The Model Is Not the Product: What a Breast Cancer Lab Workflow Taught Me About AI Architecture

We design AI systems assuming the model is the differentiator. But what happens when your model hits 62% accuracy on a workflow that demands >99%? And what happens when the fix has nothing to do with the model?

What you’ll see in this talk

If you build AI systems for regulated or procedure-driven work, this talk walks you through Chris Henry’s real experiment turning a breast cancer lab’s standard operating procedures into an LLM-guided routing system. You’ll follow the accuracy climb from 62% to 100%, see where local models broke, and take away why the harness around the model, not the model itself, is the actual product.

• How a breast cancer lab workflow runs end to end, and why patient privacy and 100% accuracy are non-negotiable when 0.2% error means 14 misrouted patients a week.
• Why feeding whole SOPs to an LLM fails, and how translating them into deterministic rules, skills, and worked examples raised reliability step by step.
• The lesson that LLMs are bad at deterministic checks (null checks, set membership, number ranges) but strong at judgment, so each job belongs to the right engine via an orchestrator.
• Building the harness: local inference inside the trust boundary, model swappability across a dozen models, observability with Langfuse, and cryptographically signed audit receipts.
• Why a model’s benchmark score is meaningless without its harness, shown by the same model scoring six times differently across tools.