Context Architecture Unpacked: Semantic Layers, and Deterministic Guardrails for AI

Show notes

In this closing episode of the context mini-series, Caspar and Russell step away from the high-level concepts of previous episodes to examine the technical architecture behind a context model. Russell opens by reflecting on how the idea of an integrated enterprise model, linking strategy to operations through balanced scorecards and process mining, has been a recurring theme in process management for decades, and asks what has truly changed now that context has become a dominant buzzword. He argues that the real shift is that AI has been applied across so many use cases that organizations now need structured, reusable context that AI agents can consume without knowing in advance which specific use case they will serve. The hosts agree that a context model cannot realistically capture everything, so they distinguish between use-case-specific guardrails and generic guardrails that keep an agent from going off course when it steps outside its intended domain. The core of the technical discussion is the semantic layer. Russell explains that rather than consolidating all enterprise content into one repository, a semantic layer is an ontology that references artifacts held in different systems, from process mining and KPI tools to documents on SharePoint, acting as a structured index that tells AI where to find context and how to interpret it. Hard rules, such as excluding draft or unpublished process models, sit alongside open rules that tolerate incomplete data; Russell illustrates this with a missing social security number, which should signal incomplete data rather than a conclusion about a person. The hosts stress that process governance and data governance are both essential, because poor data quality remains the primary threat to AI output. Caspar raises concerns about uncontrolled data accumulation, while Russell describes how restricting AI access to sensitive data often weakens its performance, and both agree that access must be governed through defined rules and approvals rather than either open access or total lockdown. The episode closes with a practical lesson from Russell’s own development work. Although AI helped him define requirements in detail, he found that a deterministic program was more reliable, cheaper, and easier to control than repeatedly prompting an AI agent for each step. Russell summarizes the principle as “think AI first, think deterministic second,” which the hosts adopt as the takeaway for the episode. They conclude the context mini-series and promise guests and follow-up sessions on the topic. 5 Key Takeaways: 1. A semantic layer is an ontology that references enterprise artifacts across systems rather than consolidating all content into one repository, giving AI a structured index to find and interpret context. 2. Context models cannot capture everything, so organizations need use-case-specific guardrails combined with generic guardrails that keep AI agents within bounds when they move beyond their intended domain. 3. Hard validity rules, such as excluding unpublished process models, should be combined with open rules that treat missing data as incomplete rather than as a conclusion about reality. 4. Data governance and process governance are equally essential, because poor-quality data produces poor AI outcomes and uncontrolled data accumulation creates serious risk. 5. Where requirements are well understood, a deterministic solution is often more reliable, cheaper, and controllable than an AI agent, so teams should think “AI first, deterministic second.” If you have suggestions or questions, please reach out to us via questions@bpm360podcast.com If you enjoy our content, please like, rate, subscribe… we do appreciate that…