95% of Companies Get Zero Return From AI. The Other 5% Ask Three Questions First

AI hype is at its absolute maximum. Huge amounts of fear and equal amounts of excitement. The people building the models are telling us that all jobs are going to disappear and every single business is going to be disrupted.

And every week, someone sits across from you with a hundred ideas of things they want to build.

Most of them don’t pass the sniff test.

You feel the pull too. A new model drops, a demo goes viral, and for an afternoon you want to build everything at once. Ten agents. A whole department, automated by Friday.

But there’s a huge contradiction in the numbers.

In 2025 a research team at MIT, Project NANDA, reviewed more than 300 public AI initiatives, interviewed 52 organisations and surveyed 153 senior leaders. Their report, The GenAI Divide, found that despite $30 to 40 billion of enterprise investment in generative AI, 95% of organisations were getting zero return. Just 5% of integrated AI pilots were extracting millions in value.

Huge piles of cash are being invested and destroyed through AI initiatives that are going wrong. That is the foundation for the next big bubble.

It doesn’t mean the tech isn’t real. Amazon and Google came out the other side of the dot-com crash, and we’re going to see companies build new products on AI that fundamentally change the way the world works.

So if 95% are getting nothing back, what do the 5% know that the rest don’t?

It isn’t the technology. Every business has access to GPT-5, Claude, Gemini, DeepSeek and hundreds of open-source models flooding the market. Switching from one model to the next latest and greatest is a single line of code change. The technology is not your moat. You can’t brute force an AI project into being successful.

The MIT team said as much. The divide isn’t driven by model quality. It “seems to be determined by approach.”

It’s about finding the right problems to apply this technology to.

Three questions before you build an AI agent

I ask three questions about any agent we want to build.

Does it solve a burning hair problem? A problem that is pressing, urgent and unignorable. Painful, like having your head on fire. For one of your customers, or for someone inside your business.

Does it unlock a core constraint? There is something holding your business back right now. If this agent directly addresses that constraint and lets you reach the next level of growth, build it. If it doesn’t, throw it onto the pile of good ideas and let it sit there. You’ll probably forget about it in a couple of months’ time.

Does it create 10 times the value for your customers? The entrepreneur’s job is very simple. Create value. As much as you can. The more value you create, the more you get to capture back as profit that flows into your bank account. If an agent isn’t increasing the value your customers get from you, don’t build it. Replace a support team with an agent that makes the service worse, and you’ll end up walking it back, losing money the whole way.

Three questions. Most ideas fail at least one of them.

Good.

Find the creakiest plank

Everyone wants to transform the business all at once, automating everything in one go. If you’ve ever worked somewhere that went through a digital transformation, you know it sucks. Nothing works. Nothing gets finished.

There’s an old Greek puzzle called the Ship of Theseus. The Athenians kept the hero’s ship for centuries, and every time a plank rotted they took it out and fitted a new one. Eventually there wasn’t a single original plank left. Was it still the same ship?

I don’t much care about the philosophy. I care about the method.

Nobody rebuilt that ship from scratch. They found the rotten plank and replaced it. Then the next one. One plank at a time, until the whole thing was new.

That’s what real AI transformation for business looks like. Other people are throwing AI at the wall and seeing what sticks. That’s why they’re in the 95%.

You’re going to operate in a smarter way. You’re going to find the creakiest plank on your ship and fix it with one agent. Deploy it. Get the return. Then find the next one. Again and again and again. You’re not going to try and build 10 agents at once. And over the next few years you’ll look back and realise you transformed your business from the inside out. A completely different machine, and it never once stopped running.

Finding the right problems to solve with AI

You won’t find the creaky planks from the captain’s chair.

Once you’ve grown a business, you’ve long since stopped being the person who runs the processes and talks to the customers. When I was younger, my teachers told me you have two ears and one mouth for a reason, and you should use them in that ratio. That is the whole skill.

Talk to your managers. Which KPIs are getting difficult to hit? Which objectives feel impossible because of the constraints they’re facing?

Then go right down. Find the people running the processes, because they know exactly which barriers they hit. Ask them one very specific question. What are you doing in your role when you’re being most impactful?

Then the follow-up. What consistently gets in the way of you doing that work?

And a bonus I absolutely love. If you had a magic wand and could fix any one thing in this business to create more value for our customers, what would it be, and why?

I cannot tell you the amount of gold I’ve found talking to the people closest to the customer.

That’s your pile of candidate planks. Run every one of them through the three questions first.

Score what survives

You don’t have unlimited resources. Entrepreneurs are terrible at prioritisation, and if you try to do everything at once you’ll end up with a pile of half-finished projects. So score whatever survives the three questions.

I use RICE, a scoring model built by the product team at Intercom. It stands for Reach, Impact, Confidence and Effort.

• Reach. How many people does this problem affect, directly and indirectly? One team member but thousands of customers? Hundreds of team members but one customer?
• Impact. How painful is it? Rate it from 0.25 to three. A show stopper that grinds the process to a halt, stops you working with a customer or causes churn is a three. A minor annoyance you can work around is 0.25.
• Confidence. A percentage between zero and 100. How sure are you of your reach and impact numbers? If it might be 10 people or might be a thousand, score it lower.
• Effort. How big is the build? An agent that replaces an entire department is a very large project. Sending an email or updating a spreadsheet is a small one.

Highest reach, highest impact, highest confidence, lowest effort. Take the top 12 and that’s your transformation roadmap.

Then build them. One at a time.

You can’t read about press-ups and get fit. You can’t have an AI-first business without building any AI agents. Design it, develop it, deploy it. Test and refine it until it’s reliable enough to do something in your business that touches the customer. Then move to the next plank.

Most people are too busy reacting out of fear. They’re running around convinced they’ll be disrupted if they don’t do something with AI, and it causes their downfall, because they’re burning money on projects that just don’t work.

Slow down. Go and find the creakiest plank in your business this week. Just one.

What is it, and who on your team already knows?