The problem
A human in the loop doesn't scale by itself
Every organization adopting AI makes the same promise: a human stays in the loop. The promise breaks on volume. AI generates work faster than reviewers can read it, and review capacity doesn't grow with it. Two failures follow. Teams try to review everything, and attention thins until review becomes a formality. Or they trust the model, and the risk stays invisible until something fails.
Human inspection has limits wherever it has been studied: error rates of 20% to 30% are commonly reported for manufacturing inspectors, and accuracy falls as defects get rarer (See, 2012). Normally the riskiest work gets the most scrutiny. With AI, that is starting to invert.
Peer review was built for the volume and pace of human authors. It doesn't scale to this, and neither does any review process built on the same assumption, in code or anywhere else AI output reaches decisions or production.
What we do
Oversight designed for the volume AI produces
We work with engineering leaders, product leaders deploying AI agents, and the risk and compliance owners accountable for the result. We find where oversight is failing, decide with you which AI outputs and actions need a human, and build the review and accountability systems to run it, so your team can keep running them after we leave. See engagement types →
The approach
Plan human attention like capacity
Human attention is the scarce resource, so we plan it rather than hope it stretches. Risk is assessed independently; AI agents review the work adversarially; people decide where the agents disagree or the risk is high. Cleared work is sampled, and every decision is recorded. AI-generated code is the hardest case we know, and the one our research is built on. Read the longer argument →
Recent writing · All research →
LinkedIn ↗
Overseeing AI-generated code
What a systematic review of 9,502 records says about deciding which AI-generated changes need a human, and six practical steps.
LinkedIn ↗
Code is outpacing our ability to check it
Verification, not generation, is now the bottleneck. Why reviewing everything and trusting the model both fail.