Everyone wants a magic wand.
I have lost count of the AI conversations I've had with business people this year. They differ in the details, but underneath they are all the same conversation. Nobody wants to understand AI. They want to wave a magic wand and watch their workload disappear.
I don't say that as criticism. It is a completely reasonable thing to want. It is what every good tool has ever promised, and these people have businesses to run.
And here is the uncomfortable part: the wand is real. You can watch a demonstration today that would have looked like sorcery three years ago. The demos work. That is exactly what makes this moment so confusing, because between owning a wand and knowing how to make one sits all the work nobody talks about.
What happens without the factory
I can tell you what unmanaged AI looks like inside a business, because I have watched it, and I have caught myself doing it. Everyone builds their own little heap. The heaps get in each other's way. Data ends up all over the place, and the AI has no way of knowing which of it is sensitive, which is critical, and who should be allowed to touch what. That sorting is a job for people. And people mostly don't do it, because it is a layer of complexity on top of everything else they are already carrying.
Then the ground moves. These tools change every week. Building on them feels like building something stable on shifting sand. If there is no structure underneath, no agreement about how the thing is used, where knowledge lives, what gets checked, it doesn't degrade gracefully. It turns into a mess with a delay on it.
A boring example, on purpose
For years I wanted a proper bookkeeping system. I attempted it a number of times. It never happened. Not because I'm stupid, but because I wasn't skilled enough in enough areas at once, and there was never a quiet month to learn them. This year, I built it.
Here is what it actually is. Invoices arrive by email, to two mailboxes, because after years of asking I still cannot get every supplier to use the dedicated one. An automation watches both. When a message with a PDF arrives, it checks the sender against a suppliers register. If it knows the sender, the PDF is stripped out, filed, and renamed to our convention: company, invoice number, month and year. A script then checks each PDF really is an invoice, archives it, and passes it on to where the receipts are collected.
Notice something about that list. Almost none of it is AI doing anything in the moment. It is an automation. It is a script. The script was written with AI. I am not a coder, and it is the best code I could ever afford. But I can read it, and it does the same thing every time. Repeatable, checkable steps. That distinction matters more than any demo you will watch this year.
AI doesn't learn. Systems do.
Here is the other thing the demos skip. AI does not learn. Its knowledge is encoded, and it operates from that knowledge. When it makes a mistake, there is no loop. It will not be wiser tomorrow. People learn from mistakes. Models don't.
But systems can, if you build the rule. In mine: if the system names an invoice correctly three times in a row, it has earned the right to handle that supplier automatically. One flagged mistake, and it starts earning again from scratch. That isn't intelligence. It's governance. It is also the only kind of learning I am prepared to let loose on my company's books.
And then, the payoff. Late one night I found a difference between what the accounts said and what the bank said. The old way of finding such an error is painful: go back to a point, months earlier, where everything still reconciled, then walk forward through every entry until it doesn't. Half a day if you are lucky, and it wants your sharpest hours, which by then you don't have. I handed it to the system. Five minutes later it had found the problem.
But be careful about what those five minutes rested on. If the bookkeeping is not precise and not consistent, your reports do not exist. The wand only worked because the factory was already built.
The river and its bed
There is a question you may be asking at this point. If the factory is the hard part, can't you just buy someone else's? Take a system that already works somewhere, drop it into your business, wave the wand?
I keep coming back to a television programme I watched a long time ago. It asked whether you could freeze a person and wake them a hundred and fifty years later, alive and well. Technically, the argument went, perhaps you could: slow the processes of life until they nearly stop. The real problem arrives at the waking. The brain that wakes was built by and for a world that no longer exists. The programme used an image I have never forgotten: a river and its bed. A river runs the way it runs because of the bed it carved, and the bed carved the river back. Move the river a hundred and fifty years forward and there is no bed that fits it. Nothing is broken. It simply doesn't fit.
Brains do this from the very beginning. As far as I understand it, we carry the most connections we will ever have at around the age of two: a buzzing jungle of wiring that doesn't yet make sense of anything. Then the environment gets to work, day after day, keeping what gets used and pruning what doesn't, until what remains fits the world it actually lives in. A brain is not shaped in general. It is shaped by its own circumstances.
A business is the same kind of thing. It is a form of life: it has spent years adapting to its own customers, its own suppliers, its own people, its own history. My bookkeeper works because it fits our bed exactly: the two mailboxes I could never consolidate, our suppliers' particular habits, our naming convention, our appetite for risk. Drop the identical system into another company and you would have the river in the wrong bed. Nothing broken. Doesn't fit.
The same idea reached me just now from a completely different world. Someone was writing about computer chips and the glossy performance charts that come with them. The point underneath was that none of those numbers mean anything unless the tests resemble the work you will actually give the thing. Someone else's proof tells you what fits someone else's circumstances.
So there is no skipping the adaptation. If you want the wand to work for you, you cannot avoid the experience of adapting: the system to your business and, just as much, your business to the system. That always takes time and it always takes effort. The wand at the end is real. But it exists only because of everything that shaped it, like everything else alive.
The four rungs
The adaptation has an order to it. The way I have come to think about it, there is a ladder, and it has four rungs.
Consult is where everyone starts. You talk to it the way you would talk to an expert with a vast library behind them. It reads for you, summarises for you, answers with sources. Most businesses are standing on this rung and believe they are near the top.
Collaborate is when it sits on your own computer. If your computer is Iron Man, AI is the suit. It works your files, tidies, maintains, builds small automations. Powerful, and personal.
Implement is when you start building pieces of systems out of parts AI helped make, like my bookkeeper. The parts are reviewable. The behaviour is repeatable. The intelligence is in the design, not in the moment.
Operate is when the system runs processes itself: it starts the work, moves it through its stages, and comes to a human at exactly the points where judgement belongs. We are building towards this. Nobody sensible skips to it.
Each rung is more delegation. And each is safe precisely in proportion to the governance underneath it: the knowledge base, the rules about what earns automation, the record of what happened, the gates where a person decides.
Why this matters most in production
Building prototypes is easy now, for everyone, to a quality that used to take teams. Which is exactly why prototypes stopped meaning anything. The hard part, the part that decides who wins, is bringing AI into production, where mistakes cost real money.
And the deepest version of that is production in the literal sense: factories, where software touches physical machinery, and where the cost of an ungoverned mistake is not measured in money at all. People will be rightly careful bringing AI anywhere near that world. The ones who can do it, the ones who can show the factory behind the wand, are going to matter a great deal.
The last thing you make
So yes, I believe in magic wands. I use one most days, often past midnight, which is its own warning. But the wand is the last thing you make. First you build the factory: data in the right places, rules that decide what earns automation, records that let you see what happened, and gates where a human decides.
That work doesn't demo well. It never will.
It is also the only part that is actually magic.