Ask AI which parts of your job it can help with, then test the answer

In this guide

1. The useful starting point
2. Describe your real week
3. Sort tasks by the kind of help they need
4. Ask a better version of the question
5. A worked example: a project coordinator
6. Choose a first experiment with clear boundaries
7. Build skills alongside the workflow
8. Recover from an overambitious recommendation
9. Questions you might be asking
10. Sources and further reading

The useful starting point

List the tasks you actually perform, ask AI to propose specific assistance for each, and test one low-risk workflow against a quality checklist. Compare total effort including review and corrections before deciding what to automate.

“What parts of my job can you do?” is a useful question to ask AI. It is also incomplete. The assistant needs to know what the work involves, what good looks like, what information it may use, and what happens if it gets something wrong.

A task audit turns the broad question into a practical experiment. You do not need to redesign your entire role. Begin with one recurring piece of work that is easy to describe and possible to check.

Describe your real week

For a few working days, keep a simple log of tasks as they happen. Include the unglamorous work: moving information, checking a record, preparing a meeting, formatting a document, or asking someone for missing details.

Write verbs rather than job titles. “Prepare a weekly client status summary from approved project notes” gives the assistant something concrete. “Be a project manager” does not.

Record frequency, approximate time, inputs, output, and the person who uses the result. Add the parts that require judgment or a relationship. The purpose is to see the work clearly, not to make every activity look automatable.

Sort tasks by the kind of help they need

Some work does not need AI. A reliable rule can be easier to inspect when the inputs and outcome are predictable. Save language-model judgment for the parts where interpretation is genuinely useful.

Ask a better version of the question

Here is my task log, the tools I use, and the information I am allowed to share. For each task, propose a specific form of assistance: a simple rule, an AI draft, an AI analysis, or no automation yet. Explain required inputs, a quality check, likely failure modes, and what must remain a human decision. Rank a few first experiments by practical usefulness and review burden. Do not claim you can access tools that are not connected.

Ask the assistant to show its assumptions. If it recommends automatic customer replies but has no approved policy, current records, or review process, the missing context is part of the task.

A worked example: a project coordinator

Imagine a fictional coordinator who spends 40 minutes preparing a weekly status update. She reads task notes, identifies blockers, checks owners, and writes a concise message.

Her first AI experiment uses an approved export and asks for a draft with separate sections for completed work, blockers, decisions needed, and missing information. The assistant is not allowed to invent an owner or a completion date.

The first attempt takes eight minutes to prepare and 20 minutes to review. The total is 28 minutes, a 12-minute reduction compared with this example's 40-minute baseline. That is more honest than claiming the task now takes eight minutes.

During review, she finds that the model treats “awaiting approval” as “complete.” She adds a clear status definition and tests again with a different project. If the same error persists, she may choose a structured template or a simpler rule for that part of the update.

These numbers illustrate how to evaluate effort. They are not a promised saving or a result from a real customer.

Choose a first experiment with clear boundaries

1. Pick one recurring output. Avoid starting with a broad role or a process involving many teams.
2. Define acceptance. Write the checks a reviewer would use to approve the work.
3. Choose representative inputs. Include an ordinary case and at least one incomplete or awkward case.
4. Keep the first result reversible. Draft a document or create a review item before sending or changing anything.
5. Measure total effort. Count preparation, generation, checking, corrections, and cleanup.
6. Decide whether to keep it. Continue only if the quality and effort justify the added complexity.

Use a plain scoring scale if it helps, but make the criteria visible. A numerical score is a decision aid, not scientific proof that one task should be automated.

Build skills alongside the workflow

As you test, learn how the task works more deeply. Identify which inputs matter, which decisions recur, and where the information breaks down. That understanding is useful even if you eventually choose a different tool.

Keep a small record of successful instructions, failed cases, and quality criteria. You are building a reusable working method, not just a collection of clever prompts. Share it with colleagues in a form they can inspect and improve.

If your organization has approved tools or data-handling requirements, use those. An assistant's ability to process a file is not permission to upload it.

Recover from an overambitious recommendation

If AI claims it can handle the entire process, ask it to break the work into observable steps. Identify which steps require access, which require human judgment, and which can be checked automatically.

If review takes longer than doing the task yourself, simplify the output or improve the source material. If the task changes every time, use AI as a flexible assistant rather than forcing a fragile automation. If an error would be costly, keep explicit review before action.

Repeat the audit as your work and tools change. The useful habit is asking where assistance can improve a specific outcome, then checking whether it actually does.

Questions you might be asking

Should I automate the task that takes the most time?

Not automatically. Consider how clearly it can be specified, the available inputs, the consequences of error, and the review effort. A smaller, repeatable task can be a better first test.

How do I measure time saved honestly?

Compare the original process with total assisted effort, including setup, review, corrections, and cleanup. Use more than one representative case before drawing a conclusion.

Does every automation need an AI model?

No. Predictable transfers, checks, and calculations often work well with ordinary rules. Use AI where interpretation or language work adds value.

Sources and further reading

• Make agent tools and scenarios
• n8n human review for AI tools
• Anthropic evaluation guidance

Sources checked October 16, 2026. Review the linked publisher information for current details.

Put one idea to work

Separate rules, AI assistance, review, and recovery before turning a useful task into an unattended process.

Map your first workflow →Explore more from Automately