What to automate (and what not to)
Three questions — how often, how rule-bound, how costly to get wrong — sort your chore list into automate-now, assist-only, and leave-alone.
The classic beginner mistake is automating the most annoying task instead of the most automatable one. Annoyance is a feeling; automatability is three measurable properties. Score any candidate task on these and the decision mostly makes itself:
- Frequency — how often does it run? Daily beats weekly beats monthly. A 5-minute task done 10× a day is ~200 hours a year; a 2-hour task done quarterly is 8. Automate the drip, not the puddle.
- Rules vs. judgment — could you write instructions a brand-new coworker could follow verbatim, with no questions? If yes, deterministic steps can do it. If the instructions keep sprouting 'it depends', that's judgment — either an AI step with checks, or a human who keeps the task.
- Cost of error — what happens when a run goes wrong? A misfiled ticket: someone re-files it. An email to the wrong customer: an apology. A wrong payment: a real incident. Error cost doesn't forbid automation — it sets how much checking (Module 4) the automation must carry.
The three verdicts
- Automate now — high frequency, pure rules, low error cost. Riley's email-to-ticket copying is the textbook case. These pay back in days and build your confidence for everything else. Every learner's first workflow should come from this bucket.
- Assist, don't replace — judgment in the middle, but AI can prepare and draft while a human decides: summarize the thread, draft the reply, pre-fill the form. The human's role shifts from author to reviewer — faster, and safer. Most white-collar wins live here.
- Leave alone — low frequency, heavy judgment, or high stakes: the annual budget, the sensitive personnel email, anything you'd want to explain in person. Automating these saves minutes and risks relationships. Write them on a do-not-automate list; it's as valuable as the to-do list.
Automation is an amplifier. If tickets are routed by a rule everyone works around, automating the rule produces wrong routing at machine speed with nobody watching. The sequence is always: fix the process on paper, run it manually once or twice, then automate the fixed version. (If your team has done process mapping — the AI for Business Analysts course teaches it — those maps are exactly the input this decision wants.)
Here are 10 recurring tasks from my week [list them, one line each: what it is, how often, roughly how long]. For each, score: frequency (high/med/low), rules-vs-judgment (could a new coworker do it from written instructions alone - yes/partly/no), and cost of a wrong run (trivial/annoying/serious). Then sort into: automate now / AI-assist with human review / leave alone. Explain the two placements I'm most likely to disagree with.
AI is a good first-pass sorter for your own chore list — and the 'most likely to disagree' line surfaces exactly the tasks worth thinking hardest about.