The scene repeats in every company this year. Someone shows an AI demo, the room lights up, ideas fly three a minute and somebody with a budget says the line, “we have to do something with this”. Six months later, the tool that came out of that meeting gathers digital dust. Nobody opens it and nobody misses it. Nobody quite remembers what it was supposed to fix.
We call that project the shiny toy. A shiny toy is a project born from what the technology can do, not from what hurts the business. It gets bought for the demonstration and abandoned once somebody does the math. And it does not fail because it is badly built. It fails because it is an orphan, because it never had a concrete pain to hold on to once the novelty wore off.
The scale of the problem is not anecdotal. Gartner put a number on it in June 2025, more than 40% of agentic AI projects, the kind that act on their own, will have been canceled before 2027 ends. Our read of that number is uncomfortable for the industry, because most of those projects did not die of bad engineering. They were born dead, chosen for the shine and not for the pain. How the ones that do launch survive is the story of keeping AI alive. This article is about the step before, choosing what deserves to launch.
The question that manufactures toys
The question “where can we put AI?” is backwards. It starts from the solution and goes hunting for problems, so it finds exactly what it looks for, places where AI fits. But fitting is not the same as paying off. The useful question starts at the other end. What is costing us money, time or customers right now? With that list on the table, AI has to compete on real pains, and it stops being the goal and becomes just another tool.
How to spot a toy before paying for it
First signal, nobody suffers it. Ask who loses hours or money today to the problem the project claims to solve. If the answer takes a while or is “everyone in general”, the pain has no owner. A project without someone who wants it gone has nobody to defend it in the second meeting.
Second signal, the gain has no number. A healthy project shows up with its math, the hours it returns, the errors it prevents, the wait times it kills. A toy shows up with adjectives. How we separate the metric that commits from the one that does not is in measure AI by profit.
Third signal, the AI is in the headline. Remove the acronym from the project’s name and see if anyone is still interested. If the answer is no, what was being bought was the headline. A tool gets bought for what it removes. A toy, for what it shows.
The antidote is a map of pains
What we do before proposing anything is map what hurts, with the people who suffer it in the room. And we separate two families that call for different urgencies. Critical problems already cost money every week. Bottlenecks do not hurt yet, but they will cap growth as soon as volume rises. You attack what burns first and you watch what smokes. Every candidate enters the list with its math done, never with its demo.
Two of our automation projects started exactly like that. At Barceloneta Premium, a real-estate agency in Barcelona, the pain was dozens of daily inquiries with five to ten minutes of checking behind each one, a fire with a number on it. At Stanton, a property management firm, the utility paperwork someone keyed in invoice by invoice. Neither started with a demo. Both started with somebody fed up and a figure on the table.
The cost that never shows on the invoice
A toy does not just burn its own budget. It burns the credibility of the next project, because the committee that buried one looks at everything that follows through a magnifying glass. The first dead toy makes every following project more expensive. That is why discarding early is not pessimism, it is protecting the budget and the committee’s patience for the case that does pay off. We do it as standard. If a case has no owned pain and no numbered gain, we say so before charging.
The test fits in the next demo you get shown. Let it finish, applaud if you must and ask one question. What pain of ours does this fix? If the room takes too long to answer, you have your diagnosis. The shine belongs to the vendor. The pain has to be yours.
If you already have the map of pains and a candidate with its math, this is how we work in AI workflow automation. And if you are still building the judgment, the AI agents guide walks through the cases that pay off and the ones that do not.