I Automated 5 Work Tasks With AI in 5 Days: What Stuck

Every week, the same half day of busywork: chasing unpaid invoices, writing the weekly status update, sorting an inbox that refills itself. So I ran an experiment: five working days, five annoying tasks, one AI automation per day. By Friday, three were still running and two were switched off for good.

That ratio is the real story. If you're a freelancer or a small team drowning in repeat work, this is the article I wish I'd read first: not how to automate work with AI, but how to know which tasks deserve it. By the end you'll have my three-question decision tree and one prompt you can steal.

Fair warning: the automation I wanted most is one of the two I killed.

Can you really automate your work with AI?

Yes, you can hand a real chunk of your weekly work to AI today, but not everything you automate deserves to stay automated. That second half is the part the tool ads skip. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear value, and weak risk controls. That's enterprise-scale, but my desk-scale week landed on almost the same number: 2 of my 5 automations got canceled too.

The difference between the survivors and the casualties wasn't the tool. It was the shape of the task.

What you need

You need three things: an AI assistant that follows written instructions (I used Claude), an automation platform to trigger it (Zapier, Make, and n8n all work; pick whichever you already know), and one afternoon per task. My total spend for the week was the AI subscription I already had plus free-tier platform accounts.

That's it. No developer badge required.

Days 1 and 2: invoice reminders and the weekly report both stuck

The first two automations survived because they share one trait: the input has the same shape every single time. An overdue invoice is always a name, a number, an amount, and a date. A weekly report is always the same five project notes squeezed into one update.

Day 1 was payment reminders. I get about 12 overdue invoices a quarter, and each polite nudge used to cost me a coffee's worth of dread. Now a trigger fires when an invoice passes its due date, and Claude drafts the email into my outbox for a final glance before sending. Here's the exact prompt:

You draft polite payment reminder emails. Input: client first name, invoice number, amount, and days overdue. Rules: under 120 words, friendly but direct, mention the invoice number and amount exactly once each, offer to resend the invoice, no guilt-tripping. If days overdue is above 30, make the tone one notch firmer. Output only the email body.

Day 2 was the weekly status update: same recipe, different filling. I paste my raw project notes in, the automation returns a three-paragraph update in my voice, I fix one sentence and hit send.

Setup for each: an afternoon. Payback: the dread is gone and the Friday report writes itself. Build it once, win every week.

First branch of the decision tree, learned here: if the task repeats in the same shape, automate the whole thing.

Day 3: inbox triage, or what silent failure feels like

Inbox triage was the automation I wanted most and the first one I turned off. The setup worked: incoming mail got read, labeled, and filed into folders. The demo was gorgeous. Then real life arrived.

Midweek I found a time-sensitive client request sitting in a folder labeled "newsletters," where it had been resting comfortably for two days. Nothing errored. Nothing warned me. The automation was working exactly as designed, just wrong.

My inbox has opinions about what counts as urgent. So did the AI. They disagreed.

The fix wasn't a better prompt. I demoted the whole idea: instead of filing mail, the AI now sends me one morning summary of what landed overnight and what looks urgent. It advises, it doesn't touch. Second branch of the tree: if a silent mistake costs more than the time saved, downgrade from automation to assistant.

Days 4 and 5: social drafts went out, file cleanup stayed

Day 4 taught me the editing test: if you rewrite more than half of what the AI produces, you haven't automated the task, you've doubled it. The social-post automation turned each new article into three ready-to-schedule posts. Technically flawless. But every draft needed so much reworking to sound like me that "reviewing" quietly became "writing, with extra steps." Off it went.

Day 5 was the sleeper hit: file cleanup. Receipts, contracts, and attachments renamed to a consistent date-client-type format and moved into the right folders. I run it as a small script through Claude Code, which is worth knowing even if you never touch a terminal; I explain what Claude Code actually is in plain English elsewhere on this site. Pure structure, zero judgment, zero opinions to get wrong.

Third and fourth branches: kill anything where review time approaches doing-it-yourself time, and keep anything that's all structure and no taste.

The decision tree that decides what stays automated

Three questions predict whether an automation survives, and you can answer them before you build anything:

1. Does the input repeat in the same shape every time? Yes: automate fully (invoices, reports, file naming). No: stop here.
2. What breaks if it fails silently? If the answer is "a client relationship" or "money," make it suggest instead of act (inbox triage).
3. Is checking the output faster than doing the task? If reviewing takes as long as writing, you've built yourself a second job (social drafts).

A task that passes all three is also the perfect candidate for something more ambitious. That's the point where I'd graduate from single automations to building an AI agent in one day, which chains these decisions together instead of automating them one by one.

The result: 3 automations and half a Friday back

Three keepers now run every week: invoice reminders, the weekly report, and file cleanup, and together they hand me back roughly half a day per week. The two rejects taught me more than the three successes, honestly. I spent one afternoon per task setting things up, so the whole experiment paid for itself within the first two weeks. That's hours back.

The morning inbox summary, the downgraded survivor of day 3, might be my favorite. It does less than I originally wanted and I trust it more because of that.

When this doesn't work

Skip AI automation for tasks that are rare, high-stakes, or different every time. A once-a-year contract negotiation has no pattern to learn. Anything involving money leaving your account deserves human fingers on the button. And if your process is currently a mess, automating it just produces mess at higher speed: fix the process first, then automate the fixed version.

Also skip it if you can't check on it weekly for the first month. Every automation drifts eventually, and the unwatched ones drift silently.

Would I do it again?

Yes, but I'd start with the decision tree instead of discovering it the expensive way. The five-day sprint format worked: one task per day is enough time to build and test without overthinking, and by Friday you have real evidence instead of good intentions.