Everybody Is Talking About AI. Almost Nobody Is Building It.

There is a number from this year that I have not been able to put down.

In the first five months of 2026, companies attributed 87,714 job cuts to artificial intelligence, more than in the whole of the year before, and AI became the single most common reason given. Set beside that, a piece of analysis published in the Harvard Business Review, which looked at what those companies had actually deployed and concluded that they are largely cutting for AI’s potential rather than its performance.

Read those two facts together and sit with what they describe. Companies are letting people go over software they have not built yet. The AI in question is not running. In many cases it has not been specified. It exists as a conviction, a slide, an item on a transformation roadmap, and it is already producing consequences in the world while producing nothing in the operation.

That is the clearest illustration I know of the thing this article is about. There has never been more conversation about AI, and there has never been a wider gap between the conversation and anything actually running inside a business. The talking is not a prelude to the building. For a great many companies it has quietly become a substitute for it.

The spending is real. The systems are not.

Look at the macro picture and the imbalance is almost comic.

The big cloud platforms are spending something in the region of $650B to $725B on AI infrastructure this year, roughly double what they spent last year. Alphabet alone guided to $180B to $190B. Enormous machines are being built, at enormous cost, in a hurry.

Now go and stand in a normal business. A twenty-person clinic, a forty-person contractor, a regional distributor. Ask what AI is actually running in the operation, not what is being discussed. In most of them the answer is that a few people have a chat subscription they use like a better search engine, and that is the whole of it. The trillion-dollar build-out has, so far, reached most businesses as a browser tab.

I could pad this section with the studies about how many corporate AI pilots fail to reach production. There are several, they are quoted constantly, and I have decided not to use them, because the most-cited one turns out to rest on a survey of about a hundred and fifty people, and we have already had to say so publicly. The argument does not need them. You do not need a study to tell you how much AI is running in the businesses around you. You can just look.

Why talking wins

It would be easy to be contemptuous about this and I do not think contempt is deserved, because the incentives are genuinely, structurally tilted toward talking. Once you see the tilt, the behaviour stops looking like laziness and starts looking almost rational.

Talking cannot be wrong. A strategy document is unfalsifiable. An AI roadmap that spans three horizons cannot fail this quarter, because it was never going to do anything this quarter. A working system, on the other hand, either saved the hours or it did not, and the number is right there, and everyone can see it. Building generates evidence, and evidence is dangerous to anybody whose position depends on not being measured.

A pilot is a way to appear to act while deferring judgment. This is the sharpest one. A pilot with a baseline, an end date, and a person who will call it either way is a real experiment and I am wholly in favour of them. A pilot without those things is theatre, and it can be extended forever, and it usually is. It produces status updates. It occupies the slot in the plan where the decision was supposed to go. Nobody has to be right and nobody has to be wrong, and the quarter closes with the item marked in progress.

The conversation is socially rewarded and the work is not. Speaking on a panel about AI transformation is a career event. Sitting with the person who does intake, timing how long they spend re-keying the same eight fields, and writing that number on a whiteboard is not a career event. It is, however, the only one of those two activities that has ever saved anybody any money.

And the vendors prefer it. A strategy engagement is a long, comfortable, high-margin conversation. A build has a defined end, at which point the client can check whether it worked. Guess which one large parts of my industry are configured to sell.

None of this requires anybody to be cynical. It only requires everybody to follow their incentives, which people reliably do.

What implementing actually consists of

Here is the part I think is worth the price of admission, and it is the part that never gets said out loud, because saying it out loud makes the whole thing sound less impressive than it is being sold as.

Implementing AI in a business is not one big mysterious act. It is a short list of unglamorous jobs, and every one of them is legible to a non-technical owner. This is the list.

Pick one workflow. Not a department. Not a function. One repeated thing that happens many times a week and annoys everybody.

Measure what it costs you today. How many times, how many minutes, what an hour of that person’s time actually costs. Write it on paper before anything changes, because a system with no baseline can never be proved to have worked, and the arithmetic is the entire question of whether it pays.

Get the data out of the inbox. Most of what a business would need to point an AI at is currently sitting in email threads, in a spreadsheet on someone’s desktop, and in the head of the person who has been there nine years. That has to become a structured record before anything can read it, and that step is the actual project more often than anyone expects.

Define what a right answer looks like. Take thirty real examples from last month and write down what the correct output would have been. That set of thirty is the difference between a system you can trust and a system you hope about, and almost nobody builds it, because it is boring and it takes an afternoon.

Build it where the work already happens. Inside the tool they already have open. A system in its own tab, requiring somebody to remember to go and use it, gets used for three weeks and then quietly does not.

Decide what happens when it is wrong. Not if. When. Who sees the exception, what they do with it, and how the system learns that they overrode it.

Cap the cost before you turn it on. Budgets, caching, a hard stop, spend you can see in real time, because an unattended loop with no circuit breaker will happily bill you all weekend.

Ship it, then go and look at the number you wrote down.

That is the whole of it. Nine steps, none of them requiring a transformation office, none of them requiring you to have an opinion about the state of the frontier. Read that list back and notice something: almost none of it is about AI. It is about knowing your own operation well enough to point at one thing, and being willing to be measured afterward. The model is close to the easiest part, and it is the part everyone spends the meeting talking about.

Not all talk is stalling

I want to be careful, because the lazy version of this argument is a demand that everybody stop thinking and start typing, and that produces exactly the expensive garbage I spend my life warning people about.

Scoping is not stalling. Working out whether the job is worth doing, whether the data exists, whether a rule would beat a model, whether the arithmetic clears, that is not delay, that is the work. We built our whole first engagement around doing precisely that, and a good share of the time it ends with us telling someone not to build the thing.

The distinction is simple and you can apply it to any AI conversation you are currently in, including one you are paying for. Does this conversation have a decision date? Talk that terminates in a decision, on a date, with a number attached, is scoping. Talk that terminates in more talk is a hobby. A year of strategy that has produced no shipped system and no written-down reason not to build one is not caution. It is the appearance of caution, and it is more expensive than a failed project, because a failed project at least teaches you something.

The honest pitch, once

This is where I say the obvious thing, and I will say it once and plainly, because you have read a thousand words of argument and you are entitled to know what I want.

We are an AI development studio. Implementing is the entire job. We are not a strategy practice with a build option attached, and the reason the list in the middle of this article is written out in full, rather than kept back as proprietary insight, is that the list has never been the hard part. Doing it is. Most of what separates a business with AI running from a business with AI on a slide is not knowledge, it is somebody actually sitting down and doing nine unglamorous things in order.

If you want that done, that is what we do, and our first step is deliberately small and priced so that finding out costs less than another quarter of meetings. If you would rather do it yourself, the list above is the list. It is genuinely all of it. Take it.

The two-week test

Here is what I would actually do if I were you, and it does not involve calling anyone.

Take the workflow that annoys you most. Time it, honestly, for a week. Count how often it happens. Multiply. Then ask one question of whoever is proposing AI to you, in-house or outside: what could we have running, in production, in the next two weeks, on that one workflow, and how will we know if it worked?

The answer to that question sorts everybody in the room instantly. Some people will have an answer, and it will be small and specific and slightly disappointing in its modesty, and those are the people who build things. Others will explain that it is more complicated than that, and propose a discovery phase, and the discovery phase will propose a roadmap, and eighteen months later there will be a very good slide about your AI journey and nothing whatsoever running in your business.

The gap between those two responses is the only gap that matters right now. It is not a gap in technology, which has never been more capable or cheaper. It is not a gap in information, which is free and everywhere and which you have just been handed nine steps of. It is a gap between the people who are still discussing this and the people who have quietly shipped something and are already looking at the number.

Everybody is talking. The talking is very good, and it is getting better, and it costs nothing, and it changes nothing. At some point somebody in your business has to go and build the thing, and the only real question left is whether that happens this quarter or whether you spend another one becoming extremely well-informed about a system you still do not have.