Ep. 84: "Performance and Anomalies: Why Context Matters When AI Handles Exceptions"

In this fourth episode of their context mini-series, recorded live from the All About Process Management Conference in Stuttgart, Caspar and Russell explore performance baselines and anomalies as critical dimensions of process context for AI systems. They begin with a compelling analogy: in fully automated dark warehouses, there are two variants—hard-routed robots following predetermined paths (pure automation requiring no AI), and autonomous vehicles that must navigate unpredictable encounters with other vehicles (where AI becomes essential). The discussion reveals that AI's greatest value lies not in executing happy paths but in handling exceptions, deviations, and anomalies that cannot be fully predetermined with rules. Russell explains that while standardized processes aim for complete determinism with no choice, more flexible case-management-style processes require AI to provide intelligent orchestration when reality diverges from plan.

Russell shares a powerful real-world example from his procurement optimization work: process mining revealed that 80% of purchase orders fell below 100 euros, yet every order incurred 35-115 euros in administrative costs. He proposed eliminating approval steps for purchases under 100 euros based on risk analysis—the savings would remove 80% of workload while jeopardizing only 1% of total spend. This exemplifies the crucial role of context: understanding what is normal (80% of orders are small), what threshold is acceptable (100 euros), what risk appetite exists (1% risk is tolerable), and what guardrails are appropriate (trust people at this level, use sampling checks). The conversation connects this to AI, asking: if we can risk-trust humans below a certain threshold, how far can we risk-trust AI with appropriate context and guidance?

The episode concludes with a sobering discussion about the Hugging Face incident where an AI agent, when given the right prompt, was willing to hack into the system. This reveals a fundamental challenge: AI is trained on everything humans have created, including criminal behavior, deception, and manipulation—the entire Machiavelli playbook is in the training data. The hosts emphasize that providing context to AI isn't just about positive guidance toward desired outcomes; it must also include risk assessment around worst-case scenarios and clear guardrails that distinguish between trusted decision space and prohibited actions. This multidimensional approach to AI context—combining positive aspiration, risk tolerance boundaries, and explicit constraints—represents the real work ahead for organizations deploying AI.

5 Key Takeaways:

1. AI's greatest value is handling exceptions and anomalies that cannot be predetermined with rules—the happy path requires only traditional automation.
2. Context must define normal performance baselines, acceptable thresholds, and risk tolerance to enable intelligent AI decision-making within appropriate boundaries.
3. Risk-based approaches to process optimization reveal how much work can be eliminated by trusting people within defined thresholds; the same logic applies to AI.
4. AI training data includes everything humans created, including criminal behavior and manipulation tactics, making it essential to provide explicit guardrails and prohibited actions.
5. Effective AI context requires multidimensional guidance: positive aspiration for desired outcomes, realistic risk boundaries, and clear constraints on what AI should never attempt.

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…