What is a 20x company, and how do you become one?

A 20x company does the work of a company twenty times its size. The phrase comes from GigaML, whose five engineers won a Fortune 500 contract against competitors carrying 500 staff (Ability). The arithmetic that makes it possible is not complicated. One person who directs six AI agents at once produces more than one person who does six tasks in sequence.

The idea is spreading because the examples are real. AI-native startups report $2m to $4m of revenue per employee, against roughly $300,000 at an average public SaaS company (Forbes, Nick Talwar). Cursor passed $2bn in annualised revenue with about 300 people. Lovable reached $100m of annual recurring revenue in eight months with 45.

The obvious conclusion, and the wrong one, is that buying the same tools produces the same result.

Why do almost no companies become one?#

Because the tools were never the constraint.

The National Bureau of Economic Research surveyed close to 6,000 chief executives, finance chiefs and senior executives across the United States, the United Kingdom, Germany and Australia. Adoption was high: 69% of firms reported using AI, rising to 78% in the US. The outcome was not. 89% reported no measurable effect on labour productivity over the previous three years, and more than 90% reported no measurable effect on employment (NBER working paper w34836).

Nearly nine in ten companies bought the technology and measured nothing.

One figure in that survey explains a good deal of it. Senior executives who use AI personally use it for around 1.5 hours a week. That is the leadership of companies expecting a change in how the whole organisation works, spending about eighteen minutes a day on the thing they expect to deliver it.

The failure is not free either. Where AI output is wrong, somebody has to check it and correct it, and that correction time comes out of the saving nobody measured in the first place.

What did the ones who got there do differently?#

They changed how work is assigned, not what software sits on the desk.

An AI-native engineering team is not a team with more licences. It is a team where the default assumption at every step is that AI does the first pass. Requirements get refined with it. Tests get scaffolded with it. Code review runs against it. Architectural decisions get recorded through it. Developers at Anthropic are reported to run between three and eight instances of Claude at the same time, which is a different job from writing code with an assistant open.

That is a habit change, and habits do not change because a tool was purchased. In simple words, the 20x company is an operating model, and the licence is the smallest part of it.

How do you become one?#

Four steps, in order, and the order matters because each one produces the input for the next.

1. Find out where the hours actually go#

Take the two or three jobs each team repeats every week. Watch how they get done now, step by step, and add up the hours. Not a survey of what people think takes long. The actual sequence, timed.

Most teams are wrong about which job is the expensive one. The job everybody complains about is often thirty minutes a week. The job nobody mentions, because it has always been done that way, is six hours.

This is the same exercise described in how to prove AI training worked, and it is here for the same reason: a number recorded afterwards is worth nothing without the one recorded first.

2. Write down the number before you change anything#

If the reporting job takes four hours a week now, record four hours. Record who does it and what the output has to look like to be acceptable.

This is the step that gets skipped, and skipping it is why 89% of firms have no measurable result. They did not fail to improve. They failed to record what things were like beforehand, so no improvement could be demonstrated afterwards, and an improvement nobody can demonstrate does not survive the next budget conversation.

3. Rebuild the biggest job with AI doing the first pass#

One job. The most expensive one from step 1. Rebuild it so the model produces the draft and the person edits, approves and handles the exceptions.

Do this inside the tools the team already runs. A new platform adds a migration and a training burden to a change that was already going to be hard, and the migration usually gets blamed when the change fails.

4. Measure the same job again, and decide honestly#

Same task, same output standard, timed again. If four hours became one, that is the number. If four hours became three and a half, say so, and consider that the job may not have been a good candidate.

Then repeat from step 1 with the next job. Companies reach 20x one job at a time. Nobody arrives there through a company-wide announcement.

What does it cost, and how long does it take?#

The articles that describe 20x companies tend to skip this, so here is the honest shape of it.

Steps 1 and 2 take a day for a single team, and the output is a list of jobs with hours attached. Step 3 takes between an afternoon and a fortnight for one job, depending on how many systems it touches. Step 4 takes as long as the job takes, plus the discipline to run it.

The expensive part is not the software. It is the attention of the people who do the work, for the day it takes to look at it properly. That is why so many companies buy licences instead: a licence can be approved in an afternoon, and a day of a team's time has to be argued for.

What this cannot do#

Two limits, both real, and the second one is the limit worth stating before anybody quotes the 20x figure at a board.

A job that is slow because two teams disagree about who owns it will not get faster with AI. It will get faster at producing the thing that then waits three days for a decision. Where the constraint is an unresolved question about ownership, the useful output of step 1 is that finding, and no amount of model quality substitutes for somebody settling the argument.

And the 20x figure itself deserves care. The companies quoted at the top of this article are young, technical, and were built this way from the start. A fifteen-year-old company with existing customers, existing systems and existing obligations is not going to reach the same ratio by following the same steps, and anybody promising that is selling something. What such a company can reasonably reach is a measured improvement on specific jobs, compounding as more jobs get rebuilt.

The part worth keeping#

The 89% figure is the most useful number in this article, and it is not a warning about AI. It is a warning about buying without measuring.

Every company in that survey had the same tools available as GigaML. The difference was that a few of them changed how the work is done and recorded what happened, and most of them bought access and hoped.

If you want to know which of your team's jobs would survive step 1, the two companies we have run this exercise with measured a 252% and a 217% increase in what their teams got through afterwards. The 217% was the one working towards being a 20x company, across the whole business rather than one department. Both started the same way: one day, the real work, timed before and timed after. That is two companies rather than a study, and your number would be your own, but the sequence is the same one described above.

Start with step 1. It costs a day and it produces a list, and the list tells you whether any of the rest is worth doing. If you would rather not run it yourself, that day is what we sell, and the free version of step 1 is a thirty-minute call.