Grown, Not Designed

First principles · Gall’s law

Working things are grown, not designed

A complex system that works has, as far as anyone has found, started as a simple one that worked. AI makes it nearly free to design the complex one from scratch, and that does not repeal the rule. It makes a simple thing that already works the scarcest thing you own.

A design is a prediction. A working system is a record.

You can generate the first. The second accrues one live case at a time.

  • Designed wholecomplete plan→first live week→patches that do not hold
  • Grownsmall thing that works→one step→still works→one step→complex, and works
The top path has every feature on day one and has met no live cases. The bottom path has met every case it has, because it was working the whole time it was growing.

The new system went live on Saturday. It has a dashboard, a mobile app and forty settings, and it was built in a fraction of the time the last one took.

At 8:14 a client calls to move her job to the afternoon. The new system cannot do it without cancelling and re-booking, which sends her three emails. Someone opens the old spreadsheet, the one with a tab called DO NOT DELETE, and moves the job in four seconds.

By noon half the office is back in the spreadsheet. Nobody decided this. It is just the thing that works.

The rule

In 1975 a pediatrician called John Gall wrote a short, funny book about why systems fail. Its most quoted line has held up better than most serious management theory: “A complex system that works is invariably found to have evolved from a simple system that worked. A complex system designed from scratch never works and cannot be patched up to make it work. You have to start over with a working simple system.”

Here is why it holds, in plain terms. A design is a prediction about every case the system will meet. The world then supplies cases the designer did not picture: the client who calls at 7 a.m., the invoice with two addresses, the holiday that falls on a Tuesday. Each one is small. Together they are most of the work. A system that grew has met these cases one at a time, while it was small enough to adjust without breaking, and each fix was tested by the next working day. A system designed whole meets all of them at once, in its first week, while everything depends on everything else. Fix one and you break two.

The spreadsheet is not working because it is clever. It is working because it has already met Tuesday.

That ugly old spreadsheet is a record of cases met. Every odd column, every tab someone added at 6 p.m. on a bad day, is the shape of a live case that arrived and got handled. It looks like a mess because the world is a mess, and it is the only description of the world that has been checked against the world.

What AI changes

AI changes one input: the cost of the design. A full system, screens and database and rules, can now be produced in an afternoon. The thing that used to take a team a year, and so happened rarely, can happen whenever someone is frustrated enough to ask.

The obvious reading is that AI beats Gall’s law, because you can design so much more, so much faster. That reading gets the rule backwards. Gall’s law was not about the cost of designing. It was about the cost of meeting the world. AI makes the prediction cheap. It does not make the cases arrive faster. Clients still call when they call. Invoices still come in odd. Tuesday still comes once a week.

So two things happen at once. First, the from-scratch rewrite gets much cheaper to start and no more likely to work, which means there will be a great many more of them. Second, the cheap design turns out to be most useful in the place Gall pointed to: growing a simple working thing in small steps, very quickly, each step checked by a working day. AI makes the step tiny and the iteration nearly free. Used that way, it does not fight the rule. It runs the rule faster than any team could.

This sits beside an earlier piece of mine, The Part That Cannot Be Split, about waits that no amount of parallel work can shorten. That piece was about how long a job takes. This one is about whether the thing works at all. The two meet in one place: live cases arrive at the world’s pace, not the machine’s.

You asked your AI to replace the old spreadsheet. It did not. It built a copy that runs beside it, quietly, and handles every job a second time in the background.

This morning it reports: last week the copy matched the spreadsheet on all but two jobs, and here are the two, and here is why. From today it takes the Tuesday jobs in earnest. The spreadsheet stays open. Nobody in the office notices anything at all.

What I expect

  1. Rewrites get cheaper, not likelier. The number of whole-system replacements started will climb sharply, and the share that are quietly abandoned in their first months will not fall. Many will end the way the Monday scene ends: with people back in the old thing by noon.
  2. The old working thing becomes the asset. The spreadsheet, the paper form, the procedure only one person fully knows will stop being treated as embarrassments to replace and start being treated as the seed to grow from. The best AI work in ordinary businesses will look like wrapping and extending what already works, not swapping it out.
  3. Side by side becomes the normal way to change anything. Running a new system next to the old one used to double the labour, so it was skipped. AI makes the shadow run nearly free. Cutovers will become non-events, done one slice of live work at a time, and a big launch day will start to read as a warning.
  4. Launches shrink and growth speeds up. The winning products and internal tools will ship smaller first versions than before and change every day, because the cheap part is now the step, not the plan. Speed will be measured in how many live cases a thing met this week, not in how many features it had on day one.
  5. “Cases met” becomes the thing you ask a vendor. When anyone can show you a finished design, the question that tells you something is how many working days, paying customers and odd exceptions the thing has already handled. That record cannot be generated. It can only be accrued.

Where I am standing

When I wrote Own Your CRM, the exciting part was that a contractor can now build the database they need with Claude and Supabase in days. The part that mattered more was the dull part: mirror it into the old system for a year so nothing breaks, and only then cut over. I did not have Gall’s words for it then. That year is not caution for its own sake. It is the time it takes for the new system to meet the cases the old one already knows.

The claim, and how it fails

Already true
A working app with screens, a database and rules can already be generated in an afternoon, and the first live week of a new system still turns up cases nobody planned for.
What has to happen
Design has to keep getting cheaper while live cases keep arriving at the pace the world sets, and people have to notice that the gap between a finished design and a working system did not close.
Where I am probably wrong
If systems generated whole start routinely working in their first live month, because simulated cases turn out to stand in for live ones well enough, then Gall’s law described the limits of human designers and not the structure of the world, and this piece is wrong at its root. I will know by the end of 2028. If by then the usual story of an AI-built replacement is a clean launch day that holds, I was wrong.

Back to Monday

It is 8:15. The job has been moved, in four seconds, by a spreadsheet nobody would design today. Before you close the tab, look at it differently. It is not the thing that failed to keep up. It is the only thing in the building that has met every Tuesday so far.

Grow the new one from there.

Background

Related: The Immutable Past · The Part That Cannot Be Split · Own Your CRM

Author: John Rector

John Rector is a Charleston-based entrepreneur, author, and AI strategist. He co-founded E2open, the supply-chain software company acquired for $2.1 billion in 2025, and in 2026 opened Charleston AI, a 3,000-square-foot lab that helps people and organizations understand and use artificial intelligence. He is the creator of The Reality Equation — a lecture series, book, and curriculum exploring attention, prediction, and how reality is experienced — and the author of more than two dozen books. He writes and speaks widely on artificial intelligence, attention, and the future of human work.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.