To automate tasks with AI, start with one you do every Friday: the same inputs, the same kind of output, the same deadline every week. Write the steps down, let ordinary automation do the fetching and moving, give AI only the steps that need reading or writing, and keep your approval on the result until it's been right for a month.
TL;DR
- A weekly task gives you four tests a month.
- Rules move the data; AI does the reading and writing.
- Approve every run for a month, then spot-check.
Why start with a task you do every Friday?
Because it repeats often enough to test and rarely enough that a mistake is cheap. A weekly job gives you four real runs a month to compare against what you'd have done yourself. Daily tasks pile up mistakes before you notice; quarterly ones take a year to learn from.
Good Friday candidates look like this: the weekly update you send a client, the list of invoices still unpaid with a reminder drafted for each, the timesheet roundup for payroll, or next week's schedule sent to the crew. For more places a small business gets real time back, see where AI for small business actually pays off.
How do you pick which Friday task to automate?
- The inputs come from the same places every week. The same inbox, spreadsheet or app.
- The output is written. A summary, a list, a set of drafts.
- It takes you real time. If it's five minutes, leave it.
- A mistake is easy to catch and undo. A draft you read before sending, not a payment.
Pick one. Not the Friday routine, one piece of it.
How do you write the steps down?
Do the task once by hand with a notepad open and write each step as one line, in order, the way you'd explain it to someone covering for you. "Open the job tracker. Filter to jobs finished this week. Copy customer, job and hours into the summary. Write two sentences on anything unusual. Email it to the client." If a step only lives in your head ("skip the ones that don't need it"), write down how you decide.
Which steps should AI do, and which should plain automation do?
Go down your list and mark each step. Steps that fetch, filter, move or format follow fixed rules, so ordinary automation handles them and does it the same way every time. Steps that read, summarize, sort by meaning or draft text are where AI helps. Anthropic describes the rule-following kind as workflows that "offer predictability and consistency for well-defined tasks" (Anthropic).
| Step in the weekly client update | Who does it |
|---|
| Pull jobs finished this week from the tracker | Automation |
| Copy customer, job and hours into a table | Automation |
| Write two sentences on anything unusual | AI |
| Check the draft and send it | You |
Keeping AI to its own steps makes the whole thing cheaper and easier to fix. Rick Kranz at AI Marketing Labs explains why he still uses plain automation next to AI: "So that's really where I think our linear automation shine is pre digesting information and getting it ready for AI."
How do you make it run every Friday?
With a schedule trigger in your automation tool. In Zapier, for example, "Schedule by Zapier is a built-in tool that lets you set up a Zap to run on a specific interval," and one of the options is to "Set up the Zap to trigger weekly" (Zapier help). Pick a time early enough that you can review the output before your own Friday deadline.
How do you keep a person in the loop?
Add an approval step before anything leaves the building. Zapier's version is "Human in the Loop," which lets you "pause a Zap at a certain step so a human can review," and its Request Approval action "pauses the Zap run and requests that a reviewer approve or change data you submit for review" (Zapier help).
Approve every run for the first month. Once four Fridays in a row come through with no changes, move to spot-checking. If you'd rather build the same thing inside an AI assistant, here's how to set up a one-task agent in Claude.
What goes wrong in the first month?
- A source changes. Someone renames a spreadsheet column and the fetch step pulls nothing.
- The AI over-writes. Two sentences become two paragraphs. Tighten the instruction and give it an example.
- The quiet week. Nothing unusual happened, and the AI invents something to say. Tell it what to write when there's nothing to report.
- Nobody approves. The run waits all weekend. Name a backup reviewer.
What should you automate next?
Measure the first one before you start another: how long Friday took before and after, and how often you changed the draft. Here's how to measure whether it's paying off once it's running. Then pick the next task the same way, or work through the first workflows worth building with AI automation.
If the task touches several systems or the business runs on it, AI Marketing Labs builds Custom AI Automations around one process, such as a quote, an intake or a report, and the system is owned by you outright.
Related articles
Frequently asked questions
Do I need Zapier to automate tasks with AI?
No. Make, n8n and the scheduled tasks in some AI assistants can do the same job. Pick the tool that already connects to the apps your task uses.
Which tasks shouldn't I automate with AI?
Anything where a mistake costs money or trust before you'd catch it: final pricing, payments, and replies to an upset customer. Let AI draft those, and keep the decision.
How long does it take to automate one weekly task?
Writing the steps is the quick part. Building and testing take longer, and the month of approvals is part of the job.
.jpg)