← Blog

Why AI projects fail to pay off: five causes we see

Nearly nine in ten companies now use AI in at least one part of the business, according to McKinsey’s State of AI survey from August 2026. Only 37% can point to any effect on profit, about the same share as the year before. In Foundry’s 2026 State of the CIO survey, 19% of technology leaders said their AI initiatives had met or exceeded their business goals.

I first wrote about this in 2019, when the topic was data science. Most of it still holds. These are the five causes I see most often, and what I’d do about each one.

1. Nobody agreed on what to improve

Companies collect a lot of data and hire people to “make things better” with it. Without a clear goal, those teams can spend months building models that don’t touch anything that makes money.

Here’s an example from our own work. We built a recommendation system for a company that distributed thousands of apps to customers around the world. At first, the company wanted more downloads. As we dug into the business, we found it really wanted engagement: people opening the apps and using them. So we measured how often people opened an app combined with how long they used it.

Even that took care. Some people open their email once and leave it open for months. Someone in Chicago checks a weather app several times a day, and someone in San Diego barely needs one. It took six weeks of work with the product team to define the goal correctly. Time spent defining the goal is the best time you’ll spend on an AI project.

What to do: pick one metric before any work starts. Agree on it with the executive responsible for it, and measure where it stands today. That’s the first step in how we work with clients.

2. Nobody planned how the results would be used

A model is only worth something once people use it. If we build a recommendation model, it pays off when the marketing team uses it to send each customer offers they’ll actually want.

Too often I hear, “Some other team will handle that when we get there.” By then the model may need data that the other team can’t supply, or answers in a format their systems can’t take.

What to do: before work starts, name the team that will use the results, how they’ll use them, and what they need from the model. Without that team and that plan, the project shouldn’t start. I go through the questions to settle in putting AI into production.

3. The data pipeline came last

The pipeline is the set of steps that brings data to the model and carries its answers to the people and systems that use them. Without it, a model works in a demonstration and nowhere else.

Getting the data ready is also where most of the time goes. I wrote about that in why so much of an AI project goes to cleaning data.

What to do: build the simplest possible version from end to end first. For a churn model, the first version might only count how often a customer visits your support page. Once data flows from your systems to the model and back to the people who use it, make the model smarter.

4. AI sits too far from the business

AI works best when the people building it sit with the teams that use it. When it’s treated as a technology project and parked in a separate department, the work drifts away from what the business needs.

Being close to the business also shows you where to reuse work. Say a streaming service spends six months building a model that recommends shows inside its app. The marketing team wants to email each subscriber shows they’ll like. With a small change, the same model can produce those lists every week. The company gets a second result from work it has already paid for.

Leadership matters here too. One of my first questions to any new client is, “How do you use data to make decisions?” When the most important decisions still come down to a gut call at the top, AI has a hard time changing anything.

What to do: put the people building your AI in the weekly meetings of the teams they serve. Have them report to a senior leader who cares about the result, and look for models you can reuse from the start.

5. Nobody measures it after launch

A model can test well on historical data and still disappoint once it’s live. The only way to know is to compare it with what happened before, usually with an A/B test, where one group of customers sees the new model and a similar group doesn’t. I explain how in better testing equals more revenue.

Models also get worse over time as customers, seasons and stock change. That’s why models need ongoing maintenance.

McKinsey found the same pattern in its 2026 survey. The 6% of companies it calls high performers redesign the work itself when they bring AI in.

What to do: report every result against the starting value you measured in step one, and keep testing after launch. That’s what Bennett ONE does on every system we put into production: it records every question, answer, approval and result, and reports every result against the value it started from.

Talk to us

If one of these sounds familiar, the first call with one of our senior engineers is free and takes 30 minutes. We’ll tell you plainly whether we can help.

Let's talk about what you want to improve

Talk to an engineer