Most AI projects fail at the first question, and it isn't "which model?"

The meeting usually starts the same way. Someone has seen a demo, a board member has asked about the AI plan, and the leadership team wants to talk about which model to use. Should it be the one from the big vendor, the open source one, or the one built into the ERP they already pay for?

That's the wrong first question. It's a fine third or fourth question. But when a team opens with it, they've skipped the one that decides whether the project should exist at all.

The first question is this: is the process repetitive, high volume, and expensive enough in staff time to justify the cost and risk of automating it? Most candidates fail that test. We think that's the single biggest reason so many AI projects end up abandoned.

The failure numbers are real, and they point upstream

You've probably seen the headlines. They're worth reading closely, because the reasons behind them matter more than the percentages.

S&P Global's survey of more than 1,000 respondents in North America and Europe, reported in March 2025, found that the share of companies abandoning most of their AI initiatives jumped to 42 percent, up from 17 percent the year before. The average organization scrapped 46 percent of its AI proofs of concept before they reached production.

In June 2025, Gartner predicted that more than 40 percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, or inadequate risk controls. Notice the middle one. Unclear business value means somebody built the thing before anyone could say what it was worth. That isn't a model problem. It's a selection problem.

And in August 2025, researchers at MIT's NANDA initiative reported that, despite an estimated $35 to $40 billion in enterprise spending on generative AI, about 95 percent of organizations were getting zero return. The study drew on interviews, a survey, and a review of public deployments, so treat the exact figure with some care. The direction is hard to argue with.

Put those together and a pattern shows up. The projects aren't mostly dying because the model was wrong. They're dying because the work they were pointed at was never a good candidate.

Why the model question feels so important

It's concrete. It has vendors attached, and vendors bring demos. Choosing a model feels like progress, and it gives everyone something to compare in a spreadsheet.

The selection question is harder. It means somebody has to go count things. How many invoices come in each month? How long does each one take? Who touches it, and what happens when they get it wrong? That's operations work, not technology work, and it doesn't come with a slide deck. So it gets skipped.

Here's the uncomfortable part. A great model pointed at a low-volume, judgment-heavy process will still lose money. A mediocre model pointed at a high-volume, rules-heavy process can pay for itself. The work you pick sets the ceiling. The model only decides how close you get to it.

The screening test

We use a simple four-part screen before any conversation about tools. You can run it yourself in an afternoon with a whiteboard and someone who actually does the work.

1. Volume

How many times does this happen in a month? Ten is not a candidate. Ten thousand might be. If the honest answer is "it depends" or "a few a week," stop here. There isn't enough work to spread the cost of building, testing, and maintaining anything.

2. Repetition

Does the task look roughly the same each time? Same inputs, same steps, same kind of output? If every instance needs a fresh judgment call, or the rules change every quarter, you're not looking at automation. You're looking at a person who needs better information.

3. Cost of staff time

Multiply volume by minutes per task by a loaded hourly cost. Be honest about the minutes. Then compare that annual figure against what it would cost to build, license, integrate, and maintain a solution, plus the time your people will spend checking its work. If the savings don't clearly beat the cost, the project is a hobby.

4. Cost of error

What happens when it gets one wrong? A misfiled support ticket is an annoyance. A wrong number on a customer contract or a missed safety flag is a different conversation. High error cost doesn't kill a project outright, but it means more human review, and human review eats the savings you counted in step three.

A strong candidate is high on the first three and low, or at least manageable, on the fourth. Expect most ideas on a typical AI wish list to fail at least one of these, and many to fail two. That's not bad news. That's the screen doing its job.

What to do with the ones that fail

Failing the screen doesn't mean the idea was dumb. It usually means one of three things.

• The process needs fixing, not automating. If a task is slow because the steps are a mess, automating it just makes the mess faster. Clean it up first. Sometimes that's the whole win.
• The feature already exists. A lot of low-volume use cases are covered by AI features already built into software you own. Turn them on, train people, and move on. No project required.
• It should stay with a person. Some work is low volume and high judgment, and that's exactly where your experienced people earn their pay. Give them better tools and leave it alone.

The ideas that pass are the ones worth a real business case, a small pilot with a measurable target, and then the model conversation everyone wanted to have in the first place.

Where we come down

We'd rather a client walk out of the first meeting with two good candidates than with ten maybes. Ten maybes become ten pilots, and from the numbers above you can guess how most of those end. Two candidates that passed the screen have a real chance, and they give you something you can measure, defend to your board, and build on.

None of this is glamorous. It won't make a great demo. But it's the difference between an AI program that shows up in your margins and one that shows up as a write-off.

The cheapest AI project you'll ever run is the one you correctly decided not to start.

Talk with us

If you have a list of AI ideas and aren't sure which ones are worth the money, we can help you run the screen, size the real savings, and tell you plainly which ones to kill. We don't resell software and we don't bill build hours, so we have no reason to talk you into a project you don't need. Start a conversation with us.