Start with frequent, predictable tasks that produce a reviewable draft and can safely fall back to a manual process.
Look for friction, not impressive demos
Spend two ordinary workdays noting tasks you repeat, information you copy, and moments when you wait for something you already know how to do. Include small annoyances: renaming downloads, preparing a status update, sorting meeting actions, or finding the same reference link.
The best first automation often feels unremarkable. It solves a recurring inconvenience with a clear rule and an output you can inspect. A complicated workflow that runs once a quarter may be a poor starting point.
Use four filters
- Frequency: How often does this happen, and how much time does it really take?
- Predictability: Are the inputs similar, or does each case require a fresh decision?
- Checkability: Can you quickly tell whether the result is correct?
- Reversibility: Can a mistake be corrected before it affects a customer, colleague, or account?
These filters are a decision aid, not a scientifically validated scoring system. When in doubt, automate preparation first and retain approval before anything is sent, deleted, purchased, or changed.
Three practical starting points
Meeting follow-up: Turn approved notes into proposed actions with owners and due dates. Missing owners stay unassigned. A person confirms the list before sharing it.
Weekly status: Assemble completed work, blockers, and upcoming decisions from a fixed template. Keep a link to each source and flag items whose status is unclear.
Research collection: Gather new material on one narrow question into a digest with dates and links. Review the sources before using the findings in a decision or public claim.
Choose ordinary automation when rules are enough
Some tasks do not need AI. A calendar reminder, a spreadsheet formula, or a rule that places files in folders may be simpler and easier to trust. Use AI where language varies, information needs synthesis, or a draft benefits from interpretation.
For example, “Remind me every Thursday” is a fixed rule. “Review these notes and propose the decisions I should bring to Thursday’s meeting” requires interpretation and deserves review.
Estimate net value before building
Use actual observations where possible. In a fictional example, a task occurs five times a week and takes ten minutes each time. A new process takes two minutes to run and three to review, saving twenty-five minutes weekly before maintenance. If setup takes three hours, it needs several weeks of use to repay that effort.
Pick one candidate and define success before changing anything: fewer missed actions, a shorter preparation time, or a more complete report. Include an error measure so faster does not quietly become worse.
Prompts to put it into practice
Replace the brackets with your context. Use information you have permission to share.
Rank my candidates
Evaluate these tasks for frequency, predictability, checkability, and reversibility: [list]. Ask for missing information rather than inventing estimates. Recommend one small pilot, explain the tradeoffs, and identify what should remain manual.
Keep the first version small
Design the smallest useful version of this workflow: [task]. Include its trigger, required inputs, draft output, approval step, failure notification, and manual fallback. Suggest a non-AI option if fixed rules would suffice.
Try it on your work
List five repetitive tasks. Choose the one with the clearest inputs and easiest review, then test it manually with AI before connecting any systems.
Check the result
- The proposed saving includes preparation and review.
- I know what happens if the automation fails.
- External actions require appropriate authorization.
