Search for an AI tool and you get listicles. Fifty tools, one paragraph each, all described as powerful and intuitive, ordered by who paid to be there. None of it helps you decide.
Here is a method that takes about ten minutes and does.
Start from the task, not the tool
The most common mistake is picking a tool and then looking for things to do with it. It goes badly because these tools are only impressive when pointed at a specific job.
Write down the actual task in one sentence, as concretely as you can. Not "help with marketing" but "turn one blog post into five social posts every week". Not "AI for video" but "put subtitles on a ten-minute clip".
That sentence usually names the category for you, and category is most of the decision.
Know which kind of tool you are looking at
Almost everything falls into one of three groups, and they fail in different ways.
General assistants — the big chat tools. Wide, shallow, free to start. Good default when the task is writing, explaining, summarising or thinking. Not the best at any one thing.
Specialists — transcription, image generation, voice, video editing, presentations. Narrow and much better than a general assistant within their lane. Worth it when the task is repeated and specific.
Built-ins — AI added to something you already use, like your email or your design tool. Almost always the right answer if it exists, because it works on your own material and adds no new place to log into.
The order to check is: built-in first, then general assistant, then specialist. Most people do it backwards and end up with six subscriptions.
Five questions worth asking
Once you have two or three candidates, these separate them faster than any review:
- Is there a real free tier? Not a trial — a tier you could keep using. If not, the tool must be clearly better, not marginally.
- Where does my data go? Look for whether your input is used to train the model, and whether you can turn that off. This matters most for anything involving work documents or other people's information.
- Can I get my work out? Export in a normal format, or is it trapped in their web app?
- Does it fit where I already work? A tool requiring a new tab, a new login and a new habit has to be much better to survive contact with a busy week.
- What happens when it is wrong? Every one of these tools is wrong sometimes. The good ones make that easy to notice — sources you can click, transcripts you can check, drafts you review. Be wary of anything that only shows you a confident final answer.
What to ignore
Benchmark scores. They measure things that rarely map onto your task, and every vendor picks the benchmark it wins.
Model names. Which underlying model a tool uses matters far less than whether the tool around it is any good. Most switch models regularly anyway.
"Powered by AI" on things that were fine before. Plenty of features are a spellchecker with a new label.
Lists that never criticise anything. If every tool in a comparison is excellent, the list is advertising.
Test it properly in ten minutes
Do not explore the interface. Take one real task you have to do anyway — a genuine one, with your own material — and complete it end to end.
Then check three things: did it actually finish the job, how long did fixing the output take, and would you reach for it again next week without being reminded?
That last question is the honest one. Most tools survive the first test and fail the second.
When free is enough
For most people, most of the time, it is. Paid tiers usually buy volume, speed and larger context rather than a fundamentally better tool.
Pay when you hit a limit that actually interrupts you — not in anticipation of hitting one. It is much easier to justify £20 a month for something you already use daily than to build a habit around something you paid for first.
---
Once you know the category you need, the n3os directory groups tools by what they do, and never ranks them by who paid. And if you are still working out which tasks are worth handing over at all, what AI can actually do for you is the better place to start.