Before you implement AI, talk to someone who isn't selling it to you.

The number they'll quote at you, and what it actually measured.

If anyone has talked to you about AI this year, chances are they dropped the line that "95% of AI projects fail". It's in every deck. And it is a perfect example of the problem, so it's worth looking at closely.

The figure comes from a July 2025 report called The GenAI Divide: State of AI in Business 2025, from MIT's Project NANDA. What it measured was not "AI projects": it was the share of integrated generative AI pilots that had produced no measurable impact on the profit and loss account. The sample: interviews with 52 organisations, 153 senior leaders surveyed, and an analysis of 300 publicly disclosed deployments. It was a working paper, not peer reviewed, written by a group that builds AI infrastructure. The original link hosted at MIT now redirects elsewhere. And a telling detail: much of the press that quoted it reported sample figures different from the ones in the document itself.

To be clear: I'm not saying the report is wrong. Its qualitative conclusions are among the most useful things I've read on this. I'm saying something more uncomfortable: almost nobody quoting that number has read it. It gets repeated because it works on a slide. If your adviser hands you that figure without knowing what it measured, you have just learned how much verification sits behind whatever they recommend next.

What is properly measured says almost the same thing.

And here's the interesting part: you don't need the 95% to reach the same conclusion. There are more solid studies, with a name, a date and a method, and they all point the same way.

Gartner, June 2025: more than 40% of agentic AI projects will be cancelled before the end of 2027, due to escalating costs, unclear business value or inadequate risk controls. A year earlier, in July 2024, it had already forecast that at least 30% of generative AI projects would be abandoned after proof of concept.

RAND, 2024: more than 80% of AI projects fail, roughly twice the rate of IT projects that don't involve AI. Their conclusion about the main cause is the part I care about, because it isn't technical: the most frequent reason is that the people involved never actually agreed on what the project was for.

S&P Global: the share of companies abandoning most of their AI initiatives before reaching production went from 17% to 42% in a single year.

Three independent sources, three different methods, the same story. And notice that none of them says "AI doesn't work". They say the projects that fail are the ones that start with the tool instead of starting with the problem. That distinction completely changes what you should do on Monday.

The question I ask before any other.

When someone writes to me saying "we want to bring in AI", my first question has nothing to do with AI. It's this: if three times as many customers walked through the door tomorrow, what would break first?

The answer nearly always lands in one of four places, and only one of them is AI territory:

Not enough people arrive. By far the most common. AI has no role here. What's missing is visibility: the half-empty Google profile, the site that doesn't show up when someone searches for your service in your area, word of mouth with nowhere to land. I have an entire local SEO guide about this, and it doesn't mention AI once.

People arrive but don't buy. Also not AI. The offer isn't understood, or the price isn't explained, or the contact page asks for nine fields. You fix that by writing better and removing friction, not by automating.

They buy, but you can't keep up. Here it starts to make sense. If the bottleneck is that you spend six hours a week copying data from emails into a spreadsheet, sorting documents, or writing the same kind of reply over and over, then yes: you have a repetitive, high-volume task with organised data. That is exactly where AI earns its keep.

You keep up, but you get paid late or lose customers. Usually pure process, not technology. No tool fixes the fact that nobody chases the invoices.

I didn't invent that order: it matches what the MIT report itself found. More than half of the generative AI budget they analysed had gone to sales and marketing, while the best return was in the back office, in the repetitive admin work nobody brags about. AI got bought where it looks good in the demo, not where it pays.

Why the right answer is often cheaper.

At this point the industry reflex is to show you a comparison table of returns. You've seen it: email returns 36 euros for every euro spent, local SEO costs a fraction per enquiry, therefore they beat AI.

I'm not going to do that, because it would contradict everything above. That 3,600% email figure comes from industry surveys, self-reported by companies that sell email tools. It is exactly the same species of number as the 95%: directionally useful, useless as a promise. If I use it to convince you, I'm doing the thing I criticised two sections ago.

The real argument isn't comparative return. It's the cost of being wrong, and there the difference isn't a matter of opinion:

Get an email campaign to your customers wrong and you know within two weeks, having spent a few tens of euros and an afternoon. Get your Google profile wrong and you know within a month, and the cost was time. Get an AI project wrong and you find out six months later, several thousand euros down, having burned your team's patience on a tool they don't use, and you walk away with the belief that "digital doesn't work in my sector" thrown in for free. That last bit of damage is the one I've seen most and the hardest to undo.

Asymmetric risk. That's why the order matters more than the choice: not because AI is bad, but because it's the expensive thing to get wrong and almost never the first thing outstanding.

The problem isn't AI. It's who advises you about AI.

Now the uncomfortable part, and the reason I wrote this.

Imagine you go to a doctor and find out they earn a commission on every operation they recommend. You wouldn't call them a criminal. You'd say, quite reasonably, that you need a second opinion from someone who gets paid the same whether you have surgery or not. Nobody argues with this in medicine. In business technology we have normalised it completely.

Almost the entire ecosystem talking to you about AI makes money if you implement: whoever sells the licence, whoever integrates it, whoever takes a cut from the vendor, whoever offers the free workshop that ends in a quote. Again, they're not bad people. It's that structurally they cannot tell you "don't do it", any more than you can ask a carpenter to assess neutrally whether you need furniture.

What you can do is ask someone who meets four conditions first, and this applies to me as much as to anyone:

One: they get paid the same if the answer is no. This is the only condition that genuinely matters. If the advice is free, the advice is advertising.

Two: they've built these things, not just read about them. I build agents and automations for clients, which is why I know where they break: the data that was worse than everyone claimed, the process nobody had ever written down, the employee still using their own spreadsheet. Someone who has never implemented sells you the demo.

Three: they know what the alternative costs. Someone who only knows AI will always conclude that you need AI. Being able to compare against advertising, SEO, email or simply fixing a process by hand is what turns advice into a decision.

Four: they leave you able to decide without them. If you depend on that person more at the end than at the start, it wasn't advice.

I've tried to put this in writing where it binds me. On my applied AI page there are two sentences that cost me money: that if you already have a tool working I won't replace it, and that if the project hasn't proven a return by the second phase, we stop. There's also a page explaining who I don't work with. Those sentences are there so you can use them against me, which is the only thing that makes them worth anything.

When the answer is AI.

Stopping at "no" would be just as dishonest. The same studies that document the failures also document where it works, and the patterns are clear.

It works when the task is repetitive, high-volume and has a verifiable output: pulling data out of invoices, sorting incoming email, preparing drafts a human then reviews. It works better in the back office than in sales. And it works more often when you buy something ready-made than when you build it: in the MIT sample, buying from a specialised vendor had around a 67% success rate, against roughly a third when the company built it in-house. If your adviser proposes building from scratch before trying anything off the shelf, ask them to explain why.

And it doesn't work without data. Gartner estimates that through 2026, 60% of AI projects without AI-ready data will be abandoned. In a small business, "ready data" doesn't mean a data lake: it means the information exists in one place, in one format, with some judgement applied. If your prices live in three spreadsheets that disagree with each other, that's your project this quarter, and it doesn't involve AI.

Incidentally, there's less of a rush than you're being told. According to the Spanish national statistics institute's survey on ICT use in business, in the first quarter of 2025 artificial intelligence was used by 21.1% of Spanish companies with ten or more employees, and 13.4% of those with fewer than ten. It's growing fast, yes. But if you run a small company and haven't done anything with AI yet, you're not last: you're the majority. Decide calmly.

Five questions before you sign anything.

Take these into the next meeting where somebody pitches you an AI project. They work on me too.

1. What specific problem does this solve, and how will we know in three months whether it did? If the answer contains the words "efficiency" or "transformation" and no numbers, there is no project yet.

2. What data does it need, and do I have it today? Not "we could get it". Today, and in what state.

3. What happens if it isn't working in three months? You want to hear an explicit stopping rule, not a plan for more phases.

4. What does it cost me to leave in a year? If my data stays inside their tool, I haven't commissioned a project: I've rented a dependency. I wrote about what happens when an AI tool disappears overnight, which is not hypothetical.

5. What do you gain if I say yes? The most useful of the five. It isn't a trap: it's about knowing where someone is speaking from. A calm, specific answer is a good sign. Discomfort is also an answer.

What I'm actually asking of you.

Not that you distrust AI. I use it daily, I implement it for clients, and I think in three years it will be about as debatable as having a website. What I'm asking is that you separate the decision from the sale, even if that's a fifteen-minute conversation with somebody who isn't paid for it.

Because the expensive mistake is almost never picking the wrong tool. It's never having had the conversation where someone with nothing to gain tells you your problem was a different one.

If you want the map of which use cases are worth it in a small company, go on to AI for small business: what actually works. If what's in front of you is a supplier decision, you may get more out of consultant or agency. And if you already have AI running and the legal side is what worries you, that's here.

Frequently asked questions.

Does my business actually need AI?

It depends on whether your bottleneck is one AI solves. AI is good at repetitive, high-volume tasks where the data is already organised: classifying documents, drafting replies, pulling information out of invoices or emails. It does not solve nobody finding you, your offer not being understood, or getting paid late. Identify the bottleneck before choosing a tool: if it isn't one AI solves, AI will be expensive and change nothing.

Is it true that 95% of AI projects fail?

That figure comes from The GenAI Divide, a July 2025 report by MIT's Project NANDA. It measured something much narrower: the share of integrated generative AI pilots that had produced no measurable impact on the profit and loss account. The sample was 52 organisations interviewed, 153 senior leaders surveyed and 300 public deployments, and it was a working paper with no peer review. The direction matches more rigorous studies, but 95% is not a universal fact about all AI.

How much does it cost to implement AI in a small business?

It varies so much that the figure alone tells you nothing: from subscriptions of tens of euros a month to five-figure custom builds. The number that actually matters before deciding is not the cost, it is the cost of being wrong: what you lose if the project fails, including the months of attention you will have spent on it and everything you did not do meanwhile.

What should I ask someone who wants to sell me an AI project?

Five questions: what specific problem it solves and how that is measured, what data it needs and whether I have that data today, what happens if it isn't working in three months, what it costs me to leave in a year, and what that person gains if I say yes. If the last question makes them uncomfortable, you already have your answer.

Are there cheaper marketing options than AI?

There is almost always something cheaper still outstanding: the half-empty Google Business Profile, the site that takes too long to load, the email to past customers nobody sends, the tracking that was never set up. They are not cheaper only in price but in risk: if an email campaign fails you know in two weeks and have lost very little. If an AI project fails you find out six months later.