Zaibex
Engineering··Updated ·8 min read

Why most AI projects fail, and how to de-risk yours before you spend

Most AI projects fail for reasons decided before any code is written. What the research found, and seven checks to run before you spend a dollar.

Shahzaib Shoaib
Founder and lead engineer
A clipboard headed Check First with seven ticked boxes, laid across a closed leather chequebook on a wooden desk

The AI is almost never the part that breaks. Most AI projects do fail, and the research on that is blunt. But they fail for plain business reasons, and most of those get decided in the first two weeks.

Four mistakes do most of the damage. Teams pick the wrong problem, or start with data that isn’t ready. They mistake a demo for a product, or bolt the tool onto an old way of working. Each one is cheap to fix before you spend.

There is one more finding, about who should build it, that runs against most people’s instinct. We’ll get there.

First, meet the reader I have in mind. Say you run a sales training company with 12 coaches and about 800 people in your programs each year. Call her Priya. She has set money aside for her first AI build, and she wants to know what could sink it.

Priya is a made-up example. Her choices are the real ones.

The scary numbers measure three different things

People throw the failure numbers around carelessly, so here they are with their labels on. MIT’s 2025 study of enterprise AI found that 95% of organizations were getting no measurable return on their generative AI spend. S&P Global found the share of companies abandoning most of their AI initiatives jumped to 42% in a single year, up from 17%.

Then there is Gartner. Its early 2026 survey of infrastructure leaders found that only 28% of AI use cases fully met their ROI (return on investment) expectations.

95%
of organizations saw no measurable return on generative AI spend (MIT, 2025)
42%
of companies abandoned most AI initiatives, up from 17% a year earlier (S&P Global)
28%
of AI use cases fully met their ROI expectations (Gartner, 2026)

Those are three different measures. Getting no return, giving up on a project and missing an ROI target are separate failures. So I won’t blend them into one terrifying number.

They do come from independent research, and they all point the same way. If Priya spends carelessly, the odds are against her.

The model is rarely what breaks

This is the good news hiding in those numbers. The RAND Corporation interviewed 65 experienced AI engineers about why projects collapse. Four of the five root causes they found were organizational. They were about people and decisions, and the technology came second.

MIT landed in the same place. The barrier wasn’t infrastructure, talent or model quality. Companies were buying tools without fitting them to how their business actually works.

That matters because organizational problems are the ones you can fix before you spend a dollar. The model almost always works. The open question is whether the project around it was built to succeed.

A vague goal sinks more projects than bad code

RAND found the most common cause of failure is a misunderstood problem. The business and the engineers never agreed, in plain words, on what they were solving. Close behind it is starting with the technology, chasing the newest model instead of a defined, costly problem.

Here is how that looks for Priya. “We should do something with AI agents” is a wish. “Each of my coaches spends about five hours a week answering the same practice questions by email” is a project. She can put a cost on it, and she can tell later whether it moved.

Start with the second kind of sentence. Or don’t start.

Messy data stalls projects halfway through

An AI system is only as good as the data it works from, and most companies’ data isn’t ready. Gartner found that roughly 63% of organizations either lack the data practices AI needs or don’t know whether they have them. It also predicts a large share of AI projects will be abandoned because the data under them was never AI-ready.

AI-ready doesn’t mean big. It means the data is easy to reach, defined the same way across the business, mostly complete, and looked after by someone.

Priya’s method shows the problem. It lives in slide decks, two versions of a workbook and years of call recordings, and each coach teaches it a little differently. Before an AI coach can teach it, someone has to decide which version is the method. That is project one, and it costs far less before the build than halfway through.

A great demo is the riskiest moment

This mistake fools people because it feels like success. A demo answers the questions it was built to answer, in front of people who want it to work. Production gets the question nobody planned for, phrased badly and missing context. It arrives at 2am, when some other system is down.

MIT put numbers on the gap. In 2025, around 60% of the custom AI tools companies considered got a real evaluation. Only 20% reached a pilot. Just 5% reached production.

A funnel: 60% of custom AI tools were evaluated, 20% reached a pilot, only 5% reached production.
Most custom AI tools stall after the pilot, which is where the real engineering starts. Source: MIT, State of AI in Business 2025.

The drop from pilot to live system isn’t a final coat of polish. It’s most of the engineering. Tests catch a slip in answer quality before users do. Every outside service gets error handling, and monitoring watches quality and cost as well as uptime.

The AI also has to connect to the tools people already use, and it should roll out in stages. For Priya, the demo will handle her favorite sales objection beautifully. The real test is a rep at 11pm asking about a deal the AI has never seen. A prototype that impressed the room is the cover of the product.

AI on an old process gives you a faster old process

McKinsey tested 25 organizational factors against whether companies actually saw profit from generative AI. The biggest single factor was whether they had redesigned their workflows around the AI. Only about 21% had. Nearly four in five were layering AI on top of processes they never rethought.

This one looks like success right up until it doesn’t. The AI works. The handoffs, bottlenecks and manual steps on either side stay exactly as they were. One link of the chain got faster, and the chain was never the problem.

Picture Priya’s version. Her people practice with an AI coach between sessions, yet her coaches never see how that practice went, so sessions run exactly as before. She paid for a new tool and kept the old week. The better question is what a coaching week should look like once the practice partner exists.

The model is the easy part. Projects are won or lost on the problem, the data and the workflow.

Outside specialists doubled the odds in one study

Here is the finding I promised. MIT found that AI projects built with a specialized outside partner reached production about twice as often as projects built entirely in-house. That was about 67% against 33%.

MIT adds its own caveat. It is a correlation from a modest sample, so it doesn’t prove cause. Companies that hire well may simply run projects well. I also sell this kind of work, so weigh my enthusiasm accordingly.

The reason is still plausible. A specialist has met these four mistakes before, on someone else’s budget, and designs around them from day one. A team building its first AI system learns them in production, the most expensive place to learn anything. Just check the specialist is real, because not every vendor selling “AI” is what they claim.

Seven checks belong before the first invoice

Almost every mistake above is a decision made early, while it is still free to change. These are the checks I would want Priya to run before she signs anything.

  • Write down one costed problem. Name one process and what it costs you, in words the business and the engineers both agree on.
  • Check the data first. If it is messy, scattered or missing, fixing it is project one, so budget for it.
  • Define success before any code. Pick the number that must move, by how much and by when. A fuzzy target guarantees a fuzzy verdict later.
  • Redesign the workflow along with the step. Ask what the process should look like once the AI exists, and build that.
  • Budget for production, since the demo is the cheap part. That means tests, monitoring, error handling, connections to your tools and a staged rollout.
  • Start small and prove it fast. A narrow pilot with a real path to production beats a broad program that tries everything at once.
  • Expect to iterate. The most common failure in Gartner’s 2026 survey was expecting too much, too fast.

The free audit exists for those first two weeks

Every check on that list happens before a line of code exists, while changing course costs nothing. That is why we start every engagement with a free call and a free AI audit. We look hard at the problem, the data and the workflow. Then we tell you honestly whether there is a project worth doing, and what would make it fail.

If there is one, you get a fixed quote in writing before any work starts. You own the code and the data. For a coaching business like Priya’s, here is what we build for coaching companies.

So what would sink Priya’s project? Almost certainly not the AI. It would be a fuzzy goal, or a method that lives in three places. Or a demo she trusts too much, or a coaching week that never changes.

The cheapest moment to de-risk an AI project is before it has a budget. If you are about to commission your first build, book the free call and bring the problem you would start with.

Sources

  • MIT NANDA, The GenAI Divide: State of AI in Business 2025
  • S&P Global Market Intelligence, 2025 survey on companies abandoning AI initiatives
  • Gartner, early 2026 survey of infrastructure and operations leaders on AI use case ROI
  • RAND Corporation, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed (2024)
  • Gartner, Lack of AI-Ready Data Puts AI Projects at Risk (February 2025)
  • McKinsey, The state of AI: How organizations are rewiring to capture value (March 2025)
Written by
Shahzaib Shoaib — Founder and lead engineer
Now booking new projects

Tell us how your business works. We reply within one business day.

Start with a free call. If a build makes sense, we run a free audit and send a one-page plan with a fixed quote in writing.

Free call · 30 min · with the founder

Based

  • ColoradoHQ
    Zaibex LLC
    518 Harrison Ave, Unit #2070
    Leadville, CO 80461