I want to be upfront about something before we get into this.
I'm genuinely excited about AI. I build with it every day. I've watched it go from a novelty to something that's actually changing how businesses operate. The pace of that change is faster than most people realize. Six months from now, some of what I write here will probably be out of date.
But I've also been building in this space long enough to hit real walls. Mid-project, I've recognized that the tool I planned to use wasn't reliable enough to get us where we needed to go — and had to find a different path. Not a different destination. The deliverable didn't change. The client still got what they came for. The route changed, sometimes significantly, because what AI could do in a controlled environment wasn't what it could do under real conditions with real data and real edge cases.
I don't call that failure. I call that knowing when to change roads without taking your eye off the exit.
The experience taught me a lot. Especially the six months I spent fighting ChatGPT through what I can only describe as a circular error problem. Fix one issue, a new one appears. Fix that, a third shows up. Fix the third and you're somehow back to the first — usually at the worst possible moment, the final step of a build. It was maddening. It was also one of the most useful things I've ever worked through, because it forced me to really understand where the gaps were. Working through it is what made me good at building around it.
That experience is worth something. So here's an honest look at where AI genuinely falls short right now, from someone who's been in the middle of it.
It Makes Things Up. Confidently.
This is the one that bites people the hardest because it's so counterintuitive. AI doesn't say "I'm not sure." It answers. It answers in complete sentences, with a tone of authority, whether it's right or not.
The technical term is hallucination. The practical version is: AI will sometimes give you a wrong answer that sounds completely correct, and you won't know it's wrong unless you already know the right answer.
For creative work, writing, summarizing, brainstorming: manageable. For anything that requires factual precision (specific numbers, legal language, technical specifications, client-specific details): this is a real problem that needs a human checking the output before it goes anywhere.
I've had to pull AI out of workflows specifically because the error rate on specific factual outputs was too high to trust without a review step. Which means you've traded one manual task for a different one. That's not always a win.
It Can Get Stuck in a Loop It Can't See
This one I lived for about six months and I haven't seen it written about honestly enough.
When you're working with AI on something complex — a build, a long document, a multi-step process — you can hit a point where fixing one problem creates another. Fix that one, a third appears. Fix the third and you're back to the first. Circular. Every time you push on one side of the balloon, it bulges somewhere else.
ChatGPT was especially bad at this, and especially bad at the final step. You'd be 95% of the way through a build and suddenly couldn't close it out because every correction undid something earlier. The model would optimize for your most recent feedback and quietly lose track of constraints from three rounds ago. It wasn't being difficult. It genuinely couldn't hold everything in view at once.
This has gotten meaningfully better. Context windows are larger, the models are smarter, and there are better ways to structure complex work so you're not asking a single conversation to hold everything. But it still happens. When it does, the move isn't to keep pushing through the same loop. Sometimes you step back, restructure the approach, and come at it from a different angle.
Knowing when to do that — instead of grinding the same loop for hours — is something you only learn by getting stuck in it first.
It Doesn't Know Your Business. Every Session Starts Fresh.
When you open a conversation with Claude or ChatGPT, it doesn't remember last time. It doesn't know your clients, your processes, your preferences, or anything you've told it before — unless you're using a system that's been specifically built to carry that context forward.
This is a bigger limitation than it sounds. A new employee figures out how you work over time. They stop asking the same questions. They pick up your preferences without you restating them. AI, by default, starts completely blank every single time.
There are ways to work around this. Memory systems, persistent context, RAG (Retrieval-Augmented Generation, where the AI pulls from a knowledge base you've built). These are real solutions. But they have to be built intentionally. They don't come free with the tool.
I've had clients describe situations where they'd been using AI for months and still having to re-explain their business from scratch every time. That's not a workflow improvement. That's just a fancier way to do the same work.
It Can't Fix a Broken Process. It Amplifies One.
This is the one I find myself saying in almost every intake conversation.
If your process is unclear, inconsistent, or based on tribal knowledge that lives in someone's head, AI will not clean that up. It will reflect it back to you, faster, at higher volume.
Think of it like a photocopier. A photocopier doesn't fix a crooked original. It makes a hundred crooked copies.
I've had to pause projects and say "before we can automate this, we need to document what this process actually is. Because right now three people on your team are doing it three different ways." That's not an AI problem. That's a business readiness problem. But AI makes it visible immediately, which can feel like AI is the issue when it isn't.
It Can't Verify Its Own Work
This one is subtle but important.
You can ask AI to check something it just wrote. It will tell you it looks correct. It's not actually running independent verification. It's generating a response that sounds like verification. The same reasoning process that produced the original output is reviewing the original output.
This is why human review steps aren't optional in high-stakes workflows. Not because AI is unreliable across the board, but because it genuinely cannot catch its own errors the way a second human can.
It Can't Read the Room on Relationships
AI can draft a professional email. It can hit the right tone for most situations. Where it falls apart is nuance. The history between two people, the thing that went sideways six months ago that nobody wants to bring up directly, the fact that this particular client responds better to casual language even in formal situations.
That context doesn't live in your CRM. It lives in the head of whoever manages the relationship. And right now, AI can't access that.
I've seen AI-drafted client communications that were technically correct and completely wrong for the situation. Right words, wrong read of the room. Sometimes that matters. Sometimes it really matters.
It Can't Maintain Itself
I've mentioned this in other posts but it bears repeating here because people consistently underestimate it.
An AI system (an agent, a chatbot, an automation with AI in the loop) needs ongoing attention. The model updates. The tools it connects to change. Your business changes. What was a good prompt six months ago may no longer reflect how you operate.
This isn't like installing software and walking away. It's more like a garden. You plant it, you tend it, you adjust when things grow in unexpected directions. The businesses that get long-term value from AI are the ones that plan for that ongoing relationship, not the ones that treat it as a one-time installation.
Where This Leaves Us
None of this is a reason not to use AI. I use it constantly. My clients use it. The ROI is real.
But the ROI is clearest when the expectations are accurate. AI is extraordinarily good at first drafts, research, summarizing, pattern recognition, handling volume across routine tasks, and giving you a starting point you'd have spent hours getting to on your own. It's a genuine force multiplier for the right work.
It is not a replacement for judgment. It is not self-sufficient. It is not infallible. And it is not finished. Which is honestly the most exciting part.
The capabilities I've had to work around this year, the project pivots, the "we'll come back to that when the technology catches up" conversations. A lot of those are already being resolved. Things that weren't reliable enough to put in production eight months ago are working now. Things that are still rough today will be solid by the time you read this.
We are building in a moving landscape. The job is to know where you're standing on it.
The Honest Takeaway
Use AI for what it's genuinely good at today. Build human review into anything where errors have real consequences. Don't automate a broken process and expect AI to fix it. Plan for the ongoing maintenance. And check back in every few months. Because what's true today genuinely may not be true by next quarter.
If you're not sure where AI fits into your business right now given all of this? That's the exact conversation the Automation Audit is designed for. We look at your actual operation and tell you honestly what AI can do for you today, what to wait on, and what's coming that's worth planning for.
Michelle Onizuka is co-founder and Systems Architect at Onizuka Studio. She builds AI and automation systems for small and mid-size businesses, and has the rebuilt projects to prove she's been in this long enough to know where the walls are.
[Book an Automation Audit](/automation-audit/) to get an honest read on where AI fits your business right now.