Insights & Podcasts
Go beyond the AI hype. Explore practical insights and candid conversations about what it takes to turn ambitious technology into solutions that work at scale. Read the articles or listen to the podcasts.
INSIGHT
The Correction Layer
Most AI governance programs can point to a human oversight requirement. Few can point to a logged correction from last month. That gap is not a paperwork problem. It is the difference between a requirement that satisfies an audit and a function that catches a system when it is confidently wrong. This article follows the correction layer through five environments, an automated insurance claims desk, a frontier AI product, a carrier-grade communications platform, a self-parking car, and the Swedish Army, and argues that correction is a product decision, not a compliance requirement. It closes with the four functions every correction layer performs and a three-question diagnostic leaders can run on their own AI program this week. You can transfer responsibility with a contract. You cannot transfer correction.
Read the full article or listen to the podcast version.
Article
Podcast
From Demo to Working Solution
INSIGHT
Most AI pilots do not fail because the model is weak. They stall because the organization never builds the ownership, correction mechanisms, escalation paths, and feedback loops required to operate the system in the real world. This article explores why impressive demos often collapse after launch, what separates pilots that quietly compound from those that are eventually shut down, and how leaders can determine whether a stalled AI program can still be saved. The gap between demonstration and production is not technical overhead. It is the work that turns AI into a dependable operating capability.
Read the full article or listen to the podcast version.
Article
Podcast
RESEARCH
Speed Was Not the Differentiator:
What 348 Trials Showed
Traditional SRE models tend to strain as platform complexity grows faster than the operating structure around it. A faster incident summary does not shorten an outage if nobody can act on it. Across 348 controlled trials, single-agent systems produced guidance an operator could execute 1.7% of the time. Multi-agent orchestration produced it in every trial, with zero variance. Speed was not the variable. Determinism was.
Published research: arXiv 2511.15755.