AI's semantic turn
Why trust in AI agents depends on a new architecture of meaning
TL;DR: Over the past six months, several of the most capable enterprise data platforms have converged around the same architectural idea: AI agents need access not just to organisational data, but to organisational meaning. Ontologies, context graphs, and persistent memory are all attempts to make that meaning available to agents in structured form. This is real progress. But it resolves only one half of a two-part challenge. Capturing reasoning is not the same as knowing when that reasoning applies to a new situation. Modern data platforms are well-positioned to capture reasoning. They are less well equipped to judge when that reasoning should apply to a new situation, a harder problem the knowledge management field has wrestled with for decades. Progress will come from bringing these fields together.
Meaning becomes structural
Through 2024 and 2025, much of the generative AI discourse focused on model capability. More recently, a different question has come into focus: how to make meaning persist beyond a single interaction. A model can be given context in a prompt, but that context usually disappears when the conversation ends. For AI agents to be trusted in decision-making contexts, they need to be grounded in meaning that holds across the situations they encounter.
The answer taking shape across the industry is that meaning needs to be structural, not optional.
In late 2025, Foundation Capital's Jaya Gupta and Ashu Garg proposed context graphs as 'AI's trillion-dollar opportunity',1 a missing architectural layer that captures the decision traces, exceptions, and reasoning that conventional data architectures do not represent. I wrote about it in late March, in the context of decision intelligence.2 What has become clear since then is the breadth of the convergence.
The same recognition is visible across the enterprise AI stack, with major platforms surfacing the theme in recent weeks. On 10 April, Databricks Research published an article proposing memory scaling as a new design axis for AI agents.3 Their argument is that, for real-world agents, the bottleneck is shifting from raw reasoning capacity to “grounding the agent in the correct information.”
Microsoft expanded Fabric IQ at FabCon Atlanta in March, positioning ontologies, graph capabilities, and semantic models as the foundation for agents reasoning across an enterprise data estate.45 Palantir, whose Ontology platform predates the current wave by several years, has long articulated a sharper version of the same principle: the ontology should connect data to the decisions, operations, and actions it supports.67
At the standards level, the Open Semantic Interchange, launched last September by Snowflake, Salesforce, and partners,8 has expanded to a working group of more than thirty vendors including AtScale, Databricks, dbt Labs, Atlan, and Alation. The first version of the specification was finalised in late January.9 The underlying recognition is that business logic should be defined once and travel across tools, rather than be reinvented inside each platform.
A familiar problem
For anyone who worked in knowledge management over the past thirty years, this conversation will sound familiar. KM started from the recognition that much of what matters in organisations sits outside structured systems. Insight, rationale, and experience live in documents, emails, reports, conversations, and the working habits of people who know how things get done.
Early KM systems struggled with adoption because the manual effort required to capture and maintain that knowledge was unsustainable. In practice, many retreated to improving discovery and retrieval of documented knowledge through taxonomies and enterprise search. Knowledge graphs later emerged to encode organisational relationships and context, but the richer ambition of capturing reasoning was rarely realised at scale.
AI has changed the economics of the problem. Large volumes of documents, emails, reports, messages, and system traces can now be processed, classified, linked, and summarised at a scale manual KM programmes could never sustain.
Most of the platforms facilitating this shift come from technical lineages built around large-scale numerical workloads. That lineage has produced highly effective systems for structured data, but it is less well suited to dealing with the fluid, implicit, and situation-dependent nature of context and human-centric decisions.
The interesting move is that these platforms are now engaging with questions that the knowledge management field has been working on for decades: how raw data becomes processed information, curated knowledge, and eventually experienced wisdom. This DIKW hierarchy, popularised by Russell Ackoff in 1989,10 was a staple of KM thinking for years. I wrote about it in 2005, exploring how the same set of facts can lead to different decisions depending on context.11
The appeal of the DIKW hierarchy is that it acknowledges meaning is layered. Its limitation is that it implies a clean upward path. The harder reality is that each step up requires interpretation. And interpretation becomes more contested at the higher levels. This is not a solved problem and scaling data does not make it go away.
A new architecture
The recognition that AI agents need access to meaning is driving semantic advances that KM systems could only aspire to: reasoning that can be discovered, represented, and queried at scale. Across the systems moving in this direction, three architectural ideas recur. They are not settled categories, and the boundaries between them are uneven, but each grounds AI agents in a different aspect of meaning.
Ontologies define structure. Palantir has been one of the most prominent proponents of using ontology to ground decision-making in data at scale, representing the structural reality of an organisation through objects, relationships, business rules, and the actions that connect them. Decisions have three components: data, logic, and action. Ontologies make those relationships explicit, giving systems and agents a stable foundation to reason against. Of the three approaches, ontologies are the most mature. Their strength is in making durable business structures explicit. Their limitation is that they are better at representing what an organisation knows about its operating model than how meaning shifts across situations.
Context graphs trace how meaning forms in use. The framing has been advanced most prominently by Foundation Capital's Jaya Gupta and Ashu Garg, who position context graphs as a missing architectural layer for AI agents. Where ontologies define relatively stable structures, context graphs connect fragments of information, exceptions, decision paths, and situational cues into a navigable representation of how an organisation operates in practice. Their value lies in showing what was relevant in a particular situation, why it mattered, and how that relevance may change. For now, though, context graphs remain an architectural proposition rather than a mature data layer: widely discussed, but not yet standardised or proven at scale.
Memory accumulates experience. Databricks' memory scaling post proposes a pattern in which an agent's identity lives in a persistent memory store. The premise is that everything an agent picks up through use, from individual exchanges and corrections to broader patterns of how a business operates, becomes a resource it can draw on. Such an approach would allow agents to learn from interactions over time rather than through expensive model retraining. The pattern is promising, with early components in place, but the integrated architecture is still emerging and keeping memory accurate, current, properly scoped, and safe to reuse remains ongoing research.
These three approaches are not a complete map of enterprise semantics. They are early signs of a broader shift: a new architectural layer is emerging to make meaning persistent, queryable, and usable across contexts.
The emergence of initiatives like the Open Semantic Interchange suggests that parts of this infrastructure, particularly business definitions and metadata, are moving toward open, interoperable standards. But the openness is uneven. It applies most clearly to declarative semantics, less so to context, and only weakly to accumulated experience. Memory, decision traces, and operational ontologies remain more tightly coupled to the platforms that generate them.
The harder problem
There is an assumption running through many of these efforts: that experience, once captured, can inform future decisions. If decisions, reasoning, and context can be represented, whether as decision traces, ontology objects, or memory shards, then those representations should become useful guides for future decisions.
That is a reasonable engineering response, but it treats the problem too much like one of data quality. The harder problem is that decisions made in real-world environments are not all of the same kind. They fall into at least three categories, and each behaves differently when captured and reused.
Some decisions reflect stable patterns that do generalise. Terminology, schemas, recurring processes, well-understood business rules. These are the natural territory of ontologies and structured semantic layers. Capturing them once and reusing them across the organisation is exactly the kind of work the current architectures are well suited to.
Other decisions are situational. They were appropriate to a specific set of conditions but become misleading if reused without adjustment. The 20% discount approved last quarter, the policy exception granted under unusual market conditions, the deviation from a process that worked because of who was in the room. These decisions are not wrong. They were right for their moment. But the moment was part of the decision, and a context graph that records the action without the situational sensitivity it depended on will quietly mislead any agent that reaches for it later.
A third kind of decision is irreducibly variable. Different people may interpret the same situation differently, rationalise their choices differently, and still reach defensible decisions. Even the same person may make a different judgement across situations that look similar from the outside. Such decisions are not random, but neither do they necessarily carry reusable meaning. Once captured as data, they can be mistaken for patterns that turn variability into precedent.
If semantic infrastructure treats these three categories as equivalent, it can accumulate a detailed record of the past without knowing how that past should inform the future. The harder task is to distinguish what can be reused, what must be reinterpreted, and what should remain contingent.
A path forward
Knowledge management and real-world decision-making have spent decades with the question that semantic infrastructure now has to confront: when does captured reasoning apply, and when does it not? Polanyi’s observation that “we can know more than we can tell”12 was one starting point. Gary Klein’s field studies of expert decision-making in high-pressure environments showed that much of what makes decisions work is subconscious recognition of patterns that the decision-makers themselves cannot fully articulate.13 My own doctoral research drew a related distinction between the stable, the situational, and the irreducible variability of human behaviour, and what happens when the layers are collapsed to a single value.14
That work matters now because AI agents change the stakes. Knowing when representations of organisational reasoning can be trusted has always been difficult in human decision-making. It becomes critical when those representations are used by AI agents that can act across systems, at speed, and at scale.
The path forward sits between two bodies of work: platforms that can now capture context at scale, and fields that have spent decades studying how context shapes decisions. Neither is sufficient alone. The next phase of progress will depend on joining semantic infrastructure to decision intelligence, so agents can do more than retrieve organisational history. They need ways to judge when past reasoning should shape what happens next, and when human intervention is required.
References
Gupta, J. & Garg, A. (2025). “AI’s trillion-dollar opportunity: Context graphs.” Foundation Capital. Source
Richardson, S. (2026). "AI and a new frontier for decision intelligence." Joining Dots, 30 March 2026. Source
Databricks AI Research Team (2026). “Memory Scaling for AI Agents.” Databricks, 10 April 2026. Source
Ulag, A. (2026). “FabCon and SQLCon 2026: Unifying databases and Fabric on a single, complete platform.” Microsoft Azure Blog, 18 March 2026. Source
Microsoft Fabric Blog (2026). “What’s next for Fabric IQ Ontology: The operational context that powers your AI agents (Preview).” 25 March 2026. Source
Palantir (2025). “Connecting AI to Decisions with the Palantir Ontology.” Palantir Blog, 24 November 2025. Source
Snowflake (2025). “Snowflake, Salesforce, dbt Labs, and More, Revolutionize Data Readiness for AI with Open Semantic Interchange Initiative.” Press Release, 23 September 2025. Source
Snowflake (2026). “Open Semantic Interchange (OSI) Specification Finalized.” Snowflake Blog, 28 January 2026. Source
Polanyi, M. (1966). The Tacit Dimension. University of Chicago Press.
Klein, G. (1998). Sources of Power: How People Make Decisions. MIT Press.
Richardson, S. (2020). Modelling the Social Dynamic of Urban Landscapes from Real-Time Data. PhD thesis, University College London.