Which repetitive task should you automate with AI first?

Start where the next steps are known

Choose a task you repeat often, with varied wording but known next steps: reading an enquiry, identifying its subject and preparing a reply. Anthropic distinguishes predefined workflows from agents that direct their own process. For a first pilot, favour a fixed path whose draft you can check quickly.

Follow one request from input to draft

This is an invented example, with no customer data or measured results. The request reads: “Could you quote for eight laptops? We need them next month.”

• The trigger, or event that starts the process, is your manual submission of this fictional request to a test queue. Nothing is connected to an inbox.
• The input is the request text. Ask the AI to extract the product, quantity and requested timing, then classify the subject as quotation, support or other. An absent value must remain unknown.
• Apply exact rules separately: only the quotation category with a positive whole-number quantity continues to drafting. Missing quantity, unsupported categories or a conflict flagged during extraction become exceptions: items set aside for a person to resolve.
• The output is the extracted fields beside a draft reply. Here, it should ask which delivery date “next month” means. It must not invent a price, stock level or delivery promise.

Make review and failure explicit steps

Keep the draft waiting for someone to accept, edit or reject it. Power Automate’s approval action illustrates a flow that waits for a human response. In this pilot, review is the endpoint; any sending would require a separate decision.

Also define a stopping point for technical failures. AWS Step Functions documents retry limits and alternative paths for errors. For this draft-only trial, allow one retry after a temporary connection error, then stop and flag the item. A missing fact belongs with the reviewer, not in a guessing loop.

Test whether the handoff reduces work

Before connecting a live service, try a small set of invented requests: complete, missing a quantity, contradictory and unrelated. Write the expected fields and next step before running them. Anthropic recommends task-specific evaluation, measurable criteria and difficult cases.

For each case, record wrong fields, missed exceptions, draft corrections and total time including review. Compare with manual work. A successful run alone proves no saving. Continue only if results meet your criteria; if reviewing takes longer than writing, narrow the task before adding connections.