I got a call last month from a small business owner who had spent four hours following ChatGPT's advice on an n8n workflow. The workflow was supposed to route customer inquiries to the right email inbox. It wasn't. ChatGPT told him to tear down the webhook trigger and rebuild it from scratch. The actual fix was a dropdown menu set to the wrong value. Took me about six minutes.
That is the thing about AI troubleshooting. When it works, it saves you a phone call. When it does not, it costs you an afternoon and sometimes makes the original problem worse.
I have been on the receiving end of these calls more often this past year. Someone spends hours with ChatGPT or Gemini, follows instructions that sound perfectly reasonable, and ends up deeper in the hole than when they started. But I have also had people tell me they fixed something in ten minutes because they asked an AI and got a straight answer. Both happen. The difference is knowing which situation you are in before you commit to it.
AI is good at problems that have been solved a thousand times before. Error messages are the obvious one. You paste a traceback or a Windows error code into ChatGPT and it will usually tell you what went wrong and what to do about it. The answer already exists across dozens of Stack Overflow threads and documentation pages. The AI is pulling it together faster than you would find it yourself.
Configuration lookups work the same way. "What port does Postgres use by default?" or "How do I enable SSH on Ubuntu?" are questions with one correct answer. AI handles those without breaking a sweat.
Basic troubleshooting steps fall into the same bucket. "My printer won't connect" has a handful of causes that have not changed in fifteen years. AI can walk you through them just fine. But the moment your problem gets specific to your network, your hardware revision, or your particular software stack, the ground gets shaky.
AI does not know your network. It has no idea what firewall rules you have set, what VLAN your printer sits on, or whether your ISP does something weird with DNS. It cannot look at your setup. So it guesses. And it guesses with absolute confidence, which is the part that gets people into trouble.
Studies from 2025 found that ungrounded language models hallucinate in 15 to 30 percent of support responses, depending on how complex the question is. ChatGPT itself scores under 40 percent accuracy on general question-answering benchmarks. That means on harder questions, it is wrong more often than it is right.
The problems that need a real person are the ones where context matters. Your n8n workflow talks to your CRM, which talks to your email server, which sits behind a reverse proxy you set up two years ago and forgot about. AI cannot trace that chain. It cannot SSH into your machine and read the logs. It will not notice that the real issue is a stale API token expiring silently every 72 hours because nobody set up token rotation.
Hardware is worse. "My server is slow" could be a dying drive, a bad RAM stick, thermal throttling, or a dozen other things that require someone to physically open the case. AI will tell you to clear caches and restart services. Sometimes that helps. Sometimes the actual problem is a fan clogged with cat hair and the CPU is cooking itself at 95 degrees.
If you do not like sending your tech problems to OpenAI or Google (and I get why you wouldn't), you can run a local AI on your own hardware using Ollama. It is free, open source, and runs on Linux, Mac, and Windows.
The catch is hardware. To run a model worth talking to, you need a GPU with at least 8GB of VRAM and 16GB of system RAM. An NVIDIA RTX 3060 runs around $300 for just the card. Smaller models can limp along on CPU only, but expect 5 to 15 tokens per second on a good processor. That translates to waiting ten to twenty seconds for a short paragraph. It is usable, but just barely.
I will absolutely set up a local AI for a client who wants one. But I am honest about it. If you are spending $500 or more on hardware to run an AI that still gives you the same 40 percent accuracy on the hard questions, the math does not add up. A single service call costs less and gets you someone who can actually see what is happening.
The best use of AI is not as a replacement for a technician. It is preparation for one.
When something breaks, ask AI first. Get the vocabulary. Understand the general shape of the problem. If the AI says "this sounds like a DNS resolution issue," now you have something useful to tell the tech when you call. That saves both of you time, and it means you are not paying for the first thirty minutes of basic diagnosis.
I have had clients show up to calls with a ChatGPT transcript of everything they already tried. That is genuinely helpful. It tells me what did not work, which narrows the problem down fast. Even when the AI suggestions were wrong, knowing they were tried and failed gives me real data to work with.
Use AI for research. Call a tech for execution. That split actually works.
I run Sabyla Solutions. I help people with the stuff AI cannot solve because it needs someone who can see the actual setup, trace the actual connections, and test the actual behavior. A lot of my recent calls start with "I already tried ChatGPT." That is fine. I would rather you tried first.
If you are stuck and the AI is not getting you there, reach out. I will look at it, tell you what is going on, and fix it. No subscription. No monthly fee. Just the work.
Pick a tech problem you have been putting off. Something small. A printer that will not scan, a setting you cannot find, a notification that keeps coming back. Open ChatGPT, Gemini, or Claude and describe the problem the way you would describe it to a friend. Follow the steps it gives you. If it works, you just saved yourself a call. If it does not, pay attention to where it fell apart. What was the AI guessing about? What did it assume about your setup that was not true? That awareness, the ability to feel when an AI is confident but wrong, is the actual skill worth building.
Hi, I'm Justin, an IT consultant and open-source advocate. I specialize in building robust self-hosted infrastructure, developing tools and exploring AI-driven workflows. I write about deploying solutions to common issues, the occasional hardware failure, and everything I learn along the way.