Why do most AI projects fail?
Most AI projects fail because they were set up wrong, not because the AI didn't work. On Tech for Non-Techies, host Sophia Matveeva walks through five reasons the majority of enterprise AI initiatives stall: the wrong problem, messy data, chasing technology instead of need, missing infrastructure, and problems AI simply can't solve. She contrasts a failed $600K New York City chatbot with a Korean steel plant that got it right.
From the episode
321. Why 80% of AI projects fail, and it's not the AI's fault
Sep 16, 2026 · 17 min
Reason one: nobody defined the problem or the metric
Sophia's first failure mode is the most basic. Teams start with "we should do something with AI" and never pin down what they're trying to improve or how they'd know if it worked. There's no baseline, no target number, and no agreement on what a win looks like.
That sounds obvious, but she argues it's the most common reason projects quietly die. Without a clear problem and a way to measure it, the project can't be judged, so it drifts until the budget runs out.
Her framing throughout the episode is that AI is execution and product thinking is judgment. If the judgment at the start is poor, the AI just executes that poor judgment faster.
"People aren't identifying the problem correctly and aren't measuring what they're trying to improve."
Reasons two and four: the data and infrastructure aren't ready
The second and fourth reasons are related. Organizations want an AI transformation, but their data is scattered across systems, inconsistent, and would need months of cleaning before a model could use it. Leaders tend to underestimate this because the cleaning work is invisible and unglamorous.
Beyond the data itself, Sophia points out that many organizations lack the underlying data infrastructure to make AI work at all. You need pipelines, storage, and access controls in place before the interesting part starts.
For a non-technical leader, the practical question is: before we talk about models, can we actually get the relevant data into one place in a usable form? If the honest answer is no, that's the project, not the AI.
"Organizations want to go through AI transformation, but their data is all over the place and needs lots of time to clean."
Reasons three and five: chasing the tech, and picking impossible problems
Reason three is chasing the technology rather than the need. A board hears about AI, a budget appears, and someone goes looking for a place to spend it. Sophia's advice is to invert that: focus on the problem, not the technology, and ask whether this problem is genuinely a priority worth solving. Sometimes the right answer is a spreadsheet.
Reason five is the one people don't want to hear. Some problems are just too difficult for AI to solve. The tools are impressive, but there are things they cannot do reliably, and a project built on one of those will fail no matter how well it's run.
The $600K New York City chatbot she describes is the cautionary tale here, an expensive deployment that didn't deliver because the setup was flawed from the beginning.
"Sometimes the problem is just too difficult for AI to solve. AI is great, really good, but there are just some things it cannot solve."
What the Korean steel plant did differently
The success story in the episode is a Korean steel plant, and Sophia pulls three lessons from it. First, they got the technologists and the domain experts speaking the same language so they actually worked together on the problem, instead of one group building something and handing it to the other.
Second, the domain experts were treated with respect, as the people whose knowledge made the system work, not as staff about to be replaced by an algorithm. That changed how willing they were to contribute what they knew.
Third, they tested in real production for five months before rolling it out fully. Not a one-time demo with five people. Five months of real use, real edge cases, and real feedback. Sophia's point for non-technical leaders is that none of these three moves require coding. They require judgment, which is exactly the part that can't be outsourced to the model.
"They tested it on real production for five months before rolling it out fully. Not just a one-time demo with five people."
What to remember
- Most AI failures come from project setup, not the technology.
- Define the problem and the metric before anything else, or the project can't be judged.
- Messy data and missing data infrastructure sink projects before the model matters.
- Chase the need, not the technology, and accept that some problems aren't solvable with AI yet.
- The successful rollout paired technologists with respected domain experts and tested in production for five months.
People also ask
Do I need to be technical to lead an AI project?
No. Sophia's argument is that the decisive work is judgment: picking the right problem, setting the metric, and getting experts and engineers collaborating. AI handles execution.
How long should we test an AI system before rolling it out?
The steel plant example in the episode ran in real production for five months before a full rollout. The point is sustained real-world testing rather than a single demo.
What should I check before starting an AI project?
Whether the problem is clearly defined and measurable, whether the data is clean and accessible, whether the infrastructure exists, and whether the problem is one AI can actually solve.
Based on the Tech for Non-Techies episode "321. Why 80% of AI projects fail, and it's not the AI's fault," released September 16, 2026, hosted by Sophia Matveeva.
