The pile of half-finished projects is arithmetic, not a character flaw. Starting something used to be expensive enough to make you serious about it; now it costs an afternoon and a prompt, so you started nine things instead of one. What didn't get cheaper is the last ten per cent. Nine things at seventy per cent return nothing. One thing at a hundred per cent returns something. Same hours spent, completely different answer — and the feeling of being behind comes from having done the work without collecting the result.

Why does cheap starting make you feel worse, not better?

Because the cost of starting used to do a job for you, quietly, and now it doesn't.

When a workflow needed a developer and a fortnight, that price tag acted as a filter. You thought hard before committing, because committing was visible and expensive. And once committed, you finished — abandoning it halfway would have been an obvious loss that someone would have asked about.

Cheap starting removes the filter. Nothing replaces it. So the discipline that used to be enforced by cost now has to be supplied by you, deliberately, against an instinct that is being rewarded every time a new tool makes something feel possible in an afternoon.

That's the trap in one line: the thing that got cheaper was the enjoyable part.

The tedious residue — edge cases, error handling, deciding where the output actually lands, telling someone the thing exists — costs exactly what it always did. It's just that it now sits at the end of an afternoon rather than the end of a fortnight, which somehow makes it easier to skip.

Is there evidence, or is this just a vibe?

There's evidence, and it's worth being careful about what it covers.

Gartner found that over half of generative-AI projects were abandoned after proof of concept by the end of 2025 — poor data quality, insufficient risk controls, rising costs, unclear business value. Its own earlier prediction had been thirty per cent, so reality overshot the pessimistic forecast. Gartner separately expects over forty per cent of agentic-AI projects to be cancelled by the end of 2027. MIT's Project NANDA found 95% of enterprise generative-AI pilots delivered no measurable return, and that internally built systems succeeded around a third as often as bought ones — because organisations underestimated the engineering needed to carry a pilot the last mile.

Every one of those studies measures enterprises. Companies with procurement, steering committees and pilot budgets. A one-person business is not a miniature enterprise, and none of that data was collected about you. Anyone quoting it as though it describes your situation is being sloppy, and this newsletter isn't going to do that.

What survives the translation is the shape: the model was never the hard part, and the last mile is where things die. That holds at every size, because it's a fact about attention rather than organisational design. Big companies just fail at it more expensively and write it up afterwards.

What does the maths actually say to do?

One thing at a time, to completion, and treat "done" as a definition rather than a feeling.

The reason the nine-at-seventy-per-cent state is so seductive is that it looks like more. Nine projects underway reads as ambition and momentum. But value doesn't accrue in proportion to effort spent — it accrues when something crosses into working, and not one minute before. A workflow that almost runs saves nothing at all. There is no partial credit.

So the reframe worth carrying: started things decay, finished things compound. An unfinished automation gets less likely to be finished every week, as you forget how it works and the tool it was built on changes underneath it. A finished one keeps paying, quietly, without further attention.

Only one of those is an asset.

None of this means you should stop being curious, and the point isn't to build less. It's to notice that you already did the hard part on several things, and that collecting on any one of them is a smaller job than starting the tenth.

Related Decodes

• Finish It: What to Do With the Half-Built AI Projects You Started This Summer — this week's hub and the two-hour clear-out.
• The Last-Mile Hour: Why the Final 10% Costs More Than the First 90 — Saturday: what that last ten per cent is actually made of.
• Eleven Models in Twenty Days: Why the Release Pace Is Now Your Problem — Monday: the supply of fresh day ones.
• AI Made You Faster. Did Your Prices Notice? — the other thing that changed when the work got cheap.

For the ten tools worth finishing with — free, vetted, 5-min read: 10 AI Tools That Save 10 Hours a Week.

AI the Brief (@aithebrief) — AI advice you can act on and trust, for time-starved founders.

Gartner and MIT figures describe enterprises, not one-person businesses; they are cited as evidence of a pattern, not as statistics about you.

TL;DR: The pile of half-finished AI projects is arithmetic, not a discipline failure — starting used to cost a developer and a fortnight, which acted as a filter that made you serious and made you finish, and cheap starting removed that filter without replacing it. The part that got cheaper was the enjoyable part; the tedious residue of edge cases, error handling, output destinations and handover costs exactly what it always did. Gartner found over half of generative-AI projects abandoned after proof of concept by end-2025 against its own 30% forecast, expects over 40% of agentic projects cancelled by end-2027, and MIT's Project NANDA found 95% of enterprise pilots produced no measurable return — all enterprise data, not solo data, but the shape travels: the last mile is where things die. Value doesn't accrue with effort spent, only when something crosses into working, so started things decay while finished things compound.