Logo

Why 95% of AI Projects Make Zero Money (And How to Be in the 5%)

MIT studied 150 companies and found that 95% of generative-AI pilots deliver no measurable impact on the bottom line. The failure is almost never the technology.

·6 min read·Nick Puruczky

MIT studied 150 companies and found that 95% of their generative-AI pilots deliver zero measurable impact on the bottom line. Not underperform. Zero. RAND puts the broader failure rate around 80% — roughly double the rate of a normal IT project.

So if you have experimented with ChatGPT, or bolted a chatbot onto your site and watched it quietly fizzle, you are not behind. You are the default outcome.

Here is the part almost nobody says out loud: it is almost never the technology. Every tool those failed companies used, you already have access to right now. They lost on architecture — how the pieces get put together — not on the model.

The three things that kill AI projects

The diagnosis is close to identical every time, and it comes in three shapes.

  1. 1They automate everything at once. The result is fifteen half-built systems nobody trusts, and three months later the whole team is quietly back to doing it by hand.
  2. 2They automate a messy process. Now it is fast *and* messy, which is worse than slow and messy. A useful test: if a brand-new hire could not run your process from a written document, AI cannot run it either.
  3. 3They start with the tool instead of the business. That is how a company ends up with 42 apps, none of them connected to each other.

BCG measured the split directly. Only 10% of getting AI right is the model. 70% is your process and your people. Most companies pour effectively all of their effort into the 10%.

What the other 5% looks like

A service company we worked with — roughly $3M in revenue, 14 people. Lead response went from over a day to under 60 seconds. The 12 hours a week their team burned on manual data entry went to zero. Close rate rose 35%.

Same team, same service, materially different business. They did not buy better tools than the ones you already have. They assembled them in the right order.

The order is the entire game

  1. 1Understand the business first. Audit where the hours actually go — which is almost never where you assume. High-volume, low-judgment work is the gold mine.
  2. 2Clean your data. Garbage in, garbage out is the single most common reason automations break in month two. You need one source of truth.
  3. 3Then pick your tools.
  4. 4Then build.

Most people start at step three. That is the 95%.

If you want the longer version of step one, we wrote about how to choose what to automate first — including why the problem people complain about is usually the wrong one to fix.

Where to start this week

The fix for step one costs nothing but a notepad and two days of paying attention. For one week, write down every task you or your team repeat more than five times. Do not automate any of it yet.

At the end of the week, the expensive problem will be obvious. It usually is not the thing anyone has been complaining about.

Key takeaways

  • MIT found 95% of generative-AI pilots produce no measurable bottom-line impact; RAND puts broader AI project failure near 80%.
  • The cause is architecture and process, not model quality. BCG attributes only 10% of AI success to the model and 70% to process and people.
  • The three most common failure modes are automating everything at once, automating a messy process, and starting from a tool instead of the business.
  • The working sequence is: audit where hours actually go, clean your data to one source of truth, then choose tools, then build.
  • If a new hire could not run the process from a written document, AI cannot run it either.

Adapted for the web from the original issue of The AI Core, first published June 18, 2026.

More on this topic

AI Implementation Consulting

Keep reading

Want this applied to your business?

Book a call and we'll map your workflows and rank them by what each one costs you in senior time.

Book a call →