Why some AI projects stick when better-funded ones fail

The real success factor isn't the model or the budget

Ask why an AI project failed and you'll hear about the technology: the model wasn't good enough, the data was messy, the integration wasn't there. Ask why one succeeded and the honest answer is almost never technical. It's organisational — someone felt the problem, controlled the workflow, and agreed up front on what winning meant.

The clearest proof I've seen this year is deliberately small. MIT Technology Review profiled an independent tutor, Sam Finnegan-Dehn, who was drowning in admin — lesson planning, chasing readings, invoicing, keeping up with research — and watched it cap how many students he could take on. He tried a few tools and settled on Notion AI as a second memory laid over his scattered notes. One person, a twenty-dollar subscription, no transformation office and no early access to anything. If AI success came from budget or model quality, this story couldn't exist. It does — which tells you where the levers actually are.

Three questions that predict whether an AI project survives

Before you spend anything, run the project through three questions. A one-person shop passes all three without trying; most enterprise pilots fail them because the answers live in different departments.

• Whose number moves? For the tutor, his own — every hour reclaimed becomes room for another student. Nobody is off hoping "the org gets more efficient" in the abstract.
• Who owns it — who changes how they work on Monday? The same person. There's no lab building something and lobbing it over a fence to a team that never asked for it.
• Who decides whether it's any good? Him again, within a week, with no steering committee and no bruised egos if he walks away.

When a pilot dies, those three jobs — who benefits, who owns it, who judges it — are scattered across three teams, and the project falls into the gaps between them. That's the failure mode, and it has nothing to do with the software. You can't collapse a four-hundred-person company into one person with a Notion account, but you can hand a pilot to a single owner who feels the problem, controls the workflow, and gets to call it. Most organisations pointedly don't. That's a choice, not a law of physics.

What to hand AI, and what to keep

Look at the split. He delegated the filing: syncing information across documents, hunting through old notes, stitching scattered records back into a thread. He kept the teaching, and every judgement about what a stuck student needs next. He gave away the admin and held on to the judgement — and that line does more work than any feature in the tool.

Three decisions you can copy

Strip the win to its frame and it rests on three choices. No clever prompting, no code — just decisions, which is the part that transfers to an operation a thousand times the size.

Point AI where being wrong is cheap

If it files a note under the wrong student, he catches it in seconds and nothing breaks. The expensive work — the teaching, where a mistake costs a kid's understanding — stays human. This is the real answer to the endless "is AI good enough yet." Good enough isn't a fixed property of the model; it depends on what a mistake costs and whether you'll notice in time. Put the machine where a slip is survivable, nowhere near where it isn't.

Let your data pick the tool

He'd tried Claude and ChatGPT and landed on Notion AI — not because it won a benchmark, but because his notes already lived in Notion. The tool that can already see your stuff beats the cleverer one that can't, nearly every time. Plenty of buyers do the reverse: fall for whatever demos best, then meet the integration bill after they've signed.

Decide where the freed time goes before you free it

Reclaimed hours are worth nothing if they quietly refill with other admin. His didn't, because the destination was set in advance: grow the roster. Time came out of the filing and went straight into taking on students. That wasn't a happy accident at the end — it's the reason there's a story at all.

Measure the job, not the task

Hours saved is a task number — the "sixty to eighty percent faster" that looks great on a vendor slide and tells you almost nothing about whether the business is better off. The number that mattered here was whether he could take on more students: the job number. So before you start, write down not "how much time will this save?" but "what does the freed-up time let us actually do, and how will we know it landed?" A big improvement on a task nobody was stuck on is the most tempting dead end in this business.

If you're not a one-person shop

Notice that not one of those decisions is really about AI. They're about choosing the right problem, fitting the thing into how you already work, and following through on where the gain goes — the same unglamorous disciplines that sorted winners from losers long before any of us were talking to chatbots. A one-person shop gets that alignment for free. A big company can win the same way; it just has to do on purpose what he got by accident: one owner who feels the problem, can change the workflow, and gets to define the win. Same three moves. The technology was never the hard part.

It's the same pattern behind the four failure modes I see in stalled AI programmes, and it's exactly what an AI readiness assessment exists to surface before you spend — where the return is real, and where a mistake is cheap enough to survive.

This is the site version of a teardown I first published on my Substack — read the original essay there. Case details from "How small businesses can leverage AI," MIT Technology Review, June 2026; the argument is my own.