How to choose an AI tool for a task without drowning in testing
A hundred options in the results, each with a free trial and a ten-minute review. A week later there are twelve services in your bookmarks and the task is undone. The problem is not the searching but the absence of criteria. Below are five criteria in the order of applying them.
A hundred options in the results, each with a free trial and a ten-minute review. A week later there are twelve services in your bookmarks and the task is undone. The problem is not the searching but the absence of criteria. Below are five criteria in the order of applying them.
---
The problem is not finding but choosing
Search for "AI for video" and you get a hundred options. "AI for presentations" — another hundred. Every second one calls itself the best, each has a free trial and a ten-minute review.
Then somebody does the sensible thing: starts testing. Registers with three, tries them, compares. That takes an evening, then another. A week later there are twelve services in the bookmarks, seven subscriptions in the inbox, and the task it all started for is undone.
The problem is not the searching. The problem is the choosing. And it does not get solved by more thorough trawling — the trawling is what creates it.
Why choosing takes so long
The root cause: no criteria. When it is unknown what to compare on, the only method is to try everything. And everything cannot be tried.
The misconception: "I need to find the best one". There is no best — there is one that is sufficient for your task. The difference between the leader and a solid mid-tier tool is visible to an expert and invisible to a client, and finding it takes a month.
A second misconception: "first I'll learn the tools, then start working". The order is the reverse. The tool gets chosen for the task, not the task for the tool. While there is no task, any choice is arbitrary.
And a systemic layer. The market benefits from a long list: every new service is grounds for a review, a round-up and some advertising. Nobody has an interest in telling you "take one and do not touch the others for six months".
Five criteria
The order matters: if a tool fails the first, you do not look at the rest.
1. Availability and payment. Does it open, can you pay for it. A tool you cannot pay for is not a tool but a one-off diversion. Work with deadlines cannot be built on it.
2. Sufficiency for the task. Not "the best in the world" but whether it covers what you need. State the task concretely: not "I need AI for video" but "fifteen seconds in a vertical format with sound".
3. The barrier to entry. How long until a first usable result. At five hours a week, a tool needing two evenings to learn eats half your monthly time budget.
4. The cost per task rather than per month. The right question is not "what does the subscription cost" but "what will one job cost me". Pay-as-you-go is almost always better than a subscription while there is no flow.
5. What happens if it goes away. Is there anything to replace it with at acceptable quality. A tool with no backup route is a risk you take on silently and pay for in deadlines with a client.
The one-tool rule
The most useful thing you can do with a long list is stop reading it.
One tool per service until the first payment. New ones do not get tried at all until the current one starts bringing money.
That works because it removes the main trap: information overload feels like productivity. A review of a new service watched gives the same sense of movement as work done, requires no effort and cannot end in failure. So people are sincerely busy and go nowhere for months.
The sufficiency test: you did something in it twice and it worked — it is yours. The third time, look not for a new service but for a way to do it faster in this one.
When to take a new one: a client asks for something you cannot do. That is a legitimate reason. "Something new came out" is not.
Three typical mistakes in choosing
Choosing by reviews. A review shows what a tool can do in the hands of somebody who studied it for a month. Your result on the first evening will be different. What to look at is not the ceiling of what is possible but the barrier to entry.
Choosing by features you do not need. A service with forty capabilities looks better value than one with five. In practice you use two and pay for forty — and work out an interface designed for all forty.
Changing tool at the first difficulty. The first attempts in any service come out badly — that is a normal learning curve rather than a sign of a wrong choice. Changing tool at that stage zeroes what you accumulated and starts the loop again.
The sign it is genuinely time to change: you are doing what you need in it and running into a specific named constraint. "The clip length is not enough" is grounds. "Something about it bothers me" is not.
How to use a catalogue properly
A reference work exists not for studying but for searching by task.
The bad scenario: open a round-up of two hundred services and read through.
The good one: a specific task appeared → you opened the right category → took five to ten vetted options → chose one by the criteria above → closed it.
The difference is that in the second case you arrive with a question rather than for impressions.
How to test a tool in twenty minutes
Not a week of trials but a short procedure.
Take one real task rather than a practice one. The one you have to do today anyway.
Give the tool twenty minutes. Not to "work it out" but to get a first usable result. If in twenty minutes there is nothing resembling it, the barrier to entry is high for you, and that is already an answer.
Judge it on one question: would I deliver this to a client after a little polishing? Not "is it attractive" but is it usable.
Count what it cost. If the task will bring $40 and the spend was $3, the tool works. If it is the other way round, the issue is not the subscription.
Twenty minutes against a week of comparisons — and the decision gets made on your task rather than on somebody's review.
Where to start
The "Encyclopaedia of AI tools" express course — a base of more than 190 vetted services across 24 categories, selected on the principle of "recommended by a curator" rather than "found on the internet".
How to use it: whenever you have a task of the form "I need AI for something", you open the right category and get five to ten options with descriptions. Choosing from ten takes minutes. Choosing from a thousand never happens.
The price is $9. The course is included in the Basic plan, and registration is free and opens three days of full Basic access — meaning you can simply take it having paid nothing.
Alongside it are five more foundational courses, including prompt engineering with a map of the model landscape, the ten-lesson sales module and chats with 140+ AI assistants across 14 categories.
And a short cut for people with no time to choose. Often no tool has to be found at all: a ready assistant for the task — marketing, sales, HR, legal, analytics, e-commerce — closes it with no comparison of services. In that case the choice gets made once, by the name of a category.
Do one thing today: write on a sheet of paper one service, one tool and one place to look. Three lines. Everything that did not fit goes in a "later" folder, and you do not look in it until the first dollar earned.
---
Three places where the choice has already been made for you
The five criteria above apply when you choose yourself. Often that is not required.
More than 10 000 skills for Claude Code with semantic search: it searches by the meaning of a task rather than by matching words. Plus 20 in-house AI SKILLS pipelines.
More than 3 000 automation templates — before choosing a platform, check whether what you need has been built.
Cerebrum — if you do not know what to look for: you describe the task in ordinary words and get a chain of five to seven steps with the joins explained.
And more than 190 vetted services in the encyclopaedia — not for reading through but for searching by task.
The rule that follows. First check whether the task gets solved by something ready. Choosing a tool is work worth doing only when nothing ready exists.
---
What to read next
[The one-tool rule](/en/blog/one-tool-rule) — what to do with a long list.
[What prompt engineering is](/en/blog/what-is-prompt-engineering) — the other half of the cause of failures.
[What it costs to try](/en/blog/how-much-to-try-wallet) — how to compare by the price of a task.
[AI agents: what they are and what people pay](/en/blog/ai-agents-guide) — the full guide to agents, with market rates.
[Gemini as a working platform](/en/blog/gemini-as-work-platform) — a team of assistants instead of one chat.