First, price the boring fix
Before any AI tool, ask what the dumb solution costs.
- Same five questions every day? A pinned FAQ post and the saved-replies feature in WhatsApp Business. Free, instant, never invents a price.
- Booking chaos? A booking link, often free, USD 10 a month at the top end.
- Quotes taking too long? A template document with the boilerplate already written.
- Bookkeeping mess? A bank feed into your accounting software, which you probably already pay for.
Here is the rule of thumb that decides it. Count the variants. If the work has fewer than about ten distinct versions, a template wins — it is free, instantaneous and cannot be wrong. If the work is genuinely varied text, where every instance differs and a template would be wrong more often than right, that is where AI earns its place.
Most "we need AI for this" turns out to be nine variants and a missing template.
Then run two honest weeks
Six rules. Skipping any one of them means the trial will not tell you anything.
One named task. "Replying to Instagram messages about pricing" is a task. "Marketing" is not.
Measure the week before. This is where most trials die. For five days before you start, count: how many, how long in total, how many turned into something. Without a before, there is no after — there is only the feeling that things are better, which is what every new tool produces in week one regardless of whether it works.
One number, chosen in advance. Hours, or money, or conversions. Written down before you begin, so you cannot pick the flattering one afterwards.
Fourteen days, one tool, one way of working. Not three tools. Four days each teaches you nothing about any of them.
A kill line, written while you are unexcited. "If I am not saving at least three hours a week by day 14, I stop and cancel." Decide this on day zero. On day 14 you will be attached to it.
Count the editing. The commonest hidden cost: the draft takes two minutes and fixing it takes eight. Your total is draft plus edit plus the ones you threw away, not draft alone.
What it looks like with real numbers
A two-person cleaning company in Warsaw, trialling AI-drafted replies to enquiries.
Baseline week: 42 enquiries, 3h 10m total on replies, 11 became bookings.
Week 1: 39 enquiries, 1h 50m, 12 bookings. Week 2: 45 enquiries, 2h 05m, 13 bookings.
Average after: just under 2 hours a week, against 3h 10m before. Saving about 70 minutes a week, roughly 5 hours a month. At PLN 90 an hour for the owner's time, that is about PLN 450 of time against a tool costing PLN 80 a month.
Keep it. But read the second column too. The booking rate moved from 26% to 30% on small numbers over two weeks — that is noise, not evidence. The honest conclusion is: it saved time, it did not demonstrably sell more. Do not let yourself round that up.
Three endings, all fine
Keep. It cleared the line. Move it into normal working and set a reminder to re-check in six months.
Kill. It did not. Cancel the same day, not "next month" — next month is how a dead trial becomes a two-year subscription. Write two lines about why, so you do not buy the same thing again next year.
Narrow. The most common real outcome. It worked for one of the three things you tried it on. Keep that one, drop the rest, and stop paying for the version that does all three.
The part that does not change
You will run this loop again, because the tools keep changing. Some of what you learn here will be obsolete within a year — models will get cheaper and better, and half the specialist products on the market today will be a feature inside something you already own.
The method survives all of that. Find where your hours go. Test one thing against a measured before. Count the editing. Name the kill line in advance. That is not an AI skill. It is how you evaluate anything that wants your money.
Before you move on