There is no AI for that
If you scroll through LinkedIn these days, you might think AI can solve any problem in minutes. Summarize a document? Done. Write an email? Easy. Automate your workflow? Just connect n8n, Zapier, make or similar and let the magic happen.
But if that's true, why do most companies still struggle to move AI beyond pilots? Why does Gartner predict that over 40% of agentic AI projects will be canceled by 2027?
I wanted to find out for myself. So I picked a simple personal use case and tried to automate it end to end. What I learned surprised me.
The Use Case
It started while sorting through a pile of paper mail I had collected over a few weeks. For some time now, I've been scanning these documents, tagging them and archiving them in a note-taking app. It works, but it's tedious. Every letter needs to be handled manually. That felt like a perfect job for AI.
The goal was simple: scan a document with my phone, send it to an LLM for classification, then archive it automatically. The less I have to touch the document, the better.
The Proof of Concept
I started with a quick feasibility check. I took a few already scanned documents, wrote some instructions for ChatGPT and uploaded them. It worked beautifully. The classification was spot on. The naming was sensible. The summary was useful. In under an hour, I had validated that the core AI part of my little project worked exactly as I had hoped.
I was optimistic. If the hard part was this easy, the rest should work like a charm… until it didn't.
The Scanning Problem
First, I needed a reliable way to scan documents into a folder. AI doesn't help here. This is just tooling. I wanted proper PDFs, not blurry photos. After testing a few apps, I settled on the Google Drive scanning feature. It's free, works well and drops everything into a folder that acts as my inbox.
The Workflow Layer
Next, I needed something to pick up documents from that folder, send them to an LLM and collect the results. This is where things got complicated.
Neither ChatGPT nor Gemini offer this kind of automation out of the box. After some research and a consultation with ChatGPT itself, I landed on Zapier.
Setting up the workflow taught me a few things I hadn't anticipated. The free tier wasn't enough. Using GPT via the API is different from using the chat interface. PDFs needed OCR before the model could read them. Prompts that worked in the chat didn't transfer directly to the API.
So I added another tool for OCR, rewrote my prompts and spent several hours learning how Zapier actually works. Interestingly enough, Zapier offers an internal AI which should be able to configure the tool for you, but that turned out to be a completely different rabbit hole. Eventually, I had a flow that could fetch a document, extract the text and get a classification from the LLM.
Progress. But we're not done yet.
The Destination Problem
The final step was sending the document and its classification to a note-taking app. I chose Notion as my archive. That meant learning another API (or at least how to get Zapier to talk to Notion), figuring out whether to store the PDF in Notion or on a separate file system and extending the Zapier flow once more.
A few more hours later, I had something working. Sort of.
The Cost Reality
By this point, I had hit the limits of every free tier involved. To run this workflow for the volume I intended, I would need paid versions of GPT, Zapier, an OCR tool, Google One for storage and Notion. That adds up to roughly 50 € per month, plus processing fees. For a personal document archive.
I'm sure a more experienced builder could find a cheaper path. But for someone like me who isn't deeply technical, this was the reality.
What I Actually Learned
The AI part of this project took one hour. Everything else took days.
Scanning, workflow orchestration, API integrations, data storage, cost optimization. None of that is AI. It's traditional IT work. And it made up the vast majority of the effort.
This experience gave me a new perspective on why AI adoption in enterprises is slow. Yes, AI is a powerful tool for solving individual tasks on demand. That alone is valuable. But automating a workflow end to end in a reliable, production-ready way? That's not an AI problem. It's an integration problem.
The bottleneck isn't the model. It's everything around it. And don't get me wrong. I'm not saying AI is nonsense and that it's not worth the effort. Having the opportunity of automating such repeatable tasks in such an easy way is most likely worth the effort. But you need to be deliberate about the additional effort. Mapping out the workflow, selecting the tools and setting up a solid testing strategy still makes the difference between a nice POC and a great AI solution in production.
