The AI Automation Engineer: From Satisfying Videos to Swarms of Agents

This blog article describes the emergence and crucial role of the AI Automation Engineer, inspired by a quest to automate business processes. The author explores how AI automation is evolving from individual productivity to restructuring complete workflows, driven by tools like N8n and the introduction of agentic workflows where AI agents communicate with each other. Yilmaz emphasizes that this new role is essential to bridge the gap between AI’s potential and practical implementation within organizations, a necessity endorsed by Zapier CEO Wade Foster. Ultimately, the author advocates for a mindset shift where companies must redesign processes with AI as the foundation, in order to match the speed and scale of “speedboats.”

The Fruit-in-Glass Videos

Remember those short videos that appeared everywhere on your feed? Different fruits being sliced through glass or crystal. Or those other satisfying videos where objects transformed in hypnotic ways. Perfect loops. Mesmerizing content. Millions of views. I saw one of those videos on a YouTube short by Digital Academy. And then I read the description: “Made with N8n automation workflow.”

My first thought: “I need to be able to do that too.” But this wasn’t just curiosity. This was something deeper. I had been experimenting with AI-assisted coding, vibe coding, GitHub Copilot Workspace for months. I had seen how AI exponentially expanded my capabilities as a Product Manager. And now I saw something else: not just individual productivity, but entire processes that could be automated, optimized, and transformed.

Those satisfying videos were merely a trigger. What I was really looking for was: how can we fundamentally improve certain business processes with AI?

That quest brought me to N8n, to agentic workflows, to a completely new way of thinking about automation. And ultimately to the realization that a new role is emerging: the AI Automation Engineer. This is my second personal area of interest within AI. My first, AI-assisted coding as a Product Engineer, I wrote about in my previous blog post. This is the story of the second: AI automations, agentic workflows, and why I believe every organization will soon need an AI Automation Engineer.

Part 1: Wade Foster and the AI Automation Engineer The Post That Put Everything in Place

In July 2025, Wade Foster, CEO of Zapier, published a LinkedIn post that attracted enormous attention. The title: “The hottest job in tech right now is AI Automation Engineer.”

He wrote: “Last week I put an open call: if you have the skillset of an AI Automation Engineer, we’ll consider you for ANY open role at Zapier. Your response blew us away. Applications and messages FLOODED in.” Wade’s definition was clear: “As an AI Automation Engineer, you’ll spearhead an AI-first revolution within our company. You are a catalyst for practical innovation, turning everyday pain points into AI-powered workflows. You help your fellow humans focus on creativity, strategy, and connection (letting AI do the busywork).”

What struck me most was this sentence: “Lots of teams see AI’s potential but get stuck turning ideas into action. AI Automation Engineers close that gap.” This was exactly what I saw daily. AI adoption is happening, everyone is experimenting, but the actual added value often remains invisible within existing business processes. The gap between “we could do this with AI” and “we’re now doing this with AI” remains enormous. And Wade’s description of what this role does sounded familiar:

AI Workflow Triage: identify and prioritize tasks ready for automation
Rapid Prototyping: quickly build MVPs with Zapier, LLM APIs, agent frameworks
Embed With Teams: sit with teams for 2–4 weeks to observe and redesign workflows
Scale Internal Tools: transform prototypes into durable systems
Debug AI Failures: investigate errors, refine prompts, adjust logic
Teach and Evangelize: host workshops, build playbooks, train team leads

I read this list and thought: “Fuck, this is what I’ve been doing for the past few months.”

My Context: Product Manager with Automation Superpowers

Let me be clear: I’m not an AI Automation Engineer. Not yet. I’m a Product Manager at Infinitas Learning, where I work on ed-tech products like Edupano and Beleidswijzer. I love that role. The strategic thinking, the user research, the product discovery, that remains my core. But since I started with N8n and agentic workflows, something has changed. I’ve become a PM with automation superpowers. When colleagues from different departments ask questions like “Can we make this process faster?” or “Can AI help us with this?”, I not only see the solution, I can often build a working prototype within hours. And that’s exactly Wade’s point. It’s not about a new job title. It’s about a new capability that organizations need: people who can close the gap between AI’s potential and practical implementation. But before I continue about the role itself, let me tell you how I got here. Because like my AI-assisted coding journey, this wasn’t a straight line. This was trial and error, experimenting, and ultimately seeing a future vision I couldn’t ignore.

Part 2: The Journey, From Fruits to Genchi Genbutsu Phase 1: The First Steps

After that Digital Academy video, I did what I always do: I immediately got to work. Hands-on. Experimenting. This stems from a principle I try to follow: Genchi Genbutsu, “go and see for yourself.” This is a core principle of the Toyota Production System. Don’t theorize from your office. Go to the actual place, see the problem with your own eyes, experience it yourself. For me, that means: if I see or hear something interesting, a new AI tool, a new approach, I go directly to the website and create a free account with Google SSO. My personal Gmail account is a graveyard of hundreds of experimental accounts. But that’s how I learn. By doing. N8n was no exception. Created a free account, started following tutorials on YouTube (especially Nate Herk’s channel was fantastic for beginners), and tried to build my first workflow. And… it didn’t work. The first few days were frustrating. I didn’t understand the logic. Why does this node do this? What is a webhook actually? How do you connect things? I was used to visual no-code tools, but N8n felt different.

Phase 2: The Breakthrough, Gmail Calendar Agent

Eventually it worked. My first working workflow: an AI agent connected to my Gmail Calendar. You could talk to your calendar in natural language. “When is my next meeting?” “Make an appointment tomorrow at 2:00 PM with Jan.” “What’s on the schedule this week?” Pretty simple, but it worked. And there was a lot involved: First: Connecting an AI agent node to a chat model. For this I had to go to OpenAI.com, create an API key, and put money on my account. For many perhaps a threshold, but for me that was no problem. I was already used to paying for AI credits through my vibe coding experiments. Second: Writing a prompt for the AI agent node. And here I used a strategy I had been applying for months: Claude is my personal prompt engineering assistant. Even when I brainstorm or research within ChatGPT, I use Claude to generate my prompts. Claude understands what I mean, helps me structure my thoughts, and produces prompts that work. Third: Giving the Gmail Calendar node access to my mailbox. After some struggling with authentication, that worked too. And then I had it: a working N8n workflow. Agents communicating with my data. It felt like magic. This was almost two weeks after I had created my free N8n account. The free period was also only two weeks. I had to decide: continue with a paid account, or stop. I stopped. Not because I wasn’t interested, but because at that time I was mainly busy with my vibe-coding projects. N8n came along as curiosity, not as priority.

Phase 3: The Plantyn Moment

A few months later, I was invited to the AI working group of Uitgeverij Plantyn, part of Infinitas Learning where I work. We discussed various topics to create value with AI within Plantyn. And then it happened.

During a meeting, someone asked to investigate a pilot project that could be implemented with AI support. They were talking in terms of months. Maybe half a year. Feasibility study. Proof of concept. Then implementation.

I was sitting there listening and suddenly I saw it before my eyes: the N8n workflow that could solve this. I saw the nodes. I saw the AI agents. I saw the connections between systems. It was as if someone had projected a blueprint in my head.

I took on the project. And the first step was: installing N8n locally on my laptop via Docker. N8n is open source, so you can just do that. Free. No license costs.

What the working group expected in months, I had working as a happy flow in hours. Not production-ready, but working. Demonstrable. Testable.

This was my “€60 night” moment for automation. The moment I realized: this isn’t just a fun experiment. This is fundamental. This is going to change the way we work.

Phase 4: The MCP Server Breakthrough

But the real game-changer came later: the N8n MCP server for Claude Desktop. Let me explain what MCP servers are, because this concept is crucial for the future of how we work with software. MCP stands for Model Context Protocol, it’s essentially the USB-C port between large language models and existing software. Early 2025 I saw a video that blew my mind. A user was chatting in Blender, the 3D creation software with notoriously difficult learning curve. And while he was chatting in natural language, “make a cube,” “rotate 45 degrees,” “add a light source,” it just happened. Live. In Blender. No menus. No shortcuts. No tutorials. Just talking to your software. That’s what MCP servers enable. And it’s going to bring a completely different user experience of software. Big companies are working on this full force, but the value isn’t yet widely seen. For N8n, the MCP server meant the following: I connected N8n to Claude Desktop. I described my process in a brainstorming session. And Claude generated the workflow directly in my N8n instance according to my problem statement.

The result wasn’t perfect. I still had to tweak a lot, adjust nodes, improve prompts. But according to the 80/20 rule: with 20 percent effort (actually much less) I get an N8n flow that is workable and 80 percent correct.

This was my biggest time saver. No hours configuring nodes and making connections. Claude did it for me, and I refined it.

The Broader Context: Beyond N8n

But let me be clear: this story isn’t just about N8n. It’s about process automation within companies, broader than one tool.

From the moment I started communicating about this, questions came from colleagues from different departments. I saw that in almost all departments there were opportunities where processes could be improved or automated with AI. And meanwhile, big players have also jumped on this domain:

OpenAI launched their “Agent Builder”
Microsoft has AI Foundry within Azure
Google has Vertex AI
Tools like Dify offer specialized platforms for custom AI applications
Startups are popping up like mushrooms implementing AI automations for SMEs

This is no longer a niche. This is becoming mainstream. And Wade Foster was right: every organization will soon need people who can close this gap between AI’s potential and practical implementation.

Part 3: The Aha-Moment, Agents Talking to Each Other The Agentic RAG Workflow

There was one specific moment when I saw the future. It wasn’t during the Plantyn project. It was during a personal experiment.

I had built an Agentic RAG workflow within N8n. RAG stands for Retrieval-Augmented Generation, a technique where AI agents have access to specific knowledge sources to give better answers. But what I built went beyond standard RAG. I had multiple agents with different system prompts and different roles. And I had organized them in a hierarchy:

Agent 1: The Receiver

Received questions from users
Analyzed the type of question
Routed to the right specialist agent

Agent 2: The Researcher

Searched for relevant information in knowledge sources
Extracted facts and context
Gave source references

Agent 3: The Writer

Took the research from Agent 2
Generated a coherent answer
Adjusted tone of voice

Agent 4: The Quality Controller

Reviewed the answer from Agent 3
Checked for factual correctness
Gave feedback for improvement (back to Agent 3) or approval

When I got this working, it looked like agents were communicating with each other. Agent 1 spoke to Agent 2. Agent 2 delivered information to Agent 3. Agent 3 received feedback from Agent 4. And at that moment I had a vision.

A future where agents with different roles and hierarchies can run a company. Not as science fiction. Not in ten years. But soon. Maybe within five years.

Senior employees who maintain overview of swarms of agents. Who can review. Who determine direction. Who ensure quality. But the execution? The agents do that. This is where AI automation is heading. Not just automating individual tasks, but redesigning entire workflows, entire processes, entire departments around agentic systems.

Part 4: In Action, The Marketing Content Pipeline

Let me give a concrete example. This is a hypothetical scenario, but something that perfectly illustrates what’s possible.

The Problem: Traditional Marketing Content Production

Imagine a marketing team that needs to regularly produce content. Blog posts, social media posts, newsletters. The traditional process looks like this: Step 1: Marketing manager writes briefing for content writer Step 2: Content writer does research, reads articles, gathers info Step 3: Content writer writes first draft Step 4: Designer creates visuals Step 5: SEO specialist optimizes for keywords Step 6: Social media manager creates variants for different platforms Step 7: Marketing manager reviews and approves Time investment: 1 week per content piece. Sometimes longer if there are many feedback loops. The problem? Marketing teams want to produce more content but don’t have enough resources. And especially: many content ideas sit in Covey’s Quadrant 3 (important but not urgent) and are therefore never created.

The N8n Automation Solution

Now imagine the same workflow, but automated with N8n and AI agents: Input Node: Marketing manager fills in simple form:

Topic
Target audience
Key messages
Tone of voice
Desired output (blog, social posts, newsletter)

Research Agent Node:

Searches for relevant info via web scraping
Analyzes trends via API calls to trend tools
Retrieves keywords via SEO APIs
Collects competitor content for inspiration

Writing Agent Node:

Generates multiple drafts based on research
Adjusts tone of voice according to briefing
Creates variants for different lengths
Uses GPT-5 or Claude via API

Image Generation Agent Node:

Integrates with DALL-E, Midjourney or Nano Banana
Generates visuals based on content
Creates different styles and formats

SEO Optimization Agent Node:

Analyzes keyword density
Optimizes meta descriptions
Gives suggestions for headers
Checks readability scores

Quality Control Agent Node:

Verifies facts via fact-checking APIs
Checks brand guidelines compliance
Identifies potential issues
Gives go/no-go decision

Distribution Node: If approved:

Sends notification to Slack
Publishes draft in CMS (via API)
Schedules social posts
Archives in content library

The Magic: Agents in Hierarchy What makes this next-level isn’t just that it’s fast. It’s how the agents communicate with each other: The Supervisor Agent coordinates the entire process. If the Quality Control Agent finds issues, the Supervisor sends the content back to the Writing Agent with specific feedback. The Writing Agent adjusts. The Quality Agent checks again. This continues until it’s good. It’s like having a team that consults with each other, gives feedback, and iterates, but in minutes instead of days.

The Impact

Time Savings:  From briefing to ready: 2–4 hours instead of 1–2 weeks.

Scale Increase:  From 1 content piece per week to 10+ pieces per week with the same resources.

Making Quadrant 3 Possible: All those content ideas that were “important but not urgent”? They can now happen. Because the cost per piece drops dramatically.

Cost Savings:  Fewer external freelancers needed. Teams can focus on strategy and quality control instead of execution.

Quality Improvement:  Consistency in tone of voice, SEO optimization always applied, fewer human errors.

The Reality Check

But let me be clear about the boundaries: What agents DON’T do:

Strategic decisions about which content is important
Write original thought leadership content
Determine brand positioning
Realize creative breakthroughs
Complex stakeholder management

What people MUST do:

Write briefings that are clear
Final review and approval
Determine strategic direction
Add creative differentiation
Apply human judgment

The automation makes execution faster and more scalable. But the strategy, the creativity, the human touch, that remains crucial.

Part 5: Reality Check, Struggles and Boundaries Let me be honest: this all sounds great, but it wasn’t easy. Just like my Product Engineer journey was full of struggles, I also had my “hell moments” here.

Struggle 1: The API-Endpoint Confusion

In the beginning I struggled enormously with error messages within N8n nodes. What do these mean? Why doesn’t it work? The problem was conceptual: you need to understand that many N8n nodes actually call API-endpoints. A node is often just a visual wrapper around an API call. When you don’t understand that, the error messages are cryptic. “Unauthorized.” Okay, but why? “Bad Request.” Yes, but what’s bad about it? Only when I really understood the concept of API-endpoints, authentication, headers, request bodies, response formats, could I work well with all kinds of nodes to make workflows. This took some time. Lots of googling. Lots of trial and error.

Struggle 2: The JavaScript Nodes Nightmare

Another stumbling block: JavaScript nodes within N8n. Sometimes you need to write custom code to transform data, make loops, add complex logic. I initially tried to write this myself. Of course with help from AI. I chatted with Claude, had code generated, adjusted, tested. But it was slow. Frustrating. Lots of debugging. And then came the breakthrough: the N8n MCP server. From the moment I connected N8n to Claude Desktop, JavaScript nodes were automatically generated. I described what I wanted, and Claude wrote the node code directly in my workflow. This was my biggest time gain. No more hours debugging JavaScript. Claude did it, and it usually worked the first time.

Struggle 3: The Production Boundary (My SSO Wall)

In my Product Engineer post I wrote about my “SSO Wall”, the moment I realized that some things are beyond my expertise and I need a senior engineer. For automation, my boundary is this: N8n is fantastic for department-specific workflows and quick wins. But for complex, enterprise-wide, business-critical processes, I fall short. I experimented with Azure AI Foundry and Google Vertex AI. I could make prototypes. Good prototypes even. But to bring these to production? My knowledge fell short. Scalability, error handling, monitoring, security, compliance, this requires expertise I don’t have. I can do the architecture and analysis. I understand what’s needed. But the implementation? For that I need engineers. And that’s okay. Knowing boundaries is essential.

The Context: Where N8n Does and Doesn’t Belong

Let me nuance this, because it’s important:

For SMBs and specific departments:  N8n and similar tools are absolutely production-ready. They offer cost-effective, fast automation solutions and the flexibility of open-source. For smaller companies, startups, and departments within large enterprises, they’re perfect.

For enterprise-wide, business-critical systems:  For core systems that support the entire business operation, large enterprises often prefer their own, fully integrated solutions. Think Microsoft Azure or other robust enterprise platforms. These offer the guarantee of scalability, extensive support and governance that are crucial for mission-critical operations. It’s not that N8n is “not good enough.” It’s that different situations require different solutions. And knowing when which tool is right, that’s part of the expertise.

Part 6: The Future, Speedboats vs Tankers The Urgency That Companies Are Missing

Here’s where I’m concerned. I see AI adoption at companies. Everyone is experimenting. Every organization has an “AI working group” or “AI task force.” Pilots are being run. There’s talk about use cases. But the actual added value often remains invisible within existing business processes. Why? Because companies try to “fit” AI into their current processes. They ask: “How can we use AI to make our current process faster?” That’s the wrong question. The right question is: “If we were to start over today, knowing what AI can do, how would we design this process?”

The companies, the startups, the scale-ups, that put AI central and will redefine their business processes according to this reality, they will make big strides. The big tanker companies that don’t see the urgency of this? They will miss the boat.

The Future: Senior Reviewers and Agent Swarms

If the speed of developments within AI continues like this, and all signals point to that, then it will soon become possible to run a large company with: Senior employees who:

Maintain overview of swarms of agents
Ensure quality by reviewing
Determine strategic direction
Can switch between high-level vision and detail level where needed

Agent swarms that:

Do the actual execution
Communicate and collaborate with each other
Iterate based on feedback
Scale without linear costs

Reviewing may become the most important characteristic. Not executing yourself, but judging whether the execution is good. Giving direction. Correcting where necessary.

The Time of Speedboats

There’s a meme that perfectly summarizes the current situation: tankers vs speedboats. Big companies are tankers. They’re powerful, have lots of resources, dominate their market. But they’re slow. They can’t turn quickly. When they see the iceberg, it’s often too late. Startups are speedboats. They’re agile, fast, can change direction lightning-fast. And with AI, speedboats suddenly get the firepower of tankers.

A team of five people with the right automation can now do things that used to require a team of fifty. A startup that works AI-first can iterate faster, get to market faster, learn faster than a corporate that slaps AI as an “extra layer” on legacy processes. The time of speedboats is breaking. And companies that don’t see this? That stay in their comfort zone of “we’ve always worked this way”? They will wake up to a market that has raced past them.

My Position: PM with Automation Superpowers

Let me come back to the question: am I now an AI Automation Engineer? No. Not yet. I’m a Product Manager. And I still love my job. The strategic thinking, understanding users, product discovery, that remains my passion.

But if I ever decide not to work as a PM anymore? Then I wouldn’t hesitate to take the risk to start as an AI Automation Engineer.

Because I see the value. I see the urgency. I see that every organization, large or small, will soon need people who can close this gap between AI’s potential and practical implementation. Wade Foster was right. This is the hottest job in tech right now. And it’s only going to get hotter.

Conclusion: Automation Is Coming, Sooner or Later Let me end where I began: those satisfying videos of fruit in glass. What attracted me wasn’t the end product. It was what was behind it: the realization that with AI and automation we can make things, improve processes, create value in ways that weren’t possible a year ago. Automation is coming. Sooner or later. The question isn’t whether it’s coming, but when your organization takes the step. Companies can choose:

Wait until the market forces them
Experiment in silos without real impact
Or invest now in people and processes that put AI central

That last option? It requires courage. It requires acknowledging that the way we’ve always worked may not be the way we’ll work in the future.

It requires hiring or developing people like Wade Foster’s AI Automation Engineers. People who can close the gap. Who can prototype quickly. Who can help teams redesign. Who can evangelize and train. And it requires a mindset shift: from “how do we fit AI into our processes?” to “how do we redesign our processes now that we have AI?”

The speedboats have already departed. The question is: is your organization still on the dock, or are you already in the boat?

Practical Tips: If You Want to Start For those reading this and thinking: “I want to learn this too,” here are my tips: 1. Start With Genchi Genbutsu Don’t theorize. Do. Make a free account at N8n or Zapier. Follow one tutorial. Build one simple workflow. See it work with your own eyes. Abuse your Gmail account like I do. If you see something interesting, immediately create an account and try it out. That hands-on mentality is crucial. 2. Follow the Right People YouTube channels that helped me:

Digital Academy for inspiration and what’s possible
Nate Herk for N8n tutorials and beginner-friendly content
Matthew Berman for AI developments and tool reviews
AI Advantage for practical automation use cases

These consume what used to be your Netflix time. Accept that. 3. Experiment With Different Tools N8n is fantastic, but it’s not the only option:

Zapier for quick integrations without code
Make (formerly Integromat) for more complex flows
Dify for AI-app development (works really well!)
OpenAI Agent Builder for rapid prototyping

Try them all. See what fits your use case. 4. Use Claude as Prompt Engineer This is perhaps my most valuable tip: use Claude (or another LLM) to generate your prompts for other AI agents. Even when I work within ChatGPT, I use Claude to write my prompts. Claude understands what I mean, structures my thoughts, and produces prompts that work. This saves enormous amounts of time and frustration. 5. Start Small, Think Big Your first automation doesn’t need to be enterprise-ready. Start with something simple:

A Gmail notification workflow
A Slack bot that answers questions
A content aggregation tool for yourself

But do think about scalability. What would this look like if 100 people use it? If it becomes business-critical? Which edge cases are you missing? That mindset, starting small but thinking big, helps you grow toward more complex systems. 6. Know Your Boundaries The most important lesson: know when you can do something yourself and when you need expertise. For departments and quick wins? N8n is perfect. For enterprise-critical systems? Call in an engineer. Overconfidence is dangerous in automation. Knowing what you don’t know is a skill in itself.

Read the full article here: https://medium.com/@ozkanyilmaz66/the-ai-automation-engineer-from-satisfying-videos-to-swarms-of-agents-df9e447b0ef2