XNews

Context Graphs: capturing decision reasoning for enterprise AI

Context Graphs: capturing decision reasoning for enterprise AI

Context graphs record decision reasoning, not just facts. - Current enterprise systems record facts, not reasons - Example: Salesforce/Workday/SAP store outcomes, not why - AI executes tasks but does not record its reasoning - Context graph records inputs→rules→approvals→results - Capture event clock, not just state clock - Agents produce trajectories that reveal organizational ontology - Let agents' guided walks produce structure, not static schemas - Build a world model to support simulation and counterfactuals - PlayerZero uses production-issue trajectories to simulate code-change impact - Entrepreneurial advice: target decision-heavy, fuzzy-rule, cross-dept flows

Key insights

Context graphs capture the 'why' behind decisions

Context graphs are a proposed new software layer that records AI decision reasoning in addition to final results. Posts describe how existing enterprise systems (Salesforce, Workday, SAP) store facts such as transaction values but not the rationale behind decisions. The context graph concept records inputs, applicable rules, approvals, and resulting actions (for example: prior complaint → policy allows exception → manager approval → 10% discount). This captured reasoning enables accumulation of organizational judgment as structured, queryable traces rather than isolated facts.

Event clocks vs. state clocks: record process dynamics

The posts identify a critical gap: existing systems optimize a 'state clock' that logs current values but ignore an 'event clock' that records how and why values changed. Examples include configuration edits where only the final value is stored and the rationale is lost. The context graph approach emphasizes reconstructing event sequences and reasoning so that AI can access the full decision trace instead of only static states.

Agents' trajectories reveal organizational ontology

Rather than predefining a static schema, posts propose letting AI agents traverse systems while executing tasks; their guided trajectories—accessing APIs, data, and tools—produce repeated co-occurrence patterns that naturally surface entities and relations. These agent-generated walks are compared to graph-embedding random walks but are goal-directed, producing a usage-driven organizational ontology that emerges from problem-solving behavior.

Context graphs as a world model for simulation and counterfactuals

Accumulated agent trajectories form a context graph that functions as an organizational world model capable of simulating actions and evaluating counterfactuals. The posts state this model can encode 'organization physics'—how decisions propagate, exception handling, and change effects—and can predict outcomes like the impact of code changes on production, supporting simulation-based reasoning without continually retraining base models.

Practical startup guidance: build decision-capture systems

Posts advise entrepreneurs to prioritize building systems that capture decision processes rather than only integrating AI on top of legacy data stores. Recommended target areas are processes with frequent human judgment, fuzzy rules and exceptions, and cross-department coordination—areas where recording 'why' decisions were made delivers organizational learning and creates commercial opportunity.

Source posts