Post

The bottleneck in enterprise AI has moved. It was never the model. It was always context. Every serious AI conversation is now turning the same way: toward semantic layers, ontologies, enterprise memory. But there is a harder question hiding inside that turn.

When organisations connect their context, can they still own it, move it, and interoperate with it? Because connecting context is not enough. A context layer that cannot interoperate may look like acceleration now. Later, it becomes the wall that keeps you out.

Not a thousand disconnected copilots. Not a tabular warehouse with some semantics sprinkled on top. Not one vendor's proprietary 'Context Graph'.

The organisations that win Phase 3 will not simply be the ones with tidy internal data. They will be the ones whose meaning can cross the border — into a market, a partner, a regulator — without asking a vendor's permission first. The ones whose context cannot follow will watch deals close on the other side of a wall they paid to build.

So use the platforms. Buy the tooling. Let the vendors accelerate the work. But before you hand over your context, ask three questions…

If the answer to any of them is no, you do not own your meaning. You are renting it.

View original post on LinkedIn ↗

The Three-Phase Framework

Phase 1 — Past
Web → Model
Thirty years of open, crawlable, addressable Web compressed into model weights. Language, code, knowledge, and culture folded into the model.
Phase 2 — Present
Enterprise → Context
The enterprise folding itself into context. Giving things stable identities, making relationships first-class, formalising the meaning the business actually runs on.
Phase 3 — Future
Context → Agentic Web
Context unfolding back out into the Agentic Web — where agents cross org boundaries, only travelling where meaning is open enough to follow.

The Interoperability Test

Before you hand over your context, ask:

Comments (10 of 12 visible)

FB
François BENSID 👍 2 2h ago

Everyone is racing to give their AI better context. Few are asking who owns that context once the AI depends on it. Most organisations are renting it. The rent is invisible until you try to leave. They only discover the lease terms when they ask for the keys back.

What we've learned running this in production: the integration story only works when meaning can resolve outside the platform boundary. Open identifiers aren't an option, they're the prerequisite.

CD
Chris Day 👍 1 2h ago

Curious as to why no mention of D-PROD to help alignment of the right context with the right dataset. I deal with some truly humongous ontologies that become minuscule in an applied linguistic context.

Excellent insight! Based on your taxonomy, I believe phase 3 is which can finally help human to untangle many wicked problems at scale, caused by linear local optimization based on era of Taylorism like climate change and nature loss, and replace existing mess with intangible, carbon neutral circular economy.

This is also enormous opportunity to European industries, as our R&D have been building industrial standards based on semantic web thinking already over a decade (e.g. OPC UA Companion Specs, Asset Administration Shell, Digital Product Passport etc) — I call them 'practical domain-specific industrial ontology implementations', ready to be applied.

Great post Tony. The way I've been thinking about this concept is: Each company builds their context graph. They connect it to agents. Agents eventually become first class citizens in their business and can actually be given the freedom to act on your behalf across the digital landscape.

This forms the agent to agent economy where: Agents are social networking with other agents on behalf of companies — connecting with other agents, progressive federation and disclosure of information upon mutual interests, creating connections across context layers and taking actions on deals, partnerships, introductions, etc. Agents are given business rules, reporting structures, and guardrails on their autonomy. That's how we get to the agentic economy.

RF
Rémy Fannader 👍 1 37m ago

'Context' is so vague as being completely irrelevant: it can refer to user context, conversation, domain, business model, corporate organization, etc. While every one is relevant, no agentic cognition or collaboration is possible without explicit distinctions.

Identifier resolution is where this breaks in practice. Organizations build internally coherent ontologies but anchor everything to vendor-assigned IDs that mean nothing outside the platform. When an agent tries to cross an org boundary, there's nothing for another system to resolve against.

The semantic layer looks complete from inside and is effectively opaque from outside. The three questions at the end are the right test for exactly this: can your identifiers be resolved externally, can your ontology leave, can your graph federate. Most enterprise context layers fail all three without the teams that built them knowing it.

Thank you Tony, again a very meaningful and timely post. Describing very well where industries are heading with AI. A path that GVW is not only following, but even leading. At least for the 'Human centric, data driven, AI enabled value chain'. Partly because what we do every day with patient data and drug product data, where we now aid with point AI solutions. Mostly though through our participation in the various standards bodies. With ISO Health Informatics being the top of the pyramid. Contributing for example to ClinOps standardization and Identification of Medicinal products.

True open standards make your data expenses compound. It's the data systems. There are no data systems that we know of that have the performance and cost profile to scale semantic knowledge graphs as a primary data storage engine. But it has to be primary. Because you can't really have decisions made from derivative knowledge.

These ideas are all really nice, but at real scale your data systems run too expensive. That's why I built a labeled property graph that uses ~1/10th the memory of industry leading data systems. We can't keep going or it will break us.

Strongly agree. In pharma, the challenge isn't just building a knowledge graph — it's ensuring that meaning can flow across clinical, regulatory, safety, and partner ecosystems without being tied to a single vendor's platform. Otherwise, today's 'talk-to-your-data' APIs (being provided by more and more vendors) may easily become tomorrow's interoperability bottleneck.

Source

LinkedIn Post
linkedin.com/posts/tonyseale_the-bottleneck-in-enterprise-ai-has-moved-activity-7471109073430573056-8sPG ↗
🔗 Lookup post entity

Frequently Asked Questions

Enterprise AI context interoperability is the ability for AI agents, copilots, and models to share, understand, and build upon structured meaning across organisational and platform boundaries — without being locked into a single vendor's semantic layer. As Tony Seale argues, the bottleneck in enterprise AI has shifted from model capability to the portability of the context those models consume.
Foundation models have reached a level of capability where the quality of AI outcomes depends less on the model itself and more on the richness and portability of the contextual knowledge it can access. If that context is locked inside a proprietary system — unable to move, share, or be resolved by another agent — the model's capability cannot be applied where it matters most.
When an organisation stores its ontologies, entity definitions, and semantic relationships inside a closed AI platform, it does not truly own that meaning — it rents it. If the vendor changes terms, discontinues the product, or builds a walled garden, the organisation's structured knowledge cannot escape. Tony Seale and Kingsley Idehen identify this as a fundamental risk of the Phase 1 / Phase 2 enterprise AI architecture.
A copilot is a user-facing AI assistant embedded in a tool. A context layer is the underlying structured knowledge graph that supplies meaning to one or more agents. The post argues that organisations need to invest not just in copilots (Phase 1 / Phase 2) but in a portable, open context layer (Phase 3) that persists independently of any single copilot or vendor.
The Agentic Web is an emerging layer of the Web where software agents — not just humans — act as first-class participants: discovering, querying, reasoning over, and producing structured data. For agents to interoperate meaningfully on this web, they must share resolvable identifiers and portable ontologies. This is what Phase 3 of Tony Seale's framework enables.
When every entity in a knowledge graph is identified by a dereferenceable URI, any agent — in any organisation — can follow that link to retrieve a machine-readable description of that entity. This allows AI systems to join graphs, resolve ambiguities, and collaborate across trust boundaries without pre-negotiated proprietary integrations. As Kingsley Idehen notes, hyperlinks as stable standardised identifiers are the key.
Phase 3 is essentially the Semantic Web realised at enterprise scale for AI consumption. Technologies like RDF, OWL, and SPARQL provide the standards for portable, resolvable, globally joinable knowledge — exactly the three tests Tony Seale poses in the post.
Context portability means that the structured knowledge an AI agent relies on can be exported, transferred, and re-loaded into a different system without loss of meaning. It matters because organisations that lack it are trapped: their AI investments are tied to a vendor's platform, their data cannot join the global knowledge graph, and their agents cannot collaborate with peers outside the platform's walls.

Glossary

Context Interoperability

The capacity for AI systems from different vendors or organisations to exchange and act upon structured semantic context — ontologies, entity definitions, and relationships — without proprietary translation layers.

Ontology Portability

The ability to export, transport, and re-import an ontology — a formal specification of concepts and relationships — across different AI systems or knowledge graph platforms, retaining full fidelity of meaning.

Dereferenceable Identifier (IRI/URI)

A web address that, when resolved, returns a machine-readable description of the entity it names. The foundation of Linked Data and Phase 3 context layers — any agent can follow an IRI to discover what an entity is and how it relates to others.

Semantic Web

A W3C-defined extension of the Web where data is structured with explicit meaning, using standards such as RDF, OWL, and SPARQL, enabling machines to reason over and connect information across domains and organisations.

Context Layer

A persistent, structured knowledge graph that supplies semantic context to one or more AI agents or copilots. Unlike a copilot (which is user-facing), the context layer is infrastructure — it must be portable and open to support Phase 3 interoperability.

RDF (Resource Description Framework)

A W3C standard data model for representing structured information as subject–predicate–object triples, enabling knowledge to be expressed in a format that is both machine-readable and globally linkable.

SPARQL

The standard query language for RDF knowledge graphs, analogous to SQL for relational databases. SPARQL queries can federate across multiple distributed graph endpoints, enabling the global-graph joinability that Phase 3 requires.

Context Lock-in

A state in which an organisation's structured AI context — its ontologies, entity relationships, and semantic annotations — is trapped inside a proprietary platform and cannot be exported, resolved, or joined with external graphs.

OWL (Web Ontology Language)

A W3C ontology language built on RDF, used to formally define classes, properties, and logical constraints in a knowledge graph. OWL ontologies are machine-processable and portable — a prerequisite for Phase 3 context layers.

Agentic Web

An emergent layer of the Web in which autonomous AI agents discover, query, and produce structured linked data as first-class participants. For the Agentic Web to function, agents must share resolvable identifiers and portable ontologies — the core argument of Tony Seale's post.

HowTo: Achieve Enterprise AI Context Interoperability

A practical path from proprietary copilot context to a portable, globally joinable knowledge layer — following the framework set out by Tony Seale.

1
Inventory where structured meaning lives in your organisation: databases, wikis, SharePoint taxonomies, copilot system prompts, proprietary ontologies. Identify which of these are locked inside closed platforms and cannot be exported or dereferenced from the outside.
2
Choose an open, portable vocabulary — Schema.org, OWL, or a domain-specific standard — and express your entity definitions and relationships in RDF. Test portability: can you export the ontology to a plain Turtle file and re-import it elsewhere without data loss?
3
Every product, person, concept, or process that your AI agents reason about should have a stable, resolvable IRI — not an integer ID or vendor-assigned GUID. When an external agent dereferences that IRI it should receive a machine-readable description. This is Kingsley Idehen's "hyperlink as stable standardised identifier" principle.
4
Publish your knowledge graph through a standards-compliant SPARQL endpoint. This allows any authorised agent — internal or partner-facing — to query your context without a bespoke API integration. Endpoint federation means your graph can join queries against external graphs in real time.
5
Before declaring your context layer Phase 3-ready, run Tony Seale's three tests: (1) Can your ontology leave the platform? (2) Can another system resolve your identifiers? (3) Can your graph join the global graph? A "no" to any question means your context is still rented, not owned.
6
Use SPARQL 1.1 Federation and owl:sameAs links to connect your graph to public knowledge bases — DBpedia, Wikidata, domain ontologies — and to partner graphs. This is how siloed enterprise context becomes a node in the global graph.

SPARQL Workbench

Queries target the companion knowledge graph. Load the companion TTL file into a SPARQL endpoint (e.g. URIBurner) then click ▶ Run to execute.

PREFIX schema: <http://schema.org/>
SELECT ?position ?author ?text ?reactions
FROM <https://linkeddata.uriburner.com/DAV/demos/daas/tony-seale-enterprise-ai-context-interop-claude_sonnet_4_6-1.ttl>
WHERE {
  ?comment a schema:Comment ;
           schema:position ?position ;
           schema:author   ?person ;
           schema:text     ?text .
  ?person schema:name ?author .
  OPTIONAL {
    ?comment schema:interactionStatistic ?ctr .
    ?ctr schema:userInteractionCount ?reactions .
  }
}
ORDER BY ASC(?position)
▶ Run

Returns all 10 comments in sequence with author name and reaction count where available.

PREFIX schema: <http://schema.org/>
SELECT ?name ?profileIRI
FROM <https://linkeddata.uriburner.com/DAV/demos/daas/tony-seale-enterprise-ai-context-interop-claude_sonnet_4_6-1.ttl>
WHERE {
  ?person a schema:Person ;
          schema:name ?name ;
          schema:url  ?profileIRI .
}
ORDER BY ASC(?name)
▶ Run

Lists every contributor (author + all commenters) with their LinkedIn profile URL.

PREFIX schema: <http://schema.org/>
PREFIX owl:    <http://www.w3.org/2002/07/owl#>
SELECT ?title ?author ?url ?interactionType ?count
FROM <https://linkeddata.uriburner.com/DAV/demos/daas/tony-seale-enterprise-ai-context-interop-claude_sonnet_4_6-1.ttl>
WHERE {
  ?post a schema:SocialMediaPosting ;
        schema:name  ?title ;
        schema:author ?p ;
        schema:url    ?url ;
        schema:interactionStatistic ?ctr .
  ?p    schema:name  ?author .
  ?ctr  schema:interactionType       ?interactionType ;
        schema:userInteractionCount  ?count .
}
▶ Run

Returns post title, author, URL, and all interaction counts (reactions, comments).

PREFIX schema: <http://schema.org/>
SELECT ?position ?author ?excerpt
FROM <https://linkeddata.uriburner.com/DAV/demos/daas/tony-seale-enterprise-ai-context-interop-claude_sonnet_4_6-1.ttl>
WHERE {
  ?comment a schema:Comment ;
           schema:position ?position ;
           schema:text     ?excerpt ;
           schema:author   ?p .
  ?p schema:name ?author .
  FILTER (CONTAINS(LCASE(STR(?excerpt)), "context"))
}
ORDER BY ASC(?position)
▶ Run

Filters to comments that explicitly discuss "context" — the post's central concept.

PREFIX schema: <http://schema.org/>
DESCRIBE ?post
FROM <https://linkeddata.uriburner.com/DAV/demos/daas/tony-seale-enterprise-ai-context-interop-claude_sonnet_4_6-1.ttl>
WHERE {
  ?post a schema:SocialMediaPosting .
}
▶ Run (Turtle)

Returns a full DESCRIBE of the post node in nice Turtle format — useful for verifying the KG structure.