I spent three months watching YouTube videos about AI automation before I built a single thing. I bought courses. I made lists. I opened Claude, stared at the blank prompt, and closed the tab.
That is the real story behind most “I automated my business with AI” posts. Not a smooth launch. A long, embarrassing ramp-up period where the gap between what you imagined and what you actually know how to do is extremely wide.
Today I run 18 automated agents on two Mac Minis in my home office. They publish LinkedIn posts, write blog articles, scrape leads, send Telegram alerts, track ad performance, and manage outreach campaigns. They run while I sleep and while I am at my kids’ soccer games. I am not a programmer. I have never taken a computer science class. I built all of it over about eight months by starting with one small, ugly automation and stacking from there.
This post is what I would hand to myself on day one.
What an Automation Actually Is (No Jargon)
Strip away the AI marketing language and an automation is just this: a piece of software that does a repetitive task without you touching it.
That is it. The “AI” part just means the software can now handle tasks that used to require human judgment, like reading an email and deciding what to do with it, writing a first draft of a post, or deciding which leads are worth following up on.
Before AI, automations were limited to rules. “If the spreadsheet gets a new row, send an email.” Now they can handle ambiguity. “Read this article, extract the three most important points, and write a LinkedIn post in my voice.” That shift is what makes this era genuinely different.
But the fundamental structure is the same: a trigger (something happens), an action (the software does something), and an output (you get a result without lifting a finger).
The Mistake Almost Everyone Makes First
They try to automate something complicated.
They want to build an AI that monitors their inbox, reads every email, drafts a reply, checks their calendar, cross-references their CRM, and sends the response after a 30-minute delay to look human.
That is a ten-step automation. Each step has a failure mode. If step three breaks, steps four through ten never run. And when something breaks, you have no idea where.
I made this mistake. My first automation attempt was a full content pipeline: research a topic, write the article, generate an image, upload to WordPress, post to LinkedIn, send a Telegram confirmation. Seven steps chained together. It broke constantly. I spent more time debugging than the automation saved me.
Here is the rule I now follow: start with one step. Get that working. Add the next step. Repeat.
The 3 Automations I’d Build First
If I was starting over today with zero technical background, here is the exact sequence I would follow.
Automation 1: A Daily Briefing in Telegram
Every morning at 7 AM, my system sends me a Telegram message with: the day’s revenue snapshot, which ad campaigns are running, any alerts that fired overnight, and three things I need to do today.
This sounds fancy. The core of it is a Python script that runs on a schedule, pulls a few numbers from a spreadsheet, asks Claude to summarize them in plain English, and sends the result to my phone via Telegram’s free API.
Why start here? Because it is one script, one output, and you will actually see it every morning. Automations you never see feel pointless. This one lands in your pocket before your coffee is done.
Tools needed: Python (free, already on most computers), Anthropic API (Claude), Telegram bot (free to set up in 5 minutes). Total cost: under $2 a month in API calls if you use a small model.
Automation 2: Social Post Drafts to a Staging Area
The second thing I would build is a daily content draft. A script that wakes up, picks a topic from a list, asks an AI to write a post in your voice, and drops the draft into a folder or a scheduling tool for you to review before it goes live.
This is not fully autonomous publishing. That comes later. This is just removing the blank-page problem. The hardest part of content creation is starting. If you wake up and there is already a draft waiting, you spend five minutes editing instead of forty-five minutes staring at the cursor.
I use Typefully as the staging area for my business partner Devon’s X (Twitter) posts. The script writes five posts, loads them into Typefully, and Devon reviews and approves. He went from posting twice a week to posting daily because the friction dropped to almost zero.
Tools needed: Claude API or xAI’s Grok API (I use Grok Mini for X posts because the voice is a better fit), Typefully or a simple Google Doc as the staging area.
Automation 3: A Lead Capture and Alert
The third automation I would build is a lead alert. Something that monitors an email inbox, a web form, or a Google Sheet, and the moment a new lead comes in, sends you a Telegram message with the name, what they want, and a suggested first reply.
This one has an immediate and obvious ROI. Response time is one of the biggest factors in whether a lead converts. A 2021 Harvard Business Review study found that responding within an hour makes you seven times more likely to have a meaningful conversation than responding even one hour later. An alert that hits your phone in real time, with context already assembled, is a genuine competitive advantage.
I built this for my water treatment business. A lead fills out a form on the website. Within two minutes, I get a Telegram message with their name, address, what they asked about, and a suggested reply. I tap reply, edit two words, and send. The whole thing takes 90 seconds.
The Tools I Actually Use
I am not going to list 40 tools. Here is what is actually running in my system right now.
Claude Code (Anthropic): This is my main AI brain. It writes articles, drafts emails, analyzes ad performance, plans strategies. I run it in a terminal on my Mac. It can read files, write code, and execute commands. For anyone serious about building automations, this is the tool that changed everything for me.
Ollama on a local Mac Mini: I run a second Mac Mini as an inference server. It runs open-source models locally for free. I use it for status checks, data validation, and anything where I do not need Claude-level reasoning. A heartbeat check does not need a $20/million-token model. It needs a correct yes/no in under a second. Ollama gives me that at zero cost.
Python: Every automation I run is a Python script. I did not know Python when I started. I still would not say I “know” Python. I describe what I want to Claude, it writes the script, I run it. When it breaks, I paste the error back into Claude and ask it to fix it. This loop works. I have never written a line of Python from scratch.
Supabase: Free database in the cloud. I store leads, track which emails have been sent, log what the agents have done. You do not need a database for your first three automations, but when you want to start tracking things over time, Supabase is the easiest free option I have found.
Telegram Bot: Free, instant, reliable. Every alert and notification in my system goes to Telegram. Setting up a bot takes about eight minutes following the official docs.
Typefully: Social post scheduling. Drafts go in, I review and approve, they post on schedule. Costs $19 a month and has saved me hours every week.
What It Actually Costs
This is the part nobody is honest about, so I will be.
My monthly AI costs (August 2026):
- Anthropic (Claude) Max plan: $100/month flat. Covers all my writing, strategy, and agent work.
- xAI (Grok): Under $10/month for tweet writing (small model, high volume).
- Ollama on Mac Mini 2: $0. I own the hardware. The models are free and open source.
- Supabase: $0 (free tier covers everything I need).
- Typefully: $19/month.
- fal.ai (image generation): Under $5/month for blog images.
Total: roughly $135/month to run 18 automated agents that handle content, outreach, lead management, and reporting across four businesses.
Before I built this system, I was paying a social media manager $800/month and a VA $600/month to do tasks these agents now handle automatically. The math is not complicated.
Your first automation costs nothing. The Anthropic API has a free tier. Python is free. Telegram is free. You can build your first working automation for zero dollars today.
The Mistakes I See Most Often
Building in secret instead of building in public. The accountability loop matters. Tell someone you are building this. Post about it. The social pressure keeps you moving when the inevitable breakage happens.
Confusing complexity with value. My most valuable automation is four lines of Python. It checks a number, compares it to a threshold, and sends me a Telegram if it is out of range. I have others that are 300 lines and provide about 10% as much value.
Not reading the API docs. Every hour I have spent reading documentation has saved me ten hours of trial and error. When you use a new API, read the docs first. I know this sounds obvious. Almost nobody does it.
Storing API keys wrong. Do not put your API keys in your code. Put them in a file called .env and load them from there. If you accidentally share your code, you do not want your API keys in it. This is not advanced security hygiene. It is the baseline.
Trying to automate something you do not understand yourself. If you cannot describe the task clearly enough for a new employee to do it, an AI cannot do it either. Clarity of instruction is the skill. Build it before you build the automation.
The Mental Shift That Made Everything Click
I stopped thinking of AI as a tool I use and started thinking of it as a team member I manage.
My agents have names. They have jobs. They have a workspace (the Mac Minis). They have a daily schedule. When one of them breaks, I debug it the same way I would handle a contractor who made a mistake: figure out what went wrong, fix the process, and make sure it does not happen again.
This sounds like a weird way to think about software. But it changes how you approach building. You stop asking “how do I make this code work” and start asking “what does this agent need to know to do the job right.” The second question is easier to answer and produces better results.
What to Do Next
Here are the five steps I would take this week if I was starting from zero today:
- Set up a Telegram bot. Search “BotFather Telegram” and follow the 8-minute setup. Get your bot token. This is the output layer for almost every automation you will build.
- Get Claude API access. Go to console.anthropic.com. Add $5 in credits. This is your AI brain. The free tier covers everything you need to start.
- Ask Claude to write you a “hello world” automation. Tell it: “Write a Python script that uses the Anthropic API to summarize today’s date and one thing I should focus on, then sends the result to my Telegram bot.” Paste in your API key and bot token. Run it. When it works, you have crossed the hardest threshold.
- Pick one repetitive task you do every week. One task. Not three. Write out every step of that task in plain English. Then ask Claude: “Here is a task I do manually every week. Can you help me automate it with Python?” Work through the conversation until it builds something that runs.
- Let it run for 30 days before you add anything. Resist the urge to build more. Watch what breaks. Fix it. By month two, you will understand your system well enough to stack the next layer on top of it without everything collapsing.
The first automation takes the longest. The second one takes half the time. By the tenth, you are moving fast enough that the bottleneck is ideas, not execution.
That is where you want to be. Start small. Start today.
Jesse Navarro is a business operator and AI systems builder based in Meridian, Idaho. He runs AI agent teams for his own businesses and for clients who want results without becoming programmers.
