I've watched a fair number of companies roll out AI tools over the last few years, and I keep landing on the same conclusion: what decides whether an AI project works has almost nothing to do with the AI.
It's rarely the model. It's rarely the tool. What kills these projects is a pile of very ordinary, un-technical things. Human factors.
Almost every company is "doing AI" now. And when it comes to who owns it, most SME founders walk the same path: hand it to whoever is already doing the work, and have them improve their own patch with AI on the side. No new hire, no extra budget, and they know the work best. Looks like the cheapest, cleanest option.
But the cheapest path is usually the hardest one to actually finish. A lot of these projects are dead on arrival. They don't fail halfway; they were never set up to succeed. The catch is that you don't find out for six months. By then the money's spent, the person's worn out, and it all ends on one line: "we couldn't find the right AI tool."
So before you spend a cent, run these five checks.
Check 1: You can't say who runs this process today, or how you'd measure it.
The thing you want AI to improve: is anyone actually running it right now? Can you measure its output, its numbers? If you can't even answer that, the process hasn't taken shape yet. You're not speeding up a machine. You're parking a V8 engine on an empty lot: no car around it, and however much power it has, there are no wheels to turn. AI isn't magic. If there's nothing underneath it, it can't conjure anything out of thin air.
Check 2: You dressed up a fix that's yours to make as an "AI project" and handed it off.
Don't confuse this with Check 1. Check 1 is "there's no process underneath." This is "there's a process, it's running, but it's broken, and you already know it." When that's the case, what you actually need isn't AI, it's change management. But you don't want to run that reform yourself, so you wrap it in an "AI project" shell and pass it to a subordinate. All that does is make a broken process run faster. This time there's a car, but all four tires are flat. Dropping a V8 into it only sends it further on flat tires. The fix is new tires, not the gas pedal.
Check 3: You gave them a KPI, not the power to change things, get access, or make the call.
Landing AI is, at bottom, changing the status quo: reworking a process, getting into the data and the systems, making calls at a dozen forks in the road. None of those three is something a KPI conjures up. They need a dataset, so they go ask. They need into a system, so they go ask. They need another team to change one step, but they're not that team's boss, so they go ask again. Every move stalls on an approval, and by the time they've done the full lap, the enthusiasm's worn off and nothing has moved. You think saying "it's yours now" handed them power. It handed them a title. The authority that actually moves this stayed on your desk.
Check 4: What they're automating is their own colleagues' jobs, and you can see it but decide to deal with it "later."
What you're asking them to automate is usually work they and two or three colleagues have done for years. If it lands, what happens to those people? They know exactly what's at stake, so when you tell them to push it full speed, they're torn. Afraid of the blame (I led it, I take the fall if it breaks). Afraid of making enemies (automating someone's job away gets you resented). Afraid of doing it for nothing (even if I build it, will the team use it?). Every one of those is reasonable. And you can see this cost too. You're just betting it'll "probably work out," so you push it to launch day to face. But this doesn't sort itself out. The longer you defer, the less traction they have, and by the time you finally deal with it, the project's already cold.
Check 5: You never planned to take anything off their plate.
If you don't lift a chunk of their existing work and carve out time and space nobody else can grab, this stays the thing they get to "when they have a minute." The second the day-to-day flares up, a client pushes, or you drop a new project on them, the first thing set down is always this one, the high-risk one. It hangs there, never moves, and then everyone says AI just doesn't work here.
Look at the five
Not one of them asks which AI to use. Every one asks about you.
If you can't clear them, don't start yet.
Originally published on LinkedIn. Want to discuss this further? Get in touch.
