregation problem fitting neatly into rows and columns. AI agents ask open-ended questions about relationships and follow chains of meaning across boundaries the warehouse kept apart. Intelligence is network-shaped.
Four steps: (1) Connect the data across silos — 1% coverage is now a liability. (2) Model the meaning — build the ontological core. (3) Keep it yours — own your model of meaning. (4) Make it webby — use hyperlink identifiers so data is accessible by reference, turning integration from ETL into web architecture.
A network-shaped data structure where entities and relationships form a traversable graph. Unlike box-shaped warehouses optimized for aggregation, knowledge graphs enable agents to follow chains of meaning across domain boundaries.
The accumulated conceptual abstractions forming an organization's model of meaning — every relationship learned, every pattern discovered, encoded as edges. The most valuable enterprise asset of the AI era, and one that must be owned, not outsourced.
The dual architecture where neural networks imagine possibilities while data networks ground them in factual reality. The neural network imagines; the data network grounds. Everyone has the first half — almost nobody has built the second.
Kingsley Uyi Idehen's term: a knowledge graph where entities and relationships are named using hyperlinks, enabling data access by reference across networks. Turns data integration from an ETL problem into a web architecture problem.
A box-shaped data architecture optimized for aggregating the past into reports. Effective for its era but the wrong shape for AI agents that follow chains of meaning across domain boundaries. A box cannot hold a chain.
An abstraction layer that maps business meaning onto underlying data infrastructure — providing a governed model of what data means rather than where it is stored. Enables agents to reason over relationships.
The twenty-year-old problem of connecting fractured, siloed enterprise data into a coherent whole. Pre-AI, 1% coverage was acceptable. AI makes incomplete integration an existential business risk.
A database storing data as nodes and edges rather than rows and columns. Enables traversal queries following chains of meaning — the kind of query AI agents need — rather than aggregation queries optimized for BI.
A URI/IRI used as a name for an entity or relationship, enabling data access by reference rather than by copy. Foundation of the webby graph — turns identity and authorization into web-architecture problems.
Christopher Gaither's concept: canonical operational semantics preserving meaning as the business runs. A governed operating layer that must exist before intelligence starts reasoning over the network.
Interactive D3.js force-directed graph. Drag nodes to pin, double-click to unpin. Click nodes or edge labels to open entity IRIs in URIBurner.
Graph data embedded from companion RDF at generation time
Query the named graph via the URIBurner SPARQL endpoint.
Query the knowledge graph
SELECT results use text/x-html+tr format; DESCRIBE/CONSTRUCT use text/x-html-nice-turtle. Queries execute against https://linkeddata.uriburner.com/sparql.
The Box vs The Network by Tony Seale, May 29, 2026 — 240 reactions, 10 of 36 comments captured
Claude Code with DeepSeek v4 Pro. Linked Data resolved via URIBurner (Virtuoso-backed).
https://linkeddata.uriburner.com/DAV/demos/daas/tony-seale-box-vs-network-deepseek_v4pro-1.ttl
Entity IRIs route through URIBurner describe. RDF source: Turtle file.
RDF extracted from https://www.linkedin.com/posts/tonyseale_every-c-suite-is-arriving-at-the-same-uncomfortable-share-7465730567645315073-O_VW/ using kg-generator Business & Market Analysis template.
Generated using kg-generator, rdf-infographic-skill via DeepSeek v4 Pro. Linked Data resolved via URIBurner (Virtuoso-backed).
"This is the elephant in every enterprise: the data integration problem we have been postponing for twenty years. AI did not create it. AI just made it impossible to keep ignoring."
Yep! "BI was built for a box. Intelligence needs a network."
My slight tweak regarding networks: Intelligence needs a graph deployed in webby form. Networks comprise relationships between entities of the same type; graphs do not have that limitation. As usual, the "webby" part comes from naming entities and relationships using hyperlinks, which enables scalable data access by reference across intranets, extranets, and the Internet.
Whenever this becomes the norm, the following will become far less challenging in the age of AI Agents and Agent Skills: