Why Most AI Spending Returns Nothing

MIT's research group put a number on it that made a lot of people uncomfortable: about 95% of corporate AI spending shows no impact on profit. Boston Consulting Group got there from a different direction — roughly 70% of AI programs miss what leadership expected. Only about one pilot in twenty returns anything you could point at on a P&L.

The usual reaction is to conclude the whole thing is a bubble. I get it. I also think it's the wrong lesson, and the right one is considerably more useful to anyone running a business.

Because that number isn't a verdict on the technology. It's a verdict on what the technology was pointed at.

You cannot automate a process nobody wrote down

Here is the thing almost nobody says out loud at the demo.

Automation doesn't create order. It copies whatever order already exists, exactly, at speed. Point it at a clean, boring, repeatable process and it will do that process beautifully forever. Point it at the way three people currently handle invoices — one of whom does it differently on Fridays, one of whom has a spreadsheet nobody else can open, and one of whom just kind of knows — and you have not automated anything. You have built an expensive machine for producing the same mess faster.

This is why the failure rate clusters where it does. The pilots that work are almost always aimed at something that was already written down. The ones that die are aimed at something that only lives in somebody's head.

And in a small business, almost everything lives in somebody's head. Around 71% of small businesses depend on one or two people for their success — which is another way of saying the operating manual is those two people. That's not a criticism. It's how a business gets to twenty employees in the first place. But it does mean the ground isn't ready, and no tool is going to prepare it for you.

BCG landed in the same place, incidentally: they named culture and leadership as the root cause, not the software. The industry spent two years arguing about which model was best. It turns out the model was never the variable.

The pattern I keep seeing

It goes like this, and it's remarkably consistent.

Someone buys a tool because a competitor mentioned it, or because a vendor was persuasive, or because there's a nagging feeling that not doing anything is itself a risk. The tool gets installed. A few people try it for a fortnight. Nobody has decided which job it replaces, so it doesn't replace one — it just sits alongside the old way, which is still running, because the old way is the one that actually works.

Six months later the subscription is still being paid and the honest internal verdict is we tried AI, it didn't really do much.

Nothing was learned. Nothing was fixed. And now the organisation is slightly more cynical about the next idea, which is the genuinely expensive part.

I've written before about why good recommendations die — it's rarely because the recommendation was wrong. It's because nobody worked out who would actually do it. This is that same failure wearing newer clothes.

What I actually do about it

I start with the process, not the tool. Always, and it's not because I'm being pious about methodology — it's because there is no other order that works.

So I map what actually happens. Not what the org chart says happens. Who touches it, in what order, how long each step takes, where it stops and waits, and what it costs when it goes wrong. That part is unglamorous and it is where nearly all the value hides.

Then every step gets sorted into one of four piles.

Leave it alone. Some things are fine. A process that runs twice a year and takes an afternoon is not a problem, no matter how manual it looks. Automating it would cost more than it saves and I'd be charging you for the privilege. This pile is bigger than most consultants will admit, because everything in it is revenue they're choosing not to take.

Just fix it. A startling amount of what looks like a technology problem is a sequencing problem, or two people doing the same thing without knowing, or an approval that exists because of something that happened in 2019. No software required. These are usually the fastest wins in the whole engagement and they cost nothing but attention.

Buy something. There are genuinely excellent tools out there now — properly good, priced sensibly, solving a real and specific problem. If one of them fits your situation, I'll tell you which one and roughly what it should cost you. I don't take referral fees, which means I have no reason to name anything but the right one.

Build something. Sometimes the tool you need doesn't exist, or the ones that do want three hundred a month for the one feature you'd use. The economics of building software have changed enormously in the last couple of years — things that were a six-figure project are now genuinely not. When the case is there, I build it, and you own it.

Notice that only the last two involve buying anything, and only one involves me writing code. That's the point. If I show up determined to sell you a build, every problem starts looking like one.

The honest version of what it's good at

I'm not a sceptic. Quite the opposite — I think it's underrated by most people using it, which I've argued at some length elsewhere. The gap between casual use and skilled use is wider here than for any tool I've worked with in my career, and most of the disappointment I hear traces back to that gap rather than to the technology.

But "this is powerful" and "this will help your business next quarter" are two entirely different claims, and the industry has been running them together for three years now.

What it's genuinely good at, today, in a real business: anything repetitive with a clear right answer. Reading and sorting documents. Drafting the same category of thing over and over. Watching a data set for the pattern that means trouble. Answering the question your team gets forty times a week. Tedious, well-defined, high-volume work.

What it is not good at is being handed a vague ambition and asked to sort things out. That's still a person's job. It's my job, most weeks.

Where I land

The 95% figure isn't an argument for doing nothing. It's an argument against doing the popular thing, which is to buy first and think second.

If you want a version of this that works, the order is: understand what actually happens, fix what's simply broken, write down what remains, and only then ask what should be automated. Do it in that order and the tools work about as well as advertised. Do it in the reverse order — which is what 95% of the money did — and you get a faster version of a process that was never right.

The unfashionable truth is that the hard part was never the technology. It's the same hard part it has always been: knowing what your business actually does, step by step, and being honest about which of those steps deserve to survive.

That part doesn't have a subscription.