AI decisions, examined in depth
Longer takes on the topics that matter for AI leadership. Each article builds on a LinkedIn post with added context, evidence, and frameworks.
It Checked Its Own Work
Anthropic's Claude caught its own analysis error in a lab test, then an outside lab confirmed it was right. Here's why that matters more than speed.
Read article →
Four Levers, Not One Dial
AI autonomy isn't one dial. Anthropic's own team breaks it into four separate handoffs, and only one of them decides whether the rest are safe to hand off.
Read article →
The Silent Exit Clause
A safety check the builder can veto isn't a check. Anthropic's own policy confirms the gap, and METR's audit shows what real independence looks like.
Read article →
The Check That Left With the Afternoon
A McKinsey case describes a distributor whose system reads public building permits and drafts the sales outreach. The work it removed was somebody's afternoon, and an afternoon is never only one job: while a person was assembling the material, they were also reading it. That reading was a control nobody had written down, and it leaves with the work.
Read article →
Don't Build What You Can't Keep
McKinsey's 2026 survey found 32% of respondents saying their organization decided against buying at least one software product or feature because it could be built with coding agents. The companion post argued that the buy rule grew a second half. This article does the three things the post could not: reads the survey's size gap honestly, sorts builds into the three kinds that need different rules, and shows how to price what keeping a built tool costs every month.
Read article →
Cheap to Write, Expensive to Keep
A system that remembers across more than one layer has a stratum you can delete with no consequence and a stratum that takes everything with it. Cost asymmetry moves important knowledge into the cheap one without anyone deciding it. This is the evidence for why that happens, and why better writing tools do not fix it.
Read article →