Hi, it's Andreas here.
No issue last week. I was inside The AI Agent Cohort #2 and in the first session of an executive board AI program I launched with a very big enterprise. More on that soon. It could turn into the most transformative work I do this year.

I am also still looking for a student to support this work here: 20 to 40 hours a month, well paid, and working on the cutting edge of AI. If you know the right person, reply to this email. If they turn out to be the one, there is $500 in it for you (I prefer people in CET timezone).

A date for your calendar: Thu Sep 24, 17:00 CEST, The AI Agent Workshop comes back. Almost 2,000 of you enrolled last time and a lot of you asked for a second round. This time we will build five agents live, agents which will take hours per week from your plate. Make sure you follow along in Claude Code, Codex or your preferred harness, and you will leave with all five running. And gain 10+ hours a week back. Save your seat here.

In this issue:

Four new frontier models in one week → GPT-6 Astra, Muse Spark 1.3, Fable 5.1 and Lyria 3.5 all shipped.

Nvidia buys Hugging Face for $12.93B → 3 million models now owned by a chip vendor.

Copilot can now approve pull requests → Off by default, but it counts toward required approvals.

Stripe blocked $300M of fraud on a $10 AI plan → More fraud in days than OpenCode makes in revenue.

Your AI roadmap can hit every savings target and still miss the biggest business opportunity.

At IBM I had a front-row seat to what it takes to deliver AI at enterprise scale. Almost every one starts with the same brief: cut costs, save time, improve efficiency.

Sensible objectives, which become a strategic constraint when they define everything that gets funded.

I know how that sounds coming from me. I have spent a good part of this newsletter on the honest math behind agents, rising costs and running lean. That argument stands. Any serious AI program needs sound economics. But understanding what AI costs still leaves a bigger question unanswered: what should you build with it?

I sold the pain-point approach myself. Find a problem, see whether AI can make it hurt less, repeat. I ran plenty of those workshops. The limitation is built into the method: the existing process defines the opportunity.

That is the logic behind the left-hand curve. Pain drops, then the gains flatten. Once the form is automated, the form is still a form. There is only so much value to extract from improving the task you started with.

Now change the question: what could we offer that used to be too expensive, too slow or too difficult to deliver?

Take a services business that can only afford to provide detailed, ongoing analysis to its largest clients. The obvious AI project is to reduce the hours needed to prepare that analysis. If AI makes the work economical for smaller clients, the opportunity expands: a service once reserved for a few customers could become a viable offering for an entirely new customer segment.

The same reduction in delivery cost can improve the margin on an existing service or make a new offering viable. Which opportunity you pursue depends on the brief.

One of my clients, a large German carmaker, came in with a long list of pain points from a single function. We ran the other question over one department. Two of the three strongest ideas had been on nobody's list.

That is what the right-hand side of the chart represents: a wider set of possibilities for products, customers and ways of creating value. Some will fail the economics. Some will fail in practice. They still deserve consideration before the roadmap fills up with another round of process improvements.

Look at what your approval process rewards. If every proposal has to justify itself in hours saved, your teams will learn to propose projects that save hours. The opportunities that could change who you serve or what you sell may never reach your desk.

At your next AI program or strategy review, bring the list of processes you want to improve. Then bring the products you shelved, the customers you could not serve profitably, and the services you assumed would require too many people.

Ask which assumptions AI has actually changed. Pick the strongest opportunity and test whether customers want it, whether it works reliably and whether the economics hold.

Pain has a floor. Possibility does not.

What your agent's context window costs you → Microsoft's Jeff Hollan: cost is what enters the context each turn. Tool search alone cut input tokens 97% on big libraries.

A 76-minute course on spec-driven development → Paul Everitt of JetBrains, 15 lessons: spec, plan, implement, verify, then package the workflow as an agent skill (one of the best courses on SDD I’ve seen).

Claude Code now flags your dead-weight skills → /skill-doctor in 2.1.261 lists unused skills and what each costs you in context. Usually explains a slow, expensive session.

When AI handles incidents, engineers lose touch → Ex-LinkedIn SRE Sylvain Kalache, via Bainbridge 1983 and the FAA simulator rule: put incident simulation into on-call readiness.

I take on a limited number of advisory engagements each quarter to ensure your AI programs are successful, achieve ROI, and engage people effectively. And if you want to build agents rather than read about them, the next Agentic AI Cohort opens first to the waitlist, with 20% off.

Reply with one line on what you are seeing in your own work. I read every reply.

— Andreas