The 4-day AI quick-start plan

The 4-day AI quick-start plan is ten minutes a day, four days, one task. You time a repetitive task, hand it to an AI tool once, fix what went wrong, and then decide whether to keep it. Forty minutes in total, at the end of which you have either one job permanently off your plate or a specific reason to drop the idea.

Both are results. Most people trying to start with AI have neither: just fourteen browser tabs, a course they abandoned a third of the way through, and the same Monday as last year.

What are the four days?

One decision per day, ten minutes each. Do not skip Day 1, and do not merge the days. The gaps are part of the test.

1. Day 1 — Pick one task and time it. Something you do at least three times a week. Do it exactly as you normally would, with a stopwatch running, and write the number down. That baseline disappears the moment you change anything.
2. Day 2 — Hand it over once. Give one real, live instance of the task to an AI tool. Not a test case — the actual email, the actual notes. Time this attempt too, including the tidying up afterwards.
3. Day 3 — Fix the brief, not the tool. The first attempt is usually mediocre, usually because the tool got half the context that was in your head. Add what was missing, run it again, save the instruction.
4. Day 4 — Keep or kill, then schedule it. Compare Day 3's time against Day 1's. If it is faster and good enough to send, attach it to something that already happens in your week. If not, kill it and pick a different task.

Why does 10 minutes a day beat a weekend course?

Because what stops people is not a shortage of knowledge. It is that nothing in the working week changes. A course teaches you what the tools do. It does not put one of them between you and a task you will perform on Tuesday morning. Ten minutes a day is small enough to survive a bad week, and it touches real work from the first session.

Economists studying AI adoption in Denmark asked workers across eleven AI-exposed occupations what they were actually getting out of chatbots.

The same paper notes that controlled trials of these tools, in these occupations, routinely produce gains above 15%. The distance between 15% and 2.8% is not a measure of how good the software is. It is a measure of how rarely it gets wired into anyone's week.

Two other findings point the same way. Reported benefits are 10–40% higher where employers actively encourage use, and encouragement works less well on more experienced workers, which the authors put down to the difficulty of changing established work habits. The bottleneck is behavioural.

US survey data shows the same trial-versus-daily-use gap: about 23% of employed adults had used generative AI for work in the previous week, but only 9% used it every working day.

Day 1: which task should you pick, and why time it?

Pick the most boring thing you do most often, and time it with a stopwatch rather than an estimate. Not the most important task, the most repetitive one. Good candidates share four traits:

• You do it at least three times a week. A monthly task takes too long to prove anything either way.
• It is mostly words. Email replies, meeting notes, quotes, listings, first drafts, summarising documents. Language is where these tools are strongest.
• You could hand it to a new starter with written instructions. If you cannot explain it, you cannot delegate it to a person or a model.
• Being wrong is cheap. Nothing legal, medical, financial or client-final. Start where a bad draft costs you thirty seconds and a shrug.

Day 2: how do you actually hand a task over?

Do it once, on live work, and do not try to make it perfect. The purpose of Day 2 is to produce a specific, visible failure you can fix on Day 3.

Brief it the way you would brief a capable assistant who has never met you: what the task is, who it is for, what a good one looks like, and one example of a good one you did earlier. Most disappointing first attempts skipped the example.

Almost every bad first result is a briefing problem, not a tool problem. The model is not reading your mind. It is reading your brief.

Day 3: what do you fix when the first attempt is mediocre?

Fix the brief before you change the tool. Switching tools on Day 3 is the most common way this plan dies. You spend the next fortnight evaluating software instead of finishing a task.

Then save the working instruction somewhere you will actually reopen it. A prompt you rewrite from scratch every time is not a system, it is a hobby.

Day 4: how do you decide keep or kill?

With the two numbers you now have. If the task is meaningfully faster and the output is good enough to send, keep it. If it saved a minute or you rewrote most of it, kill it today, not eventually.

The arithmetic is small on purpose. Say the task took 12 minutes, now takes 5, and you do it four times a week: that is 28 minutes a week, or more than 20 hours a year, from forty minutes of effort. No transformation programme required.

One honest caveat: saved time does not become free time by default. In the Danish survey, about 80% of the time saved went straight back into other work tasks and fewer than 10% became breaks. Decide what the 28 minutes are for before you free them up.

So end Day 4 by attaching the task to a cue: the Monday invoicing block, the moment you come off a client call, the first coffee. That trial found no meaningful difference between routine-based and time-based cues, so use whichever you will genuinely notice. [Steve — add the task you ran this on and the before/after times.]

What if the four days don't work?

Then you have spent forty minutes and learned something specific, which beats the tab you have had open since March. Kill that task, run the plan again on a different one next week. Two or three rounds is normal.

Some tasks genuinely will not improve, and it is worth knowing which:

• Judgement-heavy work. If the hard part is deciding rather than writing, the writing was never the bottleneck.
• Anything where being wrong is expensive. It can be done, but the checking often costs more than the drafting saved.
• Tasks you cannot describe. If you cannot write the steps down, the problem is the process, not the software — and writing them down is often the win by itself.

If choosing the task is the bit you keep stalling on, that is exactly what an AI tools assessment does in 45 minutes. More on how I work, or start with the free scorecard.