I Built AI Agents From Scratch and It Cost Me More Than I Expected

I told myself at the start of this year that I was going to build something using AI. Not just use AI tools, but actually build something with it. I kept seeing people online make it sound simple enough that almost anyone could do it. So I figured, why not me?

What started as curiosity became a proper rabbit hole.

How It Started

It began with a YouTube video. I uploaded it as a reference point, a demonstration of what AI agents were capable of doing through Telegram. That video became the seed for everything that came after. I watched it, took mental notes, and thought to myself, "I can do this."

Well, I could! But not without a few surprises along the way.

Building Natalynn and Personal Assistant

What I did not expect was how quickly things could come together when you actually commit to it. In less than three hours, I had built working agents from scratch. Three of them, actually. Natalynn (client facing agent), Personal Assistant (response only to me - does my scheduling, search, read, draft emails and documents), and an Admin Agent (to handle more complex tasks). Not prototypes. Not demos. Actual, functional agents doing actual things.

That part genuinely surprised me. The tools available today make it possible to go from zero to something working in a sitting. I used Claude Code running through Windows PowerShell, set up an Ngrok server, and learned how to wire everything together working on .env files and API keys. None of that was in my vocabulary few weeks ago.

It was a real learn-by-doing experience. You figure out what each piece is for, why it matters, and how they connect, because nothing works until they all do.

Natalynn negotiating meeting times and booking it after checking with me

Personal Assistant booking my calendar and giving dinner suggestions

What Natalynn Can Do (for now)

Natalynn operates through Telegram as its primary interface. When an external party reaches out, the agent reads the intent of the message and routes it down one of three paths. For appointment requests, it checks my calendar for availability, proposes times directly to the external party, handles any back-and-forth negotiation autonomously, and only loops me in to approve or decline a finalized slot before confirming. 

For service enquiries, it presents the full list of services with pricing, collects the person's contact details as a lead, and issues a quotation if they express interest in moving forward. Beyond business, if the person is in the mood for something lighter, the agent can throw in a joke. The whole point is that I only touch the workflow at the one moment that actually requires my decision, and everything else is handled automatically end to end.

Natalynn & Personal Agent workflow (Click to expand view)

The Part Nobody Talks About Enough

Here is where I want to be straight with you, because I do not see this discussed honestly enough.

It is expensive.

Not in a casual, "oh you might spend a bit more than expected" kind of way. I mean genuinely expensive, in ways that sneak up on you.

Every time you make a change to a workflow, add new functionality, or fix a bug, you are using tokens. Every query your agent processes uses tokens. Every step in an automated workflow uses tokens. Even getting a simple reply back costs tokens as it hits the API. It all adds up, and it adds up fast.

The more you automate, the more steps a workflow takes, the more complex the task, the more it costs. Multiply the number of queries and tasks your agent handles per day across a full month, and you start doing some uncomfortable mental arithmetic.

Which led me to a question I did not expect to be asking myself after building all of this.

Is It Actually Worth It?

At a certain point, the cost of running agents and automation at scale starts to look a lot like the cost of hiring an actual person. And here is the difference: with a person, you can control the cost. A monthly salary is predictable. Token usage is not, especially when automation complexity grows over time.

What about you? Have you built an AI Agent or automate a workflow before? Would love to hear your experience.

I Automated My SEO Content. My Team Finally Stopped Complaining

Working in an AI company, using AI isn’t optional. It’s expected. It's a way of life!

So naturally, I started looking beyond just using AI for outputs. I wanted to see how it could improve the way we work. Not just faster outcomes, but smarter workflows.

I'm in a way blessed that I have a coding background, so I decided to experiment with something simple but powerful, using Google Apps Script. It’s free, flexible, and surprisingly capable if you know your way around it. I also used a combination of Claude and Minimax-M2.7 to troubleshoot and fix my code.

Me and my team now when AI and automation did all the work for us

The Problem: Content Creation Fatigue

If you’ve ever worked in marketing, you know this pain.

Coming up with SEO content ideas consistently is exhausting. Not because tools don’t exist, but because the process is fragmented:

• You brainstorm topics
• You validate keywords
• You refine prompts
• You generate content
• You edit and structure everything

Yes, AI can write. But good content still needs:

• Relevance to your brand
• Alignment with your domain expertise
• Proper structure for SEO and readability

Even with tools like Semrush helping identify keyword trends and competitor gaps, the team still had to “figure things out” every single time.

So I asked a simple question:

What if we could automate the entire workflow?

The Solution: A Lightweight Content Automation Engine

I built a simple system that connects Google Sheets, Semrush, and Gemini into one streamlined pipeline.

Here’s how it works:

Step 1: Input the Idea

The team just fills in a “Topic” column in Google Sheets.
Think of it as the marketing angle or content idea.

That’s it. No prompting gymnastics required.

Step 2: Automate Keyword Intelligence

Using Google Apps Script:

• The script runs twice a week
• It reads new topics from the sheet
• Connects to Semrush via API
• Pulls relevant and trending keywords

No manual keyword research. No switching tabs like a caffeinated octopus.

Step 3: Generate Structured SEO Content

This is where Gemini comes in. By combining those keywords from Semrush together with the given topic, I designed a structured prompt that ensures every output follows SEO best practices:

• Strong, keyword-led title
• Meta description
• Clear introduction
• 5 to 6 structured sections
• Natural keyword integration
• FAQ section
• Conclusion with CTA
• Clean markdown output
• 1200 to 1800 words

The key here isn’t just AI generation.
It’s forcing consistency in quality and structure.

Step 4: Review and Approval Loop

Once the content is generated:

• The task is marked “Completed” on Google Sheets
• An email is triggered to me and the team
• Content is reviewed, edited, and approved before publishing

Human oversight stays in place. Just without the heavy lifting.

The Outcome: Less Friction, More Output

What used to be a messy, multi-step process is now streamlined into:

1. Input topic
2. Let automation handle the heavy work
3. Review and publish

Simple. Scalable. Repeatable.

And most importantly, the team no longer dreads content creation.

Why This Matters (Especially Now)

Search behavior is changing fast.

People are:

• Starting with AI tools
• Cross-checking via reviews, Reddit, and YouTube
• Only then visiting brand websites

Which means your content needs to be:

• Structured
• Relevant
• Easy for both humans and AI to understand

This kind of workflow doesn’t just save time.
It positions your brand to show up where decisions actually happen.

Key Takeaways You Can Apply Today

1. Don’t Just Use AI. Systemise It.

Most teams use AI as a tool.
The real advantage comes when you turn it into a process.

If your team is still manually prompting every task, you’re leaving efficiency on the table.

2. Structure Beats Creativity (At Scale)

Good content isn’t just about ideas. It’s about consistency.

A strong framework ensures:

• Better SEO performance
• Easier readability
• Higher chances of being surfaced by AI

Think of structure as your unfair advantage.

3. Reduce Friction, Not Control

Automation doesn’t mean removing humans.

It means:

• Removing repetitive work
• Keeping strategic control
• Letting your team focus on refinement, not creation

The goal isn’t to replace marketers.
It’s to make them dangerously efficient.

If you’re still treating content as a manual craft, it might be time to rethink the system.

Because in an AI-first world, speed matters.
But structured speed wins.

What about you? What's stopping you from automating your workflows?

Why Learning AI Isn’t Free, But Worth It

I've spent the last few weeks playing around and creating AI agents on Relay.app, and honestly? It's been eye-opening. I'm no expert yet, but even fumbling through the basics has taught me something important: learning this stuff costs money. And yet, the possibilities for marketing are huge.

Image credit: BotSpace

Nobody Tells You About the Price Tag

AI tutorials make everything look so simple and free. Spoiler alert! It's not!

First, there's the subscriptions. ChatGPT is great for brainstorming campaign ideas and mapping out content, but if you want to do anything more serious, you need the paid version. Then there are the creative tools like Google Veo 3, PixVerse and Kling. Whatever you're using to generate or animate videos. Every time you render something, tweak the resolution, or add an animation, you're burning through credits.

Want to auto-post to your social media accounts like LinkedIn or X/Twiiter? That's another layer. You'll need to mess around with APIs and webhooks, and sometimes that means paying for extra licenses or integration connectors.

And here's the kicker…nothing works perfectly the first time. You test. You break things. You test again. Each attempt eats up more credits. The costs add up fast, and suddenly your "free learning" project has a bill attached.

What I'm Actually Using Relay.app For

Relay.app has made this whole experiment feel more practical. It connects AI to the actual work I do in marketing, which is where things get interesting.

I can set up a workflow that triggers when a blog post goes live, use AI to create different versions of copy for various channels, and add a review step so my team can approve everything before it goes out. I can plug in APIs to connect with a CRM or ad platforms, and build conditions that let me segment audiences or personalize campaigns.

It's not magic. It's just automation that actually makes sense for marketing. Instead of manually reformatting content for five different platforms, the system does the grunt work while I focus on strategy.

What I've Built So Far

Relay.app home console. As you can see, there are errors during my runs.

Here are a few workflows I've put together:

Content amplification: When a new blog post publishes (from a website of your choosing), it automatically creates posts for LinkedIn and X with AI-generated captions tailored to each platform.

Video to post: When a YouTube video goes live, it pulls the transcript, summarizes it, and sends me a draft post to review before publishing.

Campaign variations: I can generate multiple copy versions for different audience segments, ready to drop into A/B tests.

Each one costs credits. Each one takes time to set up. But once they're running, they save me hours every week.

Sample of a workflow step using prompt to summarize a transcript from a YouTube video to social posting content.

Is It Worth the Investment?

Look, not gonna lie. This isn't cheap. But for marketers, the return is real.

You can repurpose content across platforms in minutes instead of hours. You can segment audiences and personalize messaging without drowning in spreadsheets. You can move faster on campaigns because you're not bogged down in repetitive tasks.

I'm not trying to become some AI engineer here. I just want to understand how these tools can help me do better marketing at scale. Every workflow I build clarifies how automation can boost efficiency and results.

Learning AI costs money, sure. But it feels like I'm building something that'll give me an edge down the road. And in marketing, that edge matters.

What about you? Have you try bulding an AI Agent? Let me know!

P/S: This is not a paid article by relay.app