Lead, Follow, Or Get Out Of The Way

The Dinner, Not the Deck

Only 2.2% of American households pay for AI. Not 2.2% of some laggard group — 2.2% of everybody. One percent of adults personally pay for Claude. Four and a half percent have ever had an agent finish a single task. And the frontier labs are still growing revenue at a clip that would make a 1999 CFO blush. Both things are true, and the space between them is the whole story.

Here's the tell most people miss. That record frontier revenue isn't proof the frontier is winning. It's the tax every company pays for a harness that doesn't exist yet. The consensus fix — route the easy 80% to cheap open models, save the frontier for the hard 20% — is right on the whiteboard and nearly impossible in the building. The routers aren't real. Agent management is a job most shops aren't staffed for. So they overpay the frontier because they can't work the machine that would let them leave. That margin is borrowed against the non-existence of the management layer. The day a real router ships, the growth slows. Same event.

So why has coding moved and nothing else? Not because AI is best at code. Because code came with two things nothing else has: an editor the work already lived in, and a test that tells you, yes or no, whether the thing worked. Everything else — the deck, the memo, the marketing plan — has no harness and no yes-or-no. It has a senior person who eyeballs it and adds a small dollop of judgment. Encode that judgment into a test and the ball moves. Nobody wants to, because the person who can write the test is the person the test replaces. No one disintermediates themselves. You can read every word of Clayton Christensen and still not fire the version of you that formats the deck.

I ran IPO pitches on Wall Street for years. Sixty-page decks, junior analysts up all night on the footnotes. My pitch was always the same: none of this matters. The market sets your price, not our charts. The only real question is who do you want to have dinner with for the next ten years, who's picking up the phone when you call at eleven at night. Nothing in those sixty pages ever moved a decision. The dinner did.

AI didn't make the deck worthless. The deck was always worthless — AI just made it free, and free is what finally exposes it. What's left when the slop goes to zero is the only thing that was ever scarce: the person on the other end of the phone. So lead the machine, follow it, or get out of its way. Just stop hand-formatting the deck and go be human.

At COAI today: the full Signal/Noise — why the management layer, not the model, is the thing you're actually short — is live at getcoai.com.

Outsider Labs Advisory →

2.2%. That's the share of American households paying for AI — not of some laggard segment, of everyone. One percent of adults pay for Claude; 4.5% have ever let an agent finish a task. Meanwhile the frontier labs post record revenue. That's not a contradiction. It's the toll companies pay because they can't yet work the machine that would let them leave. We aren't late to AI. We're on day one, and most of us are standing in our own way.

Write the test before you point an agent at the work. Coding moved first because "done" was checkable — a passing test, a clean diff. Nothing else has that by default. Pick one workflow today and write down, in plain English, what "good" and "finished" mean: the exact checks your best person runs by eye. No oracle, no automation. That one document is your harness.

Name one owner per workflow, not per output. The reason teams quietly re-hire after deploying agents is that nobody changed who owns what. Boris Cherny's adoption ladder is blunt about it: a human owns the system that produces the work, not every piece it produces. Today, take your messiest agent and put one name on the whole pipeline. Move the throat-to-choke up a level.

Price cost per successful outcome, not per token. A support agent passed every accuracy eval and finance still killed it: $4.79 per resolved ticket against $4.20 for a human. Today, take one agent, add up its fully-loaded cost per completed, review-free task, and hold it against what a person costs. If you can't compute that number, you're flying blind on the only one that matters.

Sunday, July 19

— Harry and Anthony

Sources: