Taking clients for Q3
Jason Q. Lu
Menu
A long, almost empty gallery with one man walking through it alone
← All writing
AI for Real Businesses · 6 min

In a room of 2,000 people, one is building with AI

Eighteen in a hundred use AI. Three of those pay. Eight of those build. That is one builder in every 2,000 people, and that is where the leverage sits.

Published 15 August 2026 · Updated 12 September 2026

Eighteen people in every hundred use AI at all. Three in every hundred who use it pay for it. Eight in every hundred who pay actually build with it, wiring models into how the work gets done rather than typing into a chat box. Multiply the three together and you get roughly one person in every two thousand standing on the ground where the leverage is.

That is not a forecast about where this ends up. Nobody credible knows. It is where the room is standing right now, and it is almost entirely empty.

How many people actually use AI?

Three rooms, each one a hundred people drawn from the one before it.

100 people on earth
Use AI 18
Have never touched it 82

That is the first of the three, drawn. The table carries all three:

Put a hundred of these in a roomHow many do the next thingWhat the rest do
People on earth18 use AI82 have never touched it
People who use AI3 pay for it97 stop at the free tier
People who pay for AI8 build with it92 subscribe and use the app

Eighteen surprises most people, because if you work in technology it feels like everyone is using this. Three is the number that should stop you. Eight is the number worth sitting with.

Sources: UN World Population Prospects 2025, DataReportal Digital 2026, ITU Facts and Figures 2025, and reported user numbers from OpenAI and Microsoft. The paid and builder counts are deliberately conservative.

Does it matter which AI model you pick?

Almost not at all. Claude against Gemini, open weights against closed, whoever topped a benchmark this week: interesting, fast moving, and close to irrelevant to what you should do on Monday.

Which model you pick is the least important decision in the whole exercise. What decides the outcome is whether the work around the model gets rebuilt. Models are the cheap part now. The rebuild is the expensive part, and almost nobody writes about it.

Where is your business on this, honestly?

Nearly every business I talk to is on one of three rungs. If you work inside a large organisation rather than running your own, read "your business" as the function you own. The rungs are identical; only the budget conversation changes.

RungWhat it looks likeWhat you actually have
Someone on the team uses itThey paste into ChatGPT between tasksReal value, and it walks out the door the day they do
You pay for seatsLicences, training, a policyA rented tool. The work has not changed shape
The work runs on itThe workflow itself was rebuiltA system. This is the 8

The jump from the second rung to the third is the one that matters, and hardly anyone has made it. It is not a productivity tweak. It changes what you can take on at all.

What are the 8 actually doing?

Three systems from builds I have run this year. Different industries, same underlying shape.

  • Automated growth. A regulated field where the compliance rules run before anything is written, so one recorded conversation becomes a month of publishable work and a person signs off every piece.
  • Automated processes. Every quote scored against defined criteria before it leaves the office, so nothing goes out under-priced or off-standard.
  • Automated operations. Onboarding, deliverables, QA and the sales pipeline all run by agents. One operator, a team of them.

Underneath, all three are the same five steps:

  1. Input. A recording, a job spec, an enquiry.
  2. The rules run first. The standard is applied before the work, not checked after it.
  3. The system does the work. Drafts, scores, files, routes.
  4. A person signs off. Every time. This is the step people skip and the step that matters.
  5. Output. Published, priced, delivered.

None of that is exotic. It is ordinary business process, rebuilt so the system does the work and a person signs it off.

Why start now rather than next year?

Because building got cheap, and the businesses one rung up are quietly changing what they can deliver while the gap compounds.

A platform that needed a product manager, a designer and two engineers for six months last year takes one person about a month now. Roughly twenty to one.

The same buildThe teamThe timeRelative effort
Last yearPM, designer, two engineersSix months24
This yearOne personOne month1

The hard part moved rather than disappeared: planning at the front, sign-off at the back, and a short stretch of building in between. You do not need to be code-native to do this. You need the plan.

The part that surprised me most

Rebuild a workflow properly and you have not only solved your own problem. You have built something other businesses with the same problem would pay for. The engine I built for one business is not specific to that business. It is specific to that problem.

Inside a large organisation the same thing holds, except the other businesses are other departments and the currency is headcount you do not have to ask for.

That is productisation, and until recently it needed a development team and a year of runway. Now it is a version of the thing you already run. It tends to play out three ways:

  • Done for you. You deliver the outcome and the system stays invisible. Lowest risk, closest to what you already do.
  • Enablement. You help others build their own version. The method becomes the product.
  • Software. You license the engine. Highest ceiling, and genuinely a different company.

Then the constraint moves again. If building is cheap and everyone can build, making the thing stops being the hard part and getting anyone to know it exists becomes the hard part. Distribution is the new bottleneck, and it is the one AI helps with least.

Common questions

How many people build with AI rather than just using it? Roughly one in two thousand. Eighteen per cent of people use AI, three per cent of those pay, and eight per cent of those build systems with it.

Is it worth switching AI models? Rarely. Which model you use is close to the least important decision. What decides the result is whether the work around the model was rebuilt, and a well-built system makes changing model a settings change.

What does "building with AI" actually mean? Wiring a model into how work gets done: the rules applied before the work, the system producing drafts or scores or routing, and a person signing off before anything leaves. Not typing into a chat box.

Do I need engineers to start? No. The scarce thing is the plan, not the code. Most of the value in the first ninety days comes from connecting what you already have and measuring it.


Want to do this on your own business? The free course walks through building the first version in about twenty minutes, or book thirty minutes and bring the messiest part of your operation.

This started life as Build Notes Nº 01.