95% of Your Company's AI Projects Are Losing Money

A while back, I saw a number that startled me.

95%.

It comes from a 2025 MIT report. What it says: of all the generative AI pilot projects run by enterprises worldwide, 95% have failed.

95%. Let that sink in.

Is your company doing AI too? Is every company around you doing AI? Everyone's buying compute, hooking up APIs, running POCs, deploying agents. Billions spent — and what comes back?

What comes back is a 95% failure rate.

Today I want to talk about something deeply practical: why companies can't make money from AI, and how they actually can.

The Money Went to the Right Place — But It's Not Coming Back

Let's get one thing straight first.

It's not that companies don't want to invest in AI. Quite the opposite — they've poured in too much. Since generative AI caught fire in late 2022, it's been three years of everyone racing to get ahead. Automating workflows, building decision-support tools, boosting efficiency, cutting costs, accelerating product development — and now AI agents are the hottest thing in the room.

But when it's time to settle accounts at year-end, the CFO's face goes pale.

In Q4 2025, IBM hosted a closed-door executive gathering called Think Circle, bringing together a group of enterprise leaders to sit down and talk. The conclusion they reached was striking —

What's blocking AI returns isn't technology — it's the organization.

Corporate culture, governance structures, process design, data strategy — these "soft" things are the real bottleneck. Leaders found that the first wall AI ambition runs into isn't the model not being smart enough. It's how the company actually runs the whole thing internally.

Think about it — isn't that exactly how it is?

You buy a top-tier AI system. The model is powerful, the APIs are fast. But once you roll it out across the company, you realize nobody's changing their workflows to use it, data is scattered across a dozen systems and can't be fed in, and the business teams think "this has nothing to do with me." No matter how good the tech is, it can't move.

Three Brutal Numbers

Let me give you three more numbers.

First: Only 29% of enterprise executives dare to claim they can reliably measure AI's ROI. That's from the Think Circle report. In other words, out of ten leaders, seven can't tell you whether the AI they invested in is making money or not.

But — in that same report, 79% of companies say they've seen productivity gains.

Put those two numbers side by side. Notice something?

The efficiency is there, but efficiency doesn't convert into money. Productivity is something you feel; money is something you have to calculate. Between those two things lies a massive gap.

The second set of numbers comes from the IBM CEO Study: Only 25% of AI projects have achieved their expected returns. And the ones that have truly scaled at the company level? Just 16%.

What's the hardest thing for a CEO right now? It's being hounded by the board — "We've invested so much in AI, when do we break even?" — while simultaneously having to hold the team back and say "Don't rush, we're playing the long game."

Third: IBM's research also found that paying down the technical debt from legacy systems can boost AI returns by 29%. Because those old systems are dragging everything down — data can't flow, APIs don't connect, and everything needs constant rework.

String those three numbers together and it comes down to one sentence: AI isn't the problem — it's the soil you're planting it in.

What Even Is ROI? Let's Break Down the Math

Okay, so how do you actually calculate AI's return?

You have to approach it from two angles: "hard" and "soft."

What are hard returns?

They're the kind that can be directly calculated as money. For example, you use AI to automate IT operations — fewer outages, faster recovery, higher customer satisfaction, better renewal rates. That line connects straight to the income statement. Or AI boosts your marketing conversion rates, shortens your development cycles, opens up new revenue streams.

What are soft returns?

They're the kind that aren't easy to translate directly into money, but still benefit the company. Like employees getting comfortable with AI — job satisfaction goes up, turnover goes down. Or your decision-makers using AI-assisted analysis to make faster, more accurate calls. Or customers feeling that your personalized recommendations are more thoughtful, so the overall experience improves.

Should you pay attention to soft returns?

Of course. But here's the thing — soft returns don't convince the CFO. You say "employee happiness is up," and the CFO says "great, how much did we save on headcount?" You stammer and can't answer, and that budget line is in jeopardy.

So what you truly need is to find the metrics that can be calculated as hard numbers: how many labor hours saved, how much conversion rates improved, how much revenue grew, how much resource consumption dropped. Nail down those figures, and your AI project has a leg to stand on.

What Are the People Actually Making Money Doing Right?

So here's the question: has anyone actually made money from AI?

Yes.

The IBM Institute for Business Value has done extensive case studies. They found that in product R&D, some teams achieved a median ROI of 55%.

55%. In a landscape where 95% are failing, how did these people pull it off?

They did four things.

First, encourage feedback — don't work in silence. AI transformation isn't a one-shot deal. You have to let everyone involved speak honestly — where is this process getting stuck, how much time is being wasted at that step. The smoother the feedback flows, the less rework.

Second, take small steps and iterate. Don't try to blanket your entire product line with AI in one go — that just burns people out and magnifies risk. Test it in one workflow first, and once it works, expand.

Third, find opportunities in user data. Don't try to educate users or change their behavior. Go where the users already are. The data tells you which part of the journey users need AI help with the most — so that's where you put AI first.

Fourth, build cross-functional teams. Don't let AI become the tech department's solo project. Pull in product, design, data, and business. The moment departmental walls go up, the project slows down.

These four principles sound plain, right? But the teams that actually live by them are pulling in a 55% median ROI.

There's another domain worth mentioning: content supply chains.

IBM found that companies taking a holistic view of AI and content saw content supply chain ROI jump by 22% and generative AI integration ROI jump by 30%.

What did they do right?

Three things. First, prioritize from a bird's-eye view — don't just look at one department's needs, look at the full chain and deploy AI where the leverage is greatest. Second, take change management seriously — AI is inherently controversial, and if you don't invest in employee communication and cultivate change champions, the whole thing stalls. Third, hand low-risk work to AI and free people up for creative work — when teams don't have to worry about AI making mistakes, creativity actually gets unleashed.

But Wait — Don't Rush to Settle Accounts

At this point, you might be thinking: alright, I'll go back and follow these steps, and I want to see returns next quarter.

Hold on.

Someone disagrees with that idea. His name is Jensen Huang.

Yes, that Jensen Huang. The one whose company is now worth $4 trillion.

At the Cisco AI Summit in February 2026, Jensen Huang said something I found really interesting. He said —

Don't force engineers to justify AI projects with hard returns right from the start. It's like your kid coming to you saying they want to try a new hobby, and you asking them "what's the ROI on that hobby?"

Would you talk to your own kid like that?

No. You'd say "go try it."

He also said: "Let a thousand flowers bloom."

What does that mean? Let teams experiment. Don't rush to use ROI as a gatekeeper — to filter, to screen, to kill projects at the very beginning. The more you gatekeep, the more likely you are to kill the thing that could've been truly transformative.

Sounds a bit contradictory to what I said earlier about "calculating hard numbers"?

It's not. It depends on the stage.

Early on is the exploration phase — the priority is experimentation. If you force teams to deliver ROI in this phase, you're killing possibilities. But exploration isn't indefinite. At some point, you do need to tighten up, pick the most valuable directions, concentrate resources on them, and seriously calculate returns.

Explore broadly. Converge ruthlessly.

That's my take on it — I could be wrong.

Finally

Where is your company's AI project right now?

Is it soaking in that 95% pool, or is it on the 55% line?

If it's the former, don't blame the technology first. Go back and look at your organization, your processes, your data, your team. If the soil isn't right, the best seed in the world won't grow.

If it's the latter, congratulations. But don't get too smug — because Jensen Huang said it himself: sometimes, don't rush to settle accounts.

Maybe the best AI project is the one whose returns you can't quite calculate yet.

And I hope — that in the flood of the 95%, you find your own 5%.