Choose an AI tool by testing your task
Start with the job you need to finish. A tool that drafts a friendly email well may still fail to preserve the numbers in a spreadsheet. Broad model rankings are a useful discovery signal, but your workflow has its own constraints. This guide helps you compare a small shortlist without treating a star rating as proof.
1. Write a one-sentence task
For example: “Turn a fictional incident record into a five-line handover that preserves the ticket ID, impact, owner, unresolved action and next update time.” Use public or invented material for the first trial. Before using work information, check your organisation’s approved tools and data-handling rules.
2. Define what a pass means before trying the tools
Give every candidate the same input and instructions. For the handover task, require all five fields, no invented resolution, no new deadline, and fewer than 120 words. A critical failure, such as changing the impacted service, disqualifies that result even if the writing sounds excellent. Keep these gates separate from style preferences.
Task: Write a five-line handover using only this record. Record: INC-204. Product feed delayed. Owner: Priya. The queue is recovering after a restart. Root cause is unconfirmed. Next action: check the next scheduled feed. Next update: 16:00. Include the ticket, impact, owner, action and update time. Preserve uncertainty. Do not infer that the incident is resolved.
3. Record evidence, including failures
| Check | Candidate A | Candidate B |
|---|---|---|
| Required fields | 5 of 5 | 4 of 5; owner missing |
| Unsupported claims | None in this example | Claims the incident is resolved |
| Correction time | 2 minutes | 6 minutes |
| Decision | Eligible for another trial | Fails the uncertainty gate |
Do not conclude that A is generally better from one example. Repeat with a missing field, contradictory notes and a long record. Save the prompt, result, date and available model or plan information so another person can reproduce your comparison. Include an example where the correct behaviour is to ask for clarification.
4. Include the cost of checking and correcting
A quick first draft can be expensive if every output needs extensive repair. Estimate monthly subscription cost plus the value of correction time. The worksheet below treats every task as having the same correction time; it excludes setup, training, taxes and failed attempts. It is a planning estimate, not a price quote or a promise of savings.
Inputs stay on this page and are not saved or sent to a model. With JavaScript disabled, use: subscription + (tasks × correction minutes ÷ 60 × hourly value).
5. Make a narrow decision and revisit it
Choose the candidate that clears the quality gates at an acceptable total cost. If none passes, simplify the task or keep a human-only workflow. Check export options, access requirements and current provider terms before committing. Re-run your small test when the model, prompt or source format changes. Keep a tool only while it continues to earn its place in the workflow.