Automate the boring parts: a starter playbook
How to pick the first office tasks worth automating with AI, write the recipe, keep a human checkpoint, and know when to stop adding machinery.
The fantasy version of AI automation runs your business while you sleep. The version that actually works for a small operation is humbler: a handful of recurring chores done in a quarter of the time, with you checking the output before it counts. This playbook gets you to the humble version, which is the one that pays.
Picking the first task
Good first candidates share four properties. The task recurs at least weekly, so the setup cost amortizes. It is text-shaped or document-shaped: drafting, summarizing, extracting, classifying, reformatting. You can describe doing it as numbered steps without saying “it depends” more than once. And a mistake caught a day later is annoying rather than expensive.
Concretely, the usual winners: turning meeting notes into action lists, first drafts of recurring emails and reports, extracting fields from invoices or forms into a spreadsheet, tagging and routing incoming messages, reformatting content between systems.
The usual losers: anything involving judgment about people, exceptions to policy, final numbers that move money, or a step where you would say “I just know.” If you cannot write the recipe, the machine cannot follow it.
Write the recipe before the automation
Do the task manually one more time, and this time write down every step, every input, every decision rule, and what “done correctly” looks like, including two or three real examples of good output. This document is the automation. The tooling is just what executes it.
This is the same move as briefing the model properly in prompting fundamentals: context, constraints, and an example beat clever wording every time. A recipe that fits on one page automates well. A recipe that keeps growing footnotes is your warning that the task is judgment wearing a process costume.
The human checkpoint
Every automation gets one explicit checkpoint where a person approves the output before it takes effect: before the email sends, before the row commits, before the file goes to the client. One checkpoint, placed where the risk is, not five checkpoints sprinkled everywhere until the automation saves nothing.
The checkpoint is not a temporary training wheel. For anything customer-facing or money-adjacent it is permanent, and the verification moves in the verification habit are the checklist for what the checkpoint actually checks. The honest accounting is end-to-end: time saved doing minus time spent checking. Positive, keep. Negative after a fair trial, kill it without sentiment.
Tools: boring beats clever
You can get surprisingly far with the assistant you already pay for plus a saved recipe pasted at the start of each run. That is automation level one, and for many tasks it is enough.
Level two is a workflow tool that runs your recipe on a schedule or a trigger. The selection logic from how to choose AI tools for a small business applies unchanged, with one addition: prefer tools where your recipe stays readable and exportable as plain text. A recipe locked inside a vendor’s visual flow editor is a recipe you will rewrite from memory someday.
Resist level three, the agent that “figures it out itself,” for anything that matters. Autonomy is exactly the property that turns a small error into a hundred copies of a small error before the checkpoint sees any of them.
Run a pilot like you mean it
Two weeks, one task, logged honestly: time spent, failures caught at the checkpoint, failures that got past it. Log the failures especially. Five minutes saved drafting is not a win if review now takes ten because trust is low.
After two weeks you have real numbers and real failure examples from your own work, which beat any vendor demo. Keep, fix, or kill, and only then pick task number two.
When to stop
Automation has a natural stopping point most businesses blow past: when the remaining tasks all require judgment, relationships, or accountability. The goal was never to automate the business. It was to clear the boring layer off your week so the judgment work gets your full attention. When the weekly chores run themselves and the checkpoint takes ten minutes on a Friday, you are done. Stop adding machinery and go do the work only you can do.