ISSUE 07 · SEPTEMBER 29, 2026
AI is all we do. We keep up so you don't have to.
WHAT WE'RE SEEING IN THE MARKET
The Sentence That Stopped a Room of Founders
We ran a session recently for a strong group of founders. The topics were the market landscape, why a single knowledge base file matters, and how to keep AI grounded in real company context.
Near the end, one founder said it plainly: "This is so overwhelming, I don't know what most of this means."
THE WHY
Nobody in that room was behind
I told him he was right. It's chaos right now, it's the wild west. Nobody in that room was behind. They were seeing the market clearly.
That chaos is also why most companies can't say whether AI is paying off. When everything gets pitched as urgent at once, nobody stops to pick one task, write down how it runs today, and decide what better would look like. So the next review ends with a login count.
That's what an eval is, minus the jargon. It checks whether the AI does one job well, against a standard you set in advance. It answers the first half of the ROI question. The second half is whether doing that job well moved a number you care about, and you can't get to the second half without the first.
YOUR MOVE
Write down what good looks like
You don't have to make sense of all of it. Pick the one AI use case your team leans on most. Before your next review, write down two things: how that task went before AI, and one sentence for what a good result looks like now. Then pull five recent outputs and check them against that sentence.
That's small enough not to be overwhelming, and it gives you a real answer when someone asks what the spend bought you.
WHAT TO SKIP
The eval-platform pitch
It's being sold as the must-have layer right now, but most companies aren't at the point where they'd actually need one yet. One sentence per use case, checked against real examples, gets you further than a platform will this year.
THE ROUNDUP
Signals from the noise this week.
The backwards AI pacing debate and how far business is from the frontier (Fortune)
A useful counterweight to the pressure to keep up. Most large companies are still in the first phase of AI adoption, using assistants on top of tools they already own. If you feel behind, you're closer to the pack than the headlines suggest. Read it hereDetecting and countering misuse of AI: September 2026 (Anthropic)
Anthropic's newest threat report found attackers now go after AI API keys on purpose, not just as a side effect of a breach. If a vendor or team holds a key to your AI tools, handle it like any production credential. Read it hereFrontierMath v2 (Epoch AI)
The team behind one of the hardest AI math tests found errors in 42 percent of its own problems and reissued every score. Worth remembering the next time a vendor leads with a benchmark number you didn't check yourself. Read it here
You just did, in miniature, what we spend a month doing properly. If you want the full picture instead of one data point, that's what the Advisor is built for.
Hit reply and tell me which use case you picked. I'd like to know where people start.
Justin
Why of AI