# Engineering Agency out of the Happy Path — Matthew Jewkes, Standard Cybernetics

## Executive summary

The presentation argues that the most valuable application of AI automation is not general agent deployment, but rather achieving 'full automation' of the routine, predictable 'happy path.' By encapsulating these high-volume, low-variance tasks, organizations can recover expert attention and structure the remaining work into three distinct categories: critical errors, ambiguous/contentious cases, and unusual exceptions. This structured backlog allows domain experts to shift from reactive multitasking to proactive strategic thinking, identifying potential new product lines or market investments.

## Key takeaways

- Focus on Full Automation, Not Agents: Full automations (taking humans out of the happy path completely) are preferred over complex agent systems because they are legible, stable, and hide complexity. Agents are often difficult to maintain and increase the burden of decision-making for experts. (0:00)
- The Three Buckets of Remaining Work: After automating the routine path, remaining cases should be categorized: 1) Critical errors (where consensus exists on a misstep); 2) Undecided/Ambiguous cases (where experts disagree or the organization lacks consensus); and 3) Unusual exceptions (novel science or processes requiring deep human judgment). (6:39)
- From 95% to 99.95% Accuracy: By focusing on a specific, well-articulated set of business-critical tasks and employing advanced techniques (like 'hill climbing' and extensive evaluation), achieving high reliability (e.g., 99.95%) is possible for specific automations. This performance is claimed to be better than human performance at scale, though not for a single, deep-dive case. (9:58)
- Structuring Attention for Strategy: The primary benefit of automation is organizing the 'exhaust'—the exceptional cases—into a structured backlog. This allows experts to identify patterns, form investment theses, and advocate for new business lines, turning operational data into market signals. (12:51)

## Technical details

- Biopharma Indexing Case Study: The system processed tens of thousands of new tasks monthly, automating the indexing and relationships between drug development documents for Clarivate. The initial approach reached 80-83% performance, improving to 99.95% accuracy on the most valued business tasks. (4:18, 9:58)
- Model Failure Analysis: The speaker notes that building the system required significant investment in 'reading evals' (evaluation examples) to understand exactly why the AI failed, which was described as an exhausting process. (17:17)
- Automation Architecture: The system was designed with a robust routing layer that first checks if a task matches the 'happy path.' If not, it routes the input to specialized pipelines for errors, ambiguity, or exceptions, preventing the task from entering the main workflow. (8:52)

## Practical implications

- Focus automation efforts on processes that are scaled, have material ROI, and are associated with seven figures of revenue.
- Prioritize processes that are 'specifiable' and detail-oriented, rather than subjective judgment work.
- When making a large investment, plan for organizational change, trust building, and how the resulting efficiency will be leveraged for sales and marketing growth.
- The goal of automation should be organizational refactoring: reducing complexity and directing expert talent toward higher-value, strategic problems.

## Topics

AI Automation, Process Engineering, Domain Expertise, Business Process Optimization, Data Indexing, Machine Learning Reliability

Source: https://www.youtube.com/watch?v=wWisOHEcMPE
