Insights
Field notes from building agents that ship.
Named-source analysis on AI, search, and the work of making agents earn their place. No predictions. Only what is already happening.
-
An AI refused a sale that 92% of shoppers would wave through
MIT's Initiative on the Digital Economy priced flour a penny over a $10 limit. 92% of people bought it anyway. The AI almost never did. Why that one detail decides which work you can hand an agent, and which work has to stay with a person.
Read → -
An AI summary now answers the question your page used to
Pew tracked 68,879 Google searches. On the ones where Google showed an AI summary, click-through to a website dropped from 15% to 8%, and the businesses losing those clicks never lost a ranking. What it means if search brings you customers, and what I changed on our own site because of it.
Read → -
When AI-written code breaks, the blame floats up to the CTO
CloudBees asked 200+ enterprise tech leaders about their AI-written code. 92% trust it before it ships, 81% have already had it break in production, and only 12% have anyone whose job is to govern it. Why the hard part moved to standing behind the output, and the three questions to ask before you scale.
Read → -
Agentic AI in plain terms, from MIT's own definition
MIT Sloan defines agentic AI as software that perceives, reasons, and acts to reach a goal for you. Where generative AI writes you a draft, an agent goes and does the multi-step job around it. A spring 2025 MIT and BCG survey already put adoption at 35%, with another 44% on the way.
Read → -
MIT found the way you pair people with AI agents decides the result
MIT ran the same tasks with human teams and human-AI teams. The pairing changed the quality of the work, down to personality traits. The takeaway for a rollout: dropping an agent next to a team and hoping is the part most companies get wrong.
Read → -
53% of executives expect an AI to negotiate their contracts within a year
Icertis surveyed 1,000+ C-suite leaders, and more than half expect AI to be negotiating their contracts within twelve months. MIT tested whether the bots are any good at it, with a competition run by 300 to 400 top negotiators. What both findings say about where to start.
Read → -
Most AI pilots die because no one answered three questions first
Forrester's 2026 State of Agentic AI report found three-quarters of enterprises building agents and only a small minority running them in production. The reasons are the same everywhere, and Forrester traces all of them back to management. The three moves that separate the companies getting results.
Read → -
AI-driven layoffs backfired at most of the companies that tried them
CNBC tracked 23 S&P 500 companies that cut staff to move faster on AI. 56% saw their stock fall, by an average of 25%. Gartner found the cutters returned no better than the firms that kept their teams. What the companies actually pulling ahead do instead.
Read → -
A quarter of operational decisions are already made by AI, with no human sign-off
IBM surveyed 2,000 CEOs across 21 industries in 2026. A quarter of their operational decisions already run without a person in the loop. The handful of things the CEOs pulling ahead do consistently, and the one most mid-market businesses still get wrong.
Read →