Lead, Follow Or Get Out of the Way

THE NUMBER: 2.2%. That’s the share of American households that pay for AI. Not 2.2% of some laggard segment — 2.2% of everyone. Around it sits a whole stack of numbers that say the same thing: 0.2% of households spend more than $100 a month on it, 1% of adults personally pay for Claude, 6% use Claude at all, 4.5% have ever had an agent finish a single task, 8.3% of workers say AI lets them do work they couldn’t do before. And in the same economy, the frontier labs are posting revenue growth that would make a 1999 CFO reach for the smelling salts. Both things are true at once. The entire issue is about the gap between them, and about the fact that the gap is not a model problem. It’s a management problem, and it’s the one nobody’s been willing to name.

Ted Turner reportedly kept a sign on his desk: Lead, Follow, or Get Out of the Way. The line gets credited to Patton, to a Confederate general, to half a dozen people who never said it, which is fitting, because it’s one of those phrases that sounds like it’s about courage and is actually about traffic. It’s a sorting instruction. There are three legitimate places to stand when something big is moving through, and exactly one illegitimate one: in the middle of the road, waving your arms, adding your small personal touch to a thing that was going to happen anyway.

I spent a chunk of my career on Wall Street running IPO pitches. Taking companies public. And I want to tell you what those meetings actually were, because it explains the whole AI economy better than any benchmark.

We’d show up with a sixty-page deck. Junior analysts had been up for three nights on the footnotes, the comps, the valuation football field, the little gradient shading on the bar charts. And my pitch, every time, was to put the deck down. I’d tell the founder: none of this matters. Your company is investable or it isn’t, and the market — not our charts — is going to set the price. We can flatter you with targets all afternoon, but you’re the one who has to go tell the story and make people want it. So here’s the only real question. This is a long road, from this room to the IPO and years past it. Who do you want to have dinner with? Who do you want on the plane? Who’s picking up the phone at eleven at night when you call with a problem? Because the relationship is the entire decision. Everything else is production.

Nothing in those sixty pages ever moved an outcome. The dinner did. And here’s what took me twenty years to say out loud: the deck was slop even then. It was human slop, lovingly formatted, and it was worthless the whole time. We just couldn’t see it, because it was expensive to make, and expensive things feel valuable.

AI didn’t make the deck worthless. AI made it free. And free is what finally rips the cover off.

🪑 The Empty Party

So back to 2.2%. Because when you actually look at adoption instead of the headlines, the party is empty.

We are two-plus years into the most hyped technology since the browser, and roughly one in fifty American households has decided it’s worth a monthly credit-card charge. Bank of America, sitting on aggregated spend data for tens of millions of accounts, puts paid AI penetration around 2 to 3% of households. One percent of adults pay for Claude specifically. Four and a half percent have ever, even once, watched an agent complete a task start to finish. These are not the numbers of a technology that has arrived. These are the numbers of day one.

And yet — Anthropic went from a $9 billion revenue run-rate at the end of last year to $47 billion by late spring, and is walking into an October IPO at a valuation near a trillion dollars. OpenAI, Google, the whole frontier is compounding. So which is it? Is AI a nothing that 98% of households ignore, or is it the fastest-growing enterprise product in history?

It’s both, and the reconciliation is the most important thing I can hand you today. That record frontier revenue is not proof the frontier is winning. It’s the tax every company pays for a harness that doesn’t exist yet.

We laid the groundwork for this on Sunday the 13th, in The Man Behind the Curtain: 89 cents of every enterprise AI dollar still flows to the closed frontier, and it climbed there not because the frontier won the argument but because companies want to leave and can’t work the mechanism that would let them. Every analyst on the bleeding edge prescribes the same rational answer — route the easy 80% of your work to cheap open models, save the expensive frontier for the hard 20%. On the whiteboard it’s unimpeachable. In the building it’s nearly impossible, because the router that does that automatically isn’t a product you can buy on a Tuesday, running an open model safely on your own hardware is an engineering job most companies can’t staff, and managing a swarm of agents is a discipline almost no management team has. So they overpay the frontier. Not because it’s better. Because it’s the only thing that answers the phone.

Read the revenue that way and it inverts. The frontier’s fat margin is borrowed against the non-existence of the management layer. The day somebody ships a real router — a neutral console that makes your context portable and routes every task to the cheapest model that clears the bar — that revenue growth slows hard. And this weekend handed us the evidence that the day is coming: Moonshot’s Kimi K3, 2.8 trillion parameters, open weights shipping July 27, took the top spot in front-end coding and matches Claude Opus on measured intelligence. When the free model is at parity, the only thing propping up frontier economics is the missing harness. That’s not a moat. That’s a fuse.

🧰 Why Coding Went First

Here’s the question that cuts through all of it: if AI is so powerful, why has exactly one kind of work actually moved?

Because coding has moved. That part is not hype. Cursor, Claude Code, Replit, Codex — engineering got genuinely rewired, and it happened fast. But look at the rest of the economy and be honest: has a single large company fundamentally changed its business because of AI? Not trimmed a support budget. Changed. I can’t name one. What I can name is better code, and a lot of increasingly convincing videos of Will Smith eating spaghetti. Those are the two frontiers advancing fastest, and it is not a coincidence that they sit at opposite ends of the same spectrum.

Coding didn’t go first because AI is smartest at code. It went first because code showed up with two things nothing else has. One, the work already lived inside a harness — the IDE, the repo, version control, the whole apparatus an engineer already worked in. Two, and this is the one that matters, code can check itself. A test passes or it fails. A build compiles or it doesn’t. There is a cheap, automatic, yes-or-no oracle for “did this work,” and it runs in seconds without a human.

Now go look at the deck, the memo, the marketing plan, the diligence file. No harness. And worse, no oracle. There is no unit test for a pitch. The only quality signal is a senior person picking it up, squinting, and adding a small dollop of judgment. That judgment is the verification layer — it’s just an expensive, non-scalable, carbon-based one.

Which turns the whole “will AI transform my industry” question into something concrete and testable: your industry moves the day you can build a cheap oracle for whether the work is good. Where you can write the test, where “done” and “acceptable” can be specified and checked, the agents pour in. Where you can’t, where being right matters and there’s no automatic check, everything stalls no matter how smart the model gets. And notice where generative video sits: it needs no oracle at all, because there’s no “correct,” only vibes. So the two poles race ahead — perfect oracle and no oracle required — while the entire valuable middle, the part of the economy where being right has consequences, sits stuck. That middle is where all the money is. It’s also exactly where nobody’s built the harness.

🧾 The Bill Nobody Budgeted For

And here’s the twist that should end the “just add more AI” reflex: even the 2.2% who show up can’t make it pay.

McKinsey ran a FinOps survey this month and the numbers are brutal. 93% of enterprise AI teams are over budget. 60% of agentic spend goes to “response refinement” — the model checking and re-checking and regenerating its own work. One in five organizations has already pulled back on AI use specifically because of cost. And this is the part that breaks people’s mental model: the price of a token fell more than 99% in two years, and enterprise AI bills tripled anyway. Cheaper inputs, bigger bills, because agent architectures got hungrier faster than tokens got cheap. Tokens are not the value. Tokens are the bill.

I read a piece this weekend that made the whole thing click, from an engineer who’d published a twelve-metric quality harness for AI agents and then watched a company kill an agent that passed every single metric. A support agent, more accurate than the humans it replaced, green across the board. The CFO put one number on the screen: cost per resolved ticket, including the failed attempts, came to $4.79. The humans did it for $4.20. The agent was more accurate and underwater on every success. Finance killed it, and finance was right. Accuracy was never the survival metric. Cost per successful outcome is, and it was nowhere in the harness.

The same metric runs the other direction, which is the part to sit with. A legal team’s contract-review agent cost eighteen dollars a run and got flagged as the most expensive line in the AI budget — until someone priced the outcome. It caught a genuine risk clause on one contract in nine. Nine runs, $162, against a single missed clause that ran into six figures of exposure. On a cost-per-call dashboard it was the first thing to cut. On a cost-per-outcome basis it was the highest-return workload in the building. Same agent. Same spend. Opposite verdict, and only one of them was connected to what the work was worth.

Put the empty party next to the over-budget survey and you get the real shape of 2026. It isn’t that AI doesn’t work. It’s that operating it — knowing what to route where, measuring the outcome, deciding what a success is worth — is a management competence most organizations simply do not have yet. The intelligence is sitting on the shelf, cheap and abundant. The thing in short supply is the person who can run it.

🪜 No One Fires Themselves

Why is that competence so scarce? Because acquiring it means doing the one thing organizations are worst at, which is disintermediating themselves.

You can hand every executive in America a copy of Clayton Christensen, watch them underline “the incumbent’s greatest strength becomes the source of its failure,” nod gravely at the offsite — and none of them will go back and fire the version of themselves that formats the deck. No one disintermediates themselves. The senior banker’s touch on the slide is part real judgment, part status (“I earned the right to change the font”), and part the thing nobody says out loud: liability. Someone has to sign. You can’t fire the model, you can’t sue it, so a human stays in the loop to be the throat you can choke. Strip out the ego and the accountability, and the actual irreducible judgment is a thin slice. But we protect the whole hundred percent to keep the ten that’s real.

Boris Cherny, who built Claude Code and now runs product at Anthropic, published a map of this exact terrain this weekend that’s worth more than most consulting engagements. He lays out five stages: Stage 0, gated, zero agents, employees locked out while security and legal and procurement each hold a veto and nobody holds a green light. Stage 1, one person and one agent. Stage 2, one person orchestrating ten. Stage 3, a hundred agents in a supervised org tree. Stage 4, AI-native, a thousand-plus agents where executives steer by intent. And his diagnosis of what determines the pace is the sentence of the weekend: cultural debt. Not compute. Not model quality. The gap between what the technology can do and what the organization can absorb.

Here’s the thing that map reveals, and it’s the same insight from a different seat: managing agents is just delegation. And most companies already fail at delegation. The founder who can’t hire people better than himself and trust them to do the work — the growth ceiling a thousand companies hit and never break through — is the exact same person who can’t hand a task to an agent and leave it alone. The re-hiring you’re starting to see, where a company deploys agents and then quietly brings the humans back, isn’t proof the agents failed. It’s proof that who-owns-what never changed. They bolted a new engine onto the old org chart and the org chart won.

Which is why the radical version of this doesn’t come from inside the incumbents. It comes from the kid with three people and no cultural debt, who never built the ego or the liability structure to defend, and just trusts the output. Look at the SaaStr crew running twenty-one agents in production: they moved ten years and 450,000 records off Marketo — a migration every agency quoted at a year and a hundred grand — for $14.28 in model cost, in an hour. Their own build agent, unprompted, killed a $10,000-a-year vendor by volunteering to rebuild it over a weekend. That’s not a bigger company doing AI. That’s a company small enough that nobody had to be disintermediated for it to happen. The incumbent tinkers. The entrant leaps. Same technology, two chairs, and the chair is the whole story.

🍸 The Dinner

We’ve been here before, and the rhyme is worth saying out loud, because it tells you how long this takes and why.

Factories electrified around 1900. And for roughly two decades, the productivity gains didn’t show up. The economist Paul David wrote the classic account of why: the first thing factory owners did was rip out the central steam engine and drop in one big electric motor, then run the same old line-shafts and belts off it exactly as before. Same layout, new power source, no gain. The productivity explosion only came in the 1920s, when a new generation redesigned the whole plant around small motors on every machine — unit drive — which meant new floor plans, new workflows, and new managers who thought in electricity instead of steam. The dynamo was ready in 1900. The management wasn’t ready until 1920. The twenty-year gap was entirely human.

That’s where we are. The intelligence is the dynamo, and it’s here, cheap and humming. What’s missing is the rewired factory and the manager who thinks natively in agents. Until that arrives, it’s a long slog — not because the technology is weak, but because we keep bolting it onto the old line-shaft.

So what survives? Go back to the dinner. The reason “who do you want to have drinks with” was the only real question in that room is the same reason it’s the safe harbor now. The deck was checkable-ish and got automated. The relationship has no oracle and never will, and not because AI can’t fake warmth — it can — but because the whole point of the relationship is accountability. You’re choosing whose reputation is on the line, whose skin is in the deal, who owns it when it goes wrong. That’s the one thing an agent structurally cannot supply, because you cannot hold it responsible. The deck was always free. The dinner was always the job. AI just made the two impossible to confuse.

What This Means For You

Write the test before you point an agent at the work. Coding moved because “done” was checkable. Nothing else is, by default — so build the oracle yourself. Take one workflow this week and write down, in plain language, what “good” and “finished” mean: the exact checks your best person runs by eye. That document is the harness. No oracle, no automation, no matter how good the model gets.

Measure cost per successful outcome, not cost per token. The agent that passed every eval and still got killed did so because nobody carried the failed attempts onto the successes. Take one deployed agent, add up its fully-loaded cost per completed, review-free task, and hold it against what a human costs to do the same thing. If you can’t compute that number, you are flying blind on the only number that decides whether the agent lives.

Reassign ownership from the output to the system. The re-hiring trap is an org-chart failure, not a technology failure. Stop putting a human on every deliverable and put one named human on the whole pipeline — the owner of the system that produces the work, with the authority to stop it. That’s Cherny’s ladder in one move, and it’s the delegation muscle your company probably already needed.

Sort your own work into the deck and the dinner. Everything you produce is now one or the other. If it’s the deck — the formatting, the first draft, the slop that was always slop — get out of the way and let the machine make it for free. If it’s the dinner, the judgment and the relationship and the accountability, that’s not a task anymore; it’s your whole job. Go be excellent at the part that was always the point.

Three Questions We Think You Should Be Asking Yourself

• Can you write the test for your own best work? If you can specify what “good” looks like clearly enough that a machine could check it, your job is about to change fast. If you can’t — if the quality lives in a senior person’s gut — then you’ve found both your moat and your bottleneck. Encoding that judgment is the hardest, most valuable work in your building, and the person who can do it is the person it threatens.
• When you added your touch to the last thing you reviewed, was it judgment, or was it habit? Be honest about the dollop. Some of it is real and irreplaceable. Most of it, if you’re like the rest of us, is ritual and status and the comfort of having handled it. The middle of the road is a crowded place, and the toll is your afternoon.
• Are you leading it, following it, or standing in front of it? All three of the first are fine. Direct the machine, or trust it and ride, or step aside and let it run. The only losing move is the one almost everyone is making right now: standing in the road, hand-formatting the deck, adding a small dollop of judgment to work that was already done, and calling the traffic jam a job.

---

The deck was never the point. It took an intelligence that could produce the deck for free, in seconds, to prove what was always true: the sixty pages were slop, and the dinner was the whole decision. When the machine can do everything that was production, the only thing left to you is the thing that was never production at all.

Lead it, follow it, or get out of its way. And then go be human.

---

— Harry and Anthony

Signal/Noise by CO/AI is published most weeknights from New Canaan, Connecticut. The point is to make you the smartest person in the room without taking more than fifteen minutes of your morning. If we did, forward it to one person. If we didn’t, hit reply and tell us why.

---

Sources

• Ole Lehmann — the 2.2% adoption stack — @itsolelehmann, Jul 19, 2026 (2.2% of households pay for AI; 0.2% spend >$100/mo; 1% of adults pay for Claude; 6% use Claude; 4.5% have had an agent complete a task; 8.3% of workers say AI lets them do new work)
• Bank of America: ~3% of U.S. households pay for AI · Thrumos — AI subscriptions ~2% penetration, 155% growth · CBS News — who pays for AI subscriptions
• Boris Cherny’s Steps of AI Adoption: A Roadmap — Shelly Palmer, Jul 19, 2026 (Stage 0 → Stage 4; “cultural debt determines the pace”; the 2026 longitudinal study finding 82% of engineers doing “supervisory” work)
• Your AI Agent Passed Every Eval. Finance Still Killed It. — Towards Data Science, Jul 19, 2026 (cost per successful outcome; the $4.79-vs-$4.20 support agent; the $18/run legal agent at $162 per caught clause)
• 60% of agentic AI costs go to response refinement; 93% of teams over budget — MarketScale, on McKinsey’s Enterprise AI FinOps survey, Jul 19, 2026 (“tokens are the bill, not the value”; 1 in 5 orgs already constrained by cost)
• Claude Became Our AI VP of Product; We Moved 10 Years Off Marketo for $14 — SaaStr, Jul 17, 2026
• China’s Moonshot AI releases Kimi K3 — VentureBeat · Axios, Jul 16–17, 2026 · The State of Open Source AI V1.0 — Jul 2026 (open at parity on coding; majority of production tokens now open; deploys hard — 51% vs 63% reach production)
• Anthropic just made a move that changes the AI investing story — TheStreet/Yahoo, Jul 16, 2026 ($9B → $47B ARR; October IPO target; ~$965B valuation) · Meta in talks to rent AI infrastructure to Anthropic — CNN, Jul 17, 2026
• Paul A. David, “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox,” American Economic Review, 1990 (factory electrification and the ~20-year management lag)
• CO/AI prior issues this builds on: The Man Behind the Curtain (Jul 13 — the 89% closed-frontier figure and the harness-is-the-product thesis); A Dollar Fifty in Late Charges (Jul 15 — desire and discernment as the scarce skills); I Bought the Airline (Jul 17 — doers vs. performers)