Find one useful AI task

AI may help with part of a repeated task if you can supply the input and check the output. Start with one job, such as preparing a draft from approved notes, sorting a non-sensitive list, or summarizing a public document.

You don’t need to decide whether your whole business is ready for AI. Decide whether one task is worth testing. The free worksheet helps you make that decision without assuming a time saving or buying a new tool first.

List tasks from the week you just worked

Record the task, who performs it, how often it happens, what information goes in, and what finished work looks like. Use a recent example rather than an ideal process. Include time spent checking and correcting it.

Keep a baseline with a source and date. If you haven’t measured the work, mark the time as an estimate. Save a permitted input and the output someone considers useful. A private document doesn’t become safe to share because a task is routine.

Select a task you can evaluate

Prefer a narrow job with an observable output and an easy fallback. A public-document summary can be checked against the document. A decision about someone’s eligibility or a sensitive personal record needs a different level of review.

Anthropic’s prompting guidance begins with success criteria and a way to test them. Write those criteria before drafting a complicated instruction. If two colleagues disagree about an acceptable output, resolve the disagreement first.

For a constructed teaching example, imagine turning an approved product note into a short FAQ draft. The note lists supported file imports and their limits. The test output should preserve those facts, answer the selected question, and flag missing information. No customer, integration, or result is invented in this example.

Check the information and the chosen service

Use public or artificial information for the first trial. For real internal material, confirm the permission, product configuration, and data-handling terms. Credentials and unnecessary personal information don’t belong in the prompt.

OpenAI’s API data documentation separates training use from monitoring logs and application state. That distinction is a reason to inspect the service you’re using; it isn’t a blanket promise for all AI tools.

Record any uncertain requirement as a dependency. If the task can’t be tested without prohibited data, choose another task or obtain the relevant authorized setup. Don’t use a plausible guess as an input.

Run a small comparison

Give the tool the task, input, output format, factual boundaries, and behavior for missing information. The downloadable prompt contains those fields. Run a typical example and a sparse example, then inspect both against the original material.

Record total hands-on time, including corrections. Keep errors alongside any improvement. Use comparable input complexity when comparing with the manual process. One easy test doesn’t prove the tool handles the full task.

If the workflow involves several tools, test the failure route before connecting live accounts. Microsoft’s error-handling guidance describes explicit failure paths and retry limits. The automation manual shows how to put those decisions in a workflow map.

Decide whether to keep it

Keep the trial when it produces acceptable work, with checking effort and costs you understand. Change the task or tool when errors are difficult to detect, important inputs are unavailable, or the manual method is still more useful.

The NIST AI Risk Management Framework treats risk management as part of design, use, and evaluation. For a small trial, name an owner, an unacceptable outcome, and a way to stop or fall back.

Choose a review date appropriate to the task instead of assuming a universal trial duration. Save what you tested, what remains unknown, and the decision. Then use the tool-selection guide if another option deserves a comparison.

You can keep the task manual or run the small workflow yourself. If a clear marketing system needs implementation and handoff, Custom AI Builds can scope that specific job. Bring the pilot record, including any failures, to the conversation.

Primary guidance checked October 8, 2026.