Heidi Porter, Kevin Solorio, Chris Shinkle, and Andrew Dunlap make a practical case for bounded AI autonomy in these Midwest AI Summit talks on Practical AI. Porter separates three decisions: what performs each task, how the steps are coordinated, and what may execute without approval. Her examples put vision models on defect detection, agronomy models on crop predictions, and language models on interpreting questions or explaining retrieved evidence. Runtime checks, evaluation cases, recovery paths, and clinician or operator overrides surround the model. Solorio brings the people doing the work into that design process. Field technicians can explain why a management-friendly tool adds work, identify wrong repair advice, and turn corrections into a maintained test set. His loop runs from offline tests to sampled production runs, human annotation, and new regression cases. Shinkle and Dunlap apply the same reasoning to coding agents: choose autonomy from token budgets, organizational support, goal clarity, available domain knowledge, verifiability, and required code oversight. A two-day prototype stalled when production exposed missing knowledge and unscalable verification. A building-components team gained shared context while deliberately retaining oversight. A credit-union migration used coordinated agents with a known target and substantial backing. SEP reports a sixfold effort reduction for that migration, without a measurement protocol that would support generalizing the number. Shared knowledge, scheduled iteration, verification before speed, bottleneck removal, and scripted repetitive work matter across the spectrum. Greater autonomy has to fit the team’s ability to recognize and recover from mistakes.
“A guardrail is actual machinery, a gate, a check, a human, a limit in code.” — Heidi Porter, 08:24
“An agent without evals is a proof of concept.” — Kevin Solorio, 26:55
“Because the goal here is fit.” — Andrew Dunlap, 48:37