XenonStack BlogBeyond Context Graphs: The Four Gaps Between Decision Traces and Better Decisions

Are context graphs enough for enterprise AI? No. Context graphs are AI's trillion-dollar opportunity, and the thesis is right, but it is incomplete. Durable records of how decisions were made are necessary. Four gaps keep them insufficient: representation (YAML models arithmetic, not meaning), technology (graphs represent history, intelligence computes), operationalisation (capturing context is easy, using it efficiently is not) and evolution (imitating human decisions is phase one, optimising beyond them is the endgame).

The 60-second brief

• The consensus: AI that decides at scale needs durable, queryable records of how decisions were made: reasoning, exceptions, approvals, precedent.
• The representation gap: we encode meaning in YAML. That models arithmetic, not semantics.
• The technology gap: graphs represent history. Intelligence requires computation over evolving state.
• The operationalisation gap: capturing context is easy; using it efficiently at scale is not.
• The evolution gap: imitating human decisions is phase one. Optimising beyond them is the endgame.

The Thesis, and Where It Stops

If AI systems will make decisions at scale, we need records of how decisions were made, not just outcomes: the why behind the what. I agree completely. But the current discourse stops too early. It names what is missing, decision traces and the schema-truth gap, and leaves four gaps open. Until they close, context graphs remain necessary but insufficient.

Gap 1: Representation. Why "Semantic Layers" Are Not Semantic

The industry made a category error. We called them semantic layers, implying they capture meaning. They capture calculations:

semantic_models:
  - name: orders
    entities:
      - name: order_id
        type: primary
    measures:
      - name: revenue
        agg: sum
        expr: order_total

That tells you how to compute revenue. It does not tell you what revenue means here, what it includes or excludes, when it is recognised, or how it relates to other financial concepts. An ontology captures that:

:Revenue a :FinancialMetric ;
    :calculatedFrom :CompletedOrders ;
    :excludes :RefundedOrders, :PendingOrders ;
    :recognizedWhen :PaymentCleared ;
    :relatedTo :GrossMargin, :NetIncome ;
    :definition "Total value of completed transactions, excluding refunds and chargebacks" .

AI systems do not consume dashboards; they consume representations. If your semantic layer encodes metrics without concept hierarchies, relationship types and definitions, your AI can answer calculation questions, not reasoning questions. More in Semantic Layer for AI: Necessary, but Not Sufficient.

The real constraint is skills. This is a discipline gap, not a tooling gap: you would not ask a financial analyst to build medical ontologies, and you should not expect data engineers to model your domain's meaning.

Gap 2: Technology. Graphs Explain; Intelligence Computes

Graphs can tell you what was decided, which entities were involved, what policy applied and what precedent existed. They cannot infer latent intent, evaluate counterfactuals, optimise trade-offs or learn better strategies over time. Graphs are memory. Intelligence is stateful reasoning.

The gap is obvious outside enterprise workflows. A shopping assistant may have purchase history, product relationships and discount rules, but intelligence means inferring intent from browsing sequences, predicting conversion, balancing margin against satisfaction and learning from every interaction. That is the territory of MDPs, POMDPs and reinforcement learning, not just LLMs and graphs. The substrate is the graph; the intelligence lives above it. See Production World Models for Agentic AI.

Gap 3: Operationalisation. Capture Is Not Use

Once decision traces are durable, context stops being a prompt and becomes an accumulating substrate. More traces slow retrieval; richer reasoning burns tokens; fresh context adds infrastructure complexity. Marginal quality gains can cost orders of magnitude more in latency or spend. The bottleneck is not access. It is efficient utilisation.

Financial markets solved this long ago: P&L-explain systems reconstruct decisions in a replayable, attributable, auditable way, at millisecond latency and massive scale. Context graphs should learn from that playbook, and measure what matters:

Without discipline, context becomes technical debt: stale precedents, redundant traces and unbounded growth. The cost of knowing why can exceed the value of knowing. See Context at Inference Time and Context Freshness for AI Agents.

Gap 4: Evolution. From Mimicry to Optimisation

Most context-graph discussion ends at "we need to understand why people made decisions". Correct, and incomplete. Agents evolve through three phases, and most systems never leave the first:

AlphaGo's move 37 was not creativity; it was search in a better state space. People reason with limited memory, no counterfactual simulation, heuristic shortcuts and risk aversion. Agents need not share those limits, if the system represents state correctly, connects actions to outcomes, learns across episodes and allows safe exploration. Context graphs enable explanation. Optimisation requires learning.

The Complete Picture

The opportunity is not capturing decision traces. It is making better decisions than people could.

Here is what closing all four gaps means for a single decision:

Phase 1 explains what people did. Phase 3 does better.

The Real Question

Can this system do better than people, or only explain what people did?

If the answer is "only explain", you have built phase-one infrastructure for a phase-three world. The opportunity is not capturing decision traces. It is using them to make better decisions than people ever could, and that requires closing all four gaps. The compounding mechanism is described in The Context Flywheel.

Your turn: which of the four gaps is widest in your organisation today?

Find Your Widest Gap

ElixirData Context OS combines ontology, the context graph and decision traces, for governed agents running on ElixirClaw. In a Context Workshop, XenonStack works with your CDO, CAIO and data platform leaders to score one decision domain against the four gaps and choose which to close first. Related reading: Why RAG Isn't Enough for Enterprise AI and Feedback Loops for Enterprise AI.

Book a Context Workshop  |  Explore ElixirClaw Agentic OS

Frequently Asked Questions

What is a context graph?

A durable, queryable record of how decisions were made: the reasoning, exceptions, approvals and precedent behind each outcome.

Why are context graphs not enough?

They explain what was decided. They do not by themselves represent meaning, compute over evolving state, serve context efficiently or learn better strategies.

What is the representation gap?

Semantic layers written in YAML encode how to calculate a metric, not what it means. Meaning needs ontologies with definitions, inclusions, exclusions and relationships.

What is the difference between mimicry and optimisation?

Mimicry replicates human reasoning and is bounded by human performance. Optimisation discovers better strategies through feedback and safe exploration.

What does closing all four gaps require?

Ontologies for representation, stateful computation for reasoning, context engineering for efficient use and policy learning for improvement.

This article first appeared in Navdeep Singh Gill's LinkedIn newsletter, AI + Human = Human Squared.

Explore the XenonStack Platform and Agent Network

Platforms: ElixirData · ElixirClaw · ElixirHub · NexaStack · MetaSecure · Neural AI · AutonomousOps · CohortIQ · AI Aviator · SAIF Aviator

AI agents: AgentAnalyst · AgentEvaluation · AgentForce · AgentGRC · AgentHR · AgentInstruct · AgentLabel · AgentPixel · AgentQA · AgentRAI · AgentScrum · AgentSearch · AgentSketch · AgentSRE

XenonStack: xenonstack.com · xenonstack.ai · xenonstack.ae · xenonstack.jobs