7 AI "Automations" That Are Just Expensive Cron Jobs

I automate things for clients for a living, and here's what nobody in my corner of the industry says out loud: most of what I build has no AI in it. A script wakes up at 8 a.m., does one boring thing, sends an email, and goes back to sleep. That's the product.

But somewhere around 2023, "I set up a scheduled script for you" stopped selling, and "I built you an AI automation" started selling very well. Often it's the same script. Same schedule. New invoice.

The scheduler behind most of these automations is cron. It shipped with Unix in 1975. It's free, it has been free for fifty years, and it will happily run your entire "AI workflow" on a $5 server. The markup you're paying is for two letters in the pitch deck.

The n8n-Guru Economy

Open LinkedIn and search for "AI automation agency." You'll find thousands of people selling workflow setups built in n8n, Make, or Zapier. Now look closely at the screenshots they post. Almost every node in those workflows is deterministic: a trigger, an HTTP request, a filter, a spreadsheet row. Then, somewhere in the middle, one node calls OpenAI. That single node is what turns a few hundred dollars of Zapier configuration into a multi-thousand-dollar "AI agent."

To be clear, the tools are fine. I run this kind of software on my own server, along with most of my stack, and n8n specifically is genuinely good at what it does. The problem is the label. "AI" on the invoice has quietly become a markup, not a description. The buyer can't inspect the workflow, doesn't know what a webhook is, and has no way to tell a decision-making system from a glorified scheduled task.

The incentive is obvious once you see it. "Automation consultant" is a commodity title with commodity rates. "AI automation expert" carries 2026 pricing. Same person, same n8n canvas, different LinkedIn headline.

If someone is pitching you right now and you're not sure which kind you're dealing with, I made an honest flowchart for whether you need an AI consultant at all. Spoiler: the first question isn't about AI.

Here are the seven setups I see relabeled most often. For each one: what it gets sold as, and what it actually is. None of this is hypothetical. These are the workflows that show up again and again when I look under the hood of "AI automation" proposals.

1. The Weekly Report

Sold as: AI-powered business intelligence reporting.

What it is: pull numbers from an API, drop them into a template, email a PDF every Monday at 8 a.m. That's maybe 30 lines of Python and one cron entry. Every step is deterministic. The only place a language model genuinely helps is a short narrative summary on top of the numbers, and that's one API call costing about a cent per week. Not a platform. A paragraph.

2. Syncing Data Between Two Tools

Sold as: an intelligent data pipeline connecting your CRM to your accounting software.

What it is: field mapping. New row here becomes new row there. And here's the part the pitch skips: judgment is the last thing you want in a data sync. You want the same input to produce the same output, every single time, forever. An LLM in the middle of a sync isn't intelligence, it's a source of nondeterministic bugs.

3. Email Sorting

Sold as: an AI email triage agent.

What it is: mostly rules. Sender, subject keywords, has-attachment. Gmail filters have done this for free since 2004. There is a real AI case hiding in here: routing genuinely unstructured inquiries by intent, where rules would explode into a hundred edge cases. That part is legitimate. But check what you're actually getting, because I've seen "AI triage" that was three inbox rules and a model summarizing what the rules already caught.

4. Watching a Website for Changes

Sold as: AI-powered competitor monitoring.

What it is: fetch a page, diff it against yesterday's version, send a notification if something changed. ChangeDetection.io does exactly this, open source, and I've been running it for years. A scheduled fetch and a diff. The "AI" version does the same thing and then pays a model to tell you "the price changed" in a complete sentence.

5. Invoice Reminders

Sold as: an AI accounts receivable agent.

What it is: date math. If the due date has passed and the status isn't "paid," send a template email. That's an if-statement. When someone quotes four figures for this, ask them which part requires intelligence: the date comparison or the email?

6. Posting Your Blog to Social Media

Sold as: an AI content distribution engine.

What it is: an RSS feed and a posting schedule. The honest nuance: AI does help here, because rewriting the same announcement for X, LinkedIn, and Facebook so it doesn't read like a copy-paste job is real language work. That's one LLM call per post. It's worth an afternoon of setup. It is not worth a monthly retainer.

7. File Cleanup and Backups

Sold as: AI-driven document management.

What it is: moving files around by extension, date, or naming pattern. Fifteen lines of shell script and rsync. Backups are the strongest example on this whole list, because a backup is the one place where you absolutely do not want a model improvising. Boring and identical every night is the entire feature.

The Litmus Test: Does the Task Need Judgment, or Just a Schedule?

"Does automation require AI?" is a question people actually type into Google, often enough that it shows up in the People Also Ask boxes. The answer is no. Automation ran the world for decades before LLMs existed: payroll, backups, billing, monitoring. A script on a schedule is not the fallback option. It's the default, and it still covers most business automation today.

Here's the test I use, for client work and for my own stack. Write the task down as a checklist. Then ask: could a bored intern follow this checklist without ever making a decision?

If yes, it's a script and a cron job. Fetch, compare, format, send. Done.

If any step is "read this and figure out what it means," then that step, and only that step, is an AI candidate. Messy human language coming in. Interpretation required. Output that changes depending on context.

That's the whole AI automation vs scripts question in two sentences. Schedules are free. Judgment costs money. Pay for judgment only where judgment actually happens.

There's a cost asymmetry hiding in that rule, too. Script runs are effectively free, testable, and identical every time. LLM calls cost money per run, are occasionally wrong, and need someone to review the output. A well-built automation wants as few model calls as possible, placed exactly where the judgment lives, and nowhere else.

Where AI Actually Earns Its Keep

I'm not anti-AI. Half my agency runs on AI tools I use daily, and I sell AI consulting. In automations specifically, there are three jobs where a model genuinely earns its place:

1. Unstructured input to structured data. Turning emails, PDFs, and voice notes into clean fields. Rules can't do this. Models can.
2. Classification where rules explode. Routing by intent when the edge cases would need a hundred if-statements that break every month.
3. Summarization. Turning 500 rows of Search Console data into three sentences a client will actually read.

Here's how that looks in my own stack. The pipeline is boring on purpose: scheduled scripts fetch the data, diff it, store it. Deterministic end to end. Then exactly one step hands the result to a model to write the human-readable summary. And the design rule I never break: if the model call fails, the automation still works. I get raw numbers instead of prose. Annoying, not broken.

That's the architecture question to ask any vendor: if we removed the AI step tomorrow, what stops working? If the honest answer is "the summary gets less pretty," you now know what you're really buying.

Ask the One Question

I'll admit the awkward part: writing this is bad marketing for me. I sell AI consulting, and here I am telling you most automations need cron and 20 lines of code. But I'd rather lose the inflated project and keep the client, because the businesses that got sold a $3,000 alarm clock eventually figure it out, and they remember who told them the truth first.

So before you sign anything with "AI automation" in it, ask one question: which step in this workflow makes a decision?

If they can point to it, and the input at that step is genuinely messy, you might be buying something real. If they can't, you're buying a cron job with a markup.

Not sure which one is sitting in your inbox right now? Let's talk. I'll tell you whether your workflow needs AI or a $5 server, even when the honest answer is the cheap one.