Has AI made you more money?

Hey friends.

The AI OS letters are back.

I paused the letter for two weeks. I recharged, spent real time with family and friends, and I didn't want to send something just to send something.

The AI world didn't pause with me, of course. A wave of very capable Chinese open-weight models landed this summer. And Anthropic started watermarking Claude's output, which interests me more (that story deserves its own letter).

The break did what breaks are for. It let me zoom out.

Over the last year, I wrote a book, published six research papers, and helped dozens of companies rethink their business for the age of AI: not to do AI for the sake of doing AI, but to integrate AI where businesses can and teams can actually thrive.

It’s a harder problem than just “AI adoption”, and it has made me think about the global impact of AI on value creation.

What follows is my best current thinking, compressed: a broader look at how AI is actually hitting the economy, and why it doesn't match the version we were sold.

 

Somewhere in the last three years, we started acting as if AI had canceled every other form of leverage, and that single mistake explains most of the disappointment.

 

Business levers didn't go anywhere

Leverage is an old idea. It's the gap between what you put in and what you get out. An hour of input, a hundred hours of output. That gap is where every fortune ever built came from.

(Unfortunately, we never teach leverage in school or university. That’s why, I think, most poor people stay poor, and rich people become even richer.)

A founder I respect made this point about leverage recently, and it stuck with me because it matches what I see in client work every week.

AI is a new lever. A real one, maybe the most accessible one ever handed to ordinary people. But the old levers didn't stop working the day ChatGPT launched.

Capital still compounds. Money placed well still earns while you sleep, same as it did in 1980. No model changed that.

Distribution still compounds. This letter costs me the same effort whether 100 people read it or 15,000 do. One good piece of writing, one good product page, one good talk keeps paying you years after you made it. That math survived AI untouched.

Teams still create leverage. The irony: the frontier labs telling you agents will replace the workforce employ thousands of people! And they're ACTIVELY hiring. Reaching out to people.

If teams were obsolete, the companies building the replacement would be three founders and a GPU bill.

And the highest lever available to ANYONE, still, is a good decision. Deciding that a whole pile of work doesn't matter beats automating that pile perfectly.

Because automating a non-priority gets you exactly one thing: a faster non-priority.

AI alone can’t help with capital or distribution. It can’t replace teams, because teams are not only tasks but a bundle of tasks and ownership. But if used properly, it could help with decisions and focus on what actually moves the needle. Most of the time, though, it creates more noise than focus.

Many people hoped that vibe-coding an app could build them a fortune… Nobody I know has really made money from vibe-coding an app, or anything created with AI, without the other forms of leverage.

 

Scarcity is a healthy filter

Julien is the most AI-maximalist person I know. Every new model on launch day. Every workflow rebuilt monthly…

He’s a prophet of the one-person, one-billion company.

We caught up this summer. He runs twelve automations now. Agents drafting outreach, agents summarizing calls, an agent that watches the other agents. His token bill has climbed all year.

(Despite an unbelievable amount of token and model optimizations…)

His revenue is exactly where it was in January.

When you look closer you notice something: every automation sits downstream of a sale. Nothing sits upstream of one.

And the real bottleneck in the business is not automation, but distribution…old, fundamental, boring business levers.

I see this in client work every week. Julien is just the cleanest example. So let me say the real thing plainly:

Scarcity was doing strategic work in your company, for free.

When capacity was limited, you had to choose. The low-value report died before anyone wrote it. The nice-to-have integration stayed a nice-to-have. Nobody called that strategy. It was just what limited hours forced you to do.

Now capacity is nearly infinite, and the choosing stopped. The new hours don't go to the constraint. They go to everything you used to skip.

And you were right to skip it.

Julien isn't stupid. He's disciplined, technical, relentless. He did exactly what the discourse told him: automate everything you can. Nobody told him "can" and "should" separated a while ago.

The numbers say he has company.

MIT looked at enterprise AI pilots: tens of billions spent, and 95 percent showed no measured impact on profit and loss. The projects ran. The demos impressed. The money never showed up where money gets counted.

IBM asked two thousand CEOs. Only a quarter got the return they expected.

And the study I keep coming back to followed 25,000 workers in Denmark, in the jobs most exposed to AI. It saved them about three percent of their hours. Earnings didn't move. Hours didn't move. Nobody's.

The saved time? Workers quietly keep a good chunk of it. Some goes to life, some to balance, some to the “side AI projects”. Almost none comes back to the employer or business.

Which, honestly, fair.

Now hold it together. Massive adoption. Real time savings. No P&L moved, no paycheck moved. That's a lever pointed at the wrong thing, at economy scale.

 

This has happened before, almost beat for beat.

When factories got electricity, the productivity revolution didn't come. For decades the numbers barely moved, and economic historians went looking for the reason. What they found is almost funny: factory owners had unbolted the steam engine, installed an electric dynamo in the same spot, and changed nothing else. Same layout, same workflow. New lever, old factory.

(That's exactly the kind of AI adoption I see in every business right now.)

The gains arrived a generation later, when people finally redesigned the floor around what electricity made possible. Small motors everywhere. Machines arranged by workflow, not by distance to the power source. The factories rebuilt around the lever. Not the lever around old factories.

The lever was there. The reorganization was the actual work.

A century later the economist Robert Solow looked at the computer boom and quipped that you could see computers everywhere except in the productivity statistics. Same lesson, new machine.

We're in the dynamo years of AI. Bolting a model where a person used to sit, keeping everything else identical, and wondering why the P&L won't move.

 

Two more problems with the AI lever

There's a deeper layer under all of this, and it's the one I wrote about before the break: the verification tax, Karim's calendar full of review, the seniors turned into proofreaders. Production went to nearly free. Ownership didn't move at all. Every artifact a model generates still needs a human willing to put their name on it, and that human's attention is as scarce as it ever was. You can generate the artifact. You can't generate the liability. A lever that produces things nobody can own isn't leverage.

And the lever itself has a supply problem. Models learn from human output, and human output is thinning exactly where the models feed. Stack Overflow, the site whose answers helped train the coding models you use daily, now gets about 99 percent fewer questions than at its peak. Fewer than when it launched in 2008. The models ate the commons that made them, and clean human data is turning into the scarce, expensive asset of the whole supply chain.

I wrote about this earlier this year (cf. #68 AI is committing suicide).

 

Truth

Nobody in this industry wants to say the next part plainly, so I will.

The constraint on your business is probably not solvable with AI.

For most businesses, the constraint is demand, or conversion, or the quality of the offer, or the quality of the decisions. AI helps at the edges of each of those. It perfectly solves none of them. There is no prompt for "make people want what we sell" or “bring more attention to our products please.”

You can verify this without reading a single study. If AI solved the real constraint of most businesses, then three years in, with hundreds of millions of users, everyone using AI would be making a ton more money. They are not.

I can hear the objection, because I'd raise it too: maybe everyone else is holding the lever wrong. True. I've helped dozens of clients turn AI into real margin, and every one of them started from the constraint, never from the tool.

But three years have proved that it doesn’t happen “naturally”: just bring AI in, and somehow magically, people will start solving problems and generating more value.

That’s not going to happen.

It’s a lot of work.

So here's the working test. Answer it honestly, right now.

Has AI made you more money?

If yes: you found a constraint the lever actually touches. Push harder there.

If no: you're probably pointing it at the wrong thing. The model is fine. Your prompts are fine. The pointing is the problem.

And that stings, because pointing is the one job you can't delegate. A model will execute whatever you hand it, brilliantly. It cannot tell you that what you handed it doesn't matter.

 

What to do Monday morning

Pick one. One is enough to change your quarter.

Write your constraint in one sentence. "Revenue is limited this quarter by ___." Demand, conversion, offer, decision speed, something else. If you can't fill the blank, that's the real work, and no new automation gets built until you can.

Audit every automation against the blank. List everything AI does for you today. Next to each item, write what it would cost you to simply not do it at all. Where the honest answer is "nothing," kill it. You'll be surprised how much of your stack is faster non-priorities.

Put revenue next to the token bill. One chart, two lines, the last six months. Ten minutes of work. If one line climbs and the other doesn't, the lever is pointed at the wrong thing, and no vendor on earth is going to tell you that.

Give the saved hour a job. The saved time evaporates by default; the research above shows it plainly. So when AI frees an hour, assign it, in the calendar, to the constraint from step one. Unassigned saved time is where AI ROI goes to die.

 

Wrap Up

Most value comes from subtraction, not addition.

Keep constraints to shut down what doesn’t really help you make progress.

Shut down useless “automations.” They fool you because they work beautifully.

AI is an amplifier, not an equalizer. It multiplies whatever you point it at, priorities and non-priorities alike.

Pointing is still your job.

Have a great weekend.

Stay sharp.

— Charafeddine (CM)