How to choose an AI tool without wasting a month

The framework is ours; the plan details and privacy defaults it refers to were checked against vendor pages in September 2026. Some vendor links on this site are affiliate links; they never affect a recommendation. See our disclosure and method.

Step 1: define the job in one sentence

Write down the sentence "I need a tool that does X, from input Y, in under Z minutes, and the result goes to W." A real example: "I need a tool that turns a 40-minute meeting recording into a summary and action list, in under ten minutes, and the result goes into an email." That sentence rules out most of the market immediately. It also tells you what to test: the input is audio, so a text-only chatbot is out, and Descript or a chatbot with file upload is in. If you cannot write the sentence, you do not have a job yet, you have curiosity. Curiosity is fine, but spend it on a free tier, not a subscription.

Be honest about the volume too. "Three times a week" and "forty times a day" point to different plans and sometimes to different products. Every major vendor now meters usage in rolling windows rather than counting messages, so a job that involves long documents or reasoning models burns allowance far faster than short chat. Note the volume next to the sentence.

Step 2: check the free tier before anything else

As of September 2026 the free tiers are good enough that many people never need to pay. ChatGPT gives unlimited text chat on GPT-5.6 Luna. Claude gives Sonnet 5 with voice, memory and connectors. Gemini gives 3.6 Flash, Deep Research and 15 GB of storage. Perplexity gives near-unlimited basic search with three Pro Searches a day. Read the free tier's limits on the vendor's pricing page, not on a review site, because they change monthly and most reviews are stale within weeks.

The question to answer is not "is the free tier generous" but "can the free tier run my five-prompt test." If it can, do the test there. If the job needs a feature that is paid-only, for example a coding agent, video generation or a long context window, take one month of the cheapest plan that unlocks it and put the cancel date in your calendar the same day. Do not touch annual billing until you have paid monthly at least once.

Step 3: run the five-prompt test

Build five prompts from your real work, not from the vendor's demo. Use the actual documents, the actual client email, the actual spreadsheet, with anything confidential removed for now. The five should cover: the ordinary case you will run most days; the hard case that would take you an hour by hand; a case with a wrong assumption baked in, to see whether the tool pushes back or plays along; a case that needs a fact you can check, to see whether it makes one up; and a case that needs a specific output format, such as a table, a file or a particular length. Run all five in every tool on your shortlist on the same day, and score each answer out of five on whether you would send it with light edits.

Step 4: check data training and export

Before any real document goes in, find two settings and one page. The first setting is the training toggle. On personal plans all four major chatbots train on your conversations by default as of September 2026: ChatGPT under Settings, Data Controls; Claude under Settings, Privacy; Gemini through Keep Activity in Gemini Apps Activity; Perplexity under Preferences, AI data retention. Turn it off on day one. Business, Team and Enterprise plans, and every API, do not train by default, which is why a work account is often the right answer even for one person.

The second setting is export. Look for a way to download your chats, files and any custom setup such as projects, custom instructions or prompt libraries, in a standard format like JSON, Markdown or plain text. If you cannot get your work out, you are renting it. The page is the privacy policy: it should name a legal entity, a country, a contact address and a retention period. A tool with no privacy policy, or one that is a template with the company name pasted in, has told you everything you need to know.

Step 5: price against real usage, not the headline

Take the volume from step 1 and the limits from step 2 and work out which plan you would actually hit. Three traps. Usage multipliers such as "5x" or "20x" are relative to a base the vendor rarely quantifies, so compare against your own week on the free tier rather than the marketing. Bundles change the math: Google's $19.99 AI Pro includes 5 TB of storage and YouTube Premium Lite, which is worth real money if you would buy those anyway and nothing if you would not. And per-seat business plans usually have minimums; ChatGPT Business needs two seats, so a solo user pays double.

Then set a review date. AI pricing moves: OpenAI changed its API prices twice in six weeks over the summer, Google's Flash models double in price on January 1, 2027, and DeepSeek restructured three times in a year. Put a quarterly reminder in the calendar to re-run the five prompts in whatever is free that month. Our model tracker and free tiers and deals pages are updated for exactly that check.

Common jobs and where to start

Red flags that end a trial early

The printable checklist

1. Write the job sentence. "I need a tool that does X, from input Y, in under Z minutes, and the result goes to W." Add how often.
2. Shortlist three tools from the table above or the tool cards, and read each one's own pricing page today.
3. Confirm the free tier can run your test. If not, buy one month of the cheapest plan that can and set the cancel reminder now.
4. Build five prompts from real work: ordinary, hard, false premise, checkable fact, specific format.
5. Run all five in every tool on the same day and score each answer out of five on "would I send this with light edits."
6. Turn off training in each tool's settings before any confidential document goes in.
7. Find the export and download something. If you cannot, cross the tool off.
8. Read the privacy policy for a company name, a country, a contact and a retention period.
9. Price your real week, not the headline. Count bundle items only if you would buy them anyway.
10. Pick the highest score, pay monthly, and put a three-month re-test in the calendar.

Quick answers

How long should I trial an AI tool before paying?

One working week on the free tier, running the same five prompts you would use in real work. If the free tier cannot run your test at all, take a single month of the cheapest paid plan and set a calendar reminder to cancel. Never buy an annual plan before a monthly one.

What should I check before giving an AI tool my data?

Three things: whether the tool trains on your inputs by default and where the opt-out lives, whether you can export your chats, files and any custom setup in a standard format, and whether the vendor publishes a privacy policy that names a legal entity and a contact. If any of the three is missing, keep your real documents out of it.

Are lifetime deals on AI tools worth it?

Almost never. AI tools pay per token to a model provider every time you use them, so a one-off payment cannot cover ongoing costs. Lifetime deals usually end with quietly lowered limits, an older model, or a shutdown. Pay monthly and keep the ability to leave.

More guides

• 
  Guides

  Best AI chatbot 2026

  ChatGPT, Claude, Gemini and Perplexity in one table.

• 
  Guides

  Build your first AI stack

  Which three cards to combine and in what order.

• 
  Guides

  AI glossary

  Fifty terms in plain English, from token to red teaming.