The AI Experiments That Failed (And What I Learned)

Nobody talks about their AI failures. Everyone’s too busy posting screenshots of their perfectly formatted outputs and claiming they “10x’d their productivity.”

Let me tell you about the AI projects that crashed and burned. The ones that wasted my time, confused my clients, and taught me more than any success ever could.

If you’re experimenting with AI and feeling like you’re the only one who can’t get it to work, this article is for you.

The Automated Lead Qualifier That Qualified Nothing

The Vision: I was going to build an AI-powered lead qualification system that would score incoming inquiries, categorize them by fit, and automatically draft personalized responses. I’d wake up to a tidy inbox full of pre-written replies ready to send.

What I Built: A Google Form that fed into a Sheet, with a ChatGPT prompt that was supposed to analyze each response and populate scoring columns.

What Actually Happened:

The first problem was that ChatGPT couldn’t directly access my Google Sheet in any reliable way. I tried Zapier, I tried Make.com, I tried writing Apps Script that would pass data to the API. Each integration had its own quirks, rate limits, and mysterious failures.

When I finally got something working, the scoring was wildly inconsistent. One day it would give a consulting inquiry a 3/10. The next day, the exact same type of inquiry would get an 8/10. The AI had no memory, no consistency, no actual understanding of what made a good lead for my business.

The personalized responses? Generic garbage. “Thank you for your interest in our services. We’d love to discuss your needs further.” I could have written that template in 30 seconds without AI.

What I Learned:

  1. AI can’t maintain context across disconnected systems. Each time the API runs, it’s starting fresh. It doesn’t “remember” what a good lead looks like unless you feed it that context every single time.
  2. Simple automation beats complex AI. I eventually ditched the AI scoring entirely and just used basic Google Sheets formulas. If someone ticks certain boxes in the form, they get a high score. It’s boring, it’s deterministic, and it actually works.
  3. Personalization at scale is usually fake. Real personalization requires knowing the person. AI personalization is just mad-libs with slightly different wording. Most recipients can tell.

What I Do Now: I use a simple form with smart conditional logic that routes people to different follow-up paths based on their answers. No AI. No scoring algorithms. Just clear questions and honest triage.

The Content Repurposing Machine That Produced Garbage

The Vision: Take my podcast episodes and automatically turn them into blog posts, LinkedIn posts, email newsletters, and social media snippets. One episode becomes 20 pieces of content without lifting a finger.

What I Built: A workflow where I’d feed ChatGPT the podcast transcript and ask it to create all these different formats in one go.

What Actually Happened:

The blog posts read like they were written by a robot. Because they were. The tone was off—too formal in some places, too casual in others. The flow was choppy. Key points got lost or over-simplified.

The LinkedIn posts were worse. They all had the same structure: provocative question, three bullet points, call to action. After posting three of them, people started asking if I’d been hacked or if I’d outsourced my content to someone overseas.

The email newsletters? They didn’t sound like me at all. Subscribers started unsubscribing.

What I Learned:

  1. Your voice is harder to replicate than you think. AI can mimic surface-level style, but it can’t capture the rhythm, the specific examples, the small asides that make your content sound like you.
  2. One-shot generation is lazy and it shows. Asking AI to do everything in one go means you get AI’s default outputs, not something crafted for your audience.
  3. Repurposing isn’t automation—it’s editorial judgment. You need to decide which parts of a 30-minute podcast are actually worth expanding into a blog post, which quotes make good LinkedIn content, and which ideas need more context to work in writing.

What I Do Now: I still use AI for repurposing, but I’ve completely changed the workflow:

  • I use AI to transcribe and create a rough outline of key themes
  • I personally pick which themes deserve to become content
  • I write the first draft myself or heavily edit what AI produces
  • I use AI for specific tasks like shortening something or suggesting headline variations
  • Every piece gets a final “does this sound like me?” check before publishing

The result takes longer, but people actually read it.

The Meeting Summary Bot That Nobody Read

The Vision: I’d record client meetings, have AI transcribe them, and automatically generate summaries with action items that I’d send to the client. Professional, efficient, impressive.

What I Built: I used Otter.ai for transcription and ChatGPT to create summaries from the transcript.

What Actually Happened:

The summaries were technically accurate but completely useless. They captured what was said, but not what mattered.

AI would note that “Steve mentioned possibly updating the requirements document by next Friday” when what actually happened was the client pushed back hard on that timeline and we agreed to revisit it in two weeks.

It would list every single topic we discussed in chronological order, making a 45-minute meeting look like we covered 15 different things, when really we spent 40 minutes on one core issue.

Clients didn’t read them. I know because I’d reference “the summary I sent” in follow-up emails and get blank responses.

What I Learned:

  1. AI captures words, not meaning. It doesn’t understand emphasis, pushback, body language, or the unsaid context that shapes every business conversation.
  2. Summaries need editorial judgment. What matters in a meeting isn’t what took the most time or what was said most often—it’s what decision got made, what changed, and what happens next.
  3. Your clients want your interpretation, not a transcript summary. They were in the meeting too. They don’t need a recap. They need clarity on what it all meant and what they should do.

What I Do Now: I take notes during the meeting (usually just key points and decisions) and spend 10 minutes after writing a short, human email that says:

  • Here’s what we agreed
  • Here’s what changed from our last conversation
  • Here’s what you’re doing, here’s what I’m doing
  • Here’s when we’re talking next

AI sometimes helps me structure my rough notes, but the thinking and the framing are mine.

The AI Writing Assistant That Made Me Lazy

The Vision: Use ChatGPT to help draft everything—proposals, emails, reports. Write faster, write more, work smarter.

What I Built: Nothing complex. Just me relying heavily on AI to write my first drafts.

What Actually Happened:

My writing got worse. Not because AI writes badly (though sometimes it does), but because I stopped thinking before I wrote.

I’d fire off a half-formed thought to ChatGPT: “Write a proposal for a consulting engagement with a mid-sized tech company looking to improve their business analysis processes.”

ChatGPT would give me three pages of professional-sounding but utterly generic consulting speak. I’d edit it a bit, send it off, and lose the work.

Why? Because the proposal didn’t reflect any actual understanding of the client’s situation. It was a template with their company name swapped in.

What I Learned:

  1. AI can’t substitute for thinking. If you haven’t thought through what you’re trying to say, AI will just give you confident-sounding filler.
  2. Generic content loses to specific content every time. Clients can smell when you’ve used a template. They want to know you understand their actual problem.
  3. The writing process is where you figure out what you think. When I handed that process to AI, I stopped developing my ideas. I stopped getting clearer on my own value proposition.

What I Do Now: I write the messy first draft myself. Just brain dump onto the page. Then I might ask AI to help me restructure it, tighten it up, or check if I’ve explained something clearly. But the thinking comes from me.

The Custom GPT That Nobody Used (Including Me)

The Vision: Build a custom GPT trained on my consulting frameworks, my writing style, and my domain expertise. My personal AI consultant that could help clients even when I wasn’t available.

What I Built: I spent days uploading documents, writing detailed instructions, and testing different configurations. I created “Steve’s Business Analysis Assistant.”

What Actually Happened:

I used it twice. Clients never used it at all.

Why? Because it wasn’t actually better than just using regular ChatGPT with a good prompt. All the “custom training” didn’t make it smarter about my specific domain—it just made it repeat phrases from my documents.

When I tested it with real client questions, the answers were either obvious (things any BA would know) or generic (things regular ChatGPT would say anyway).

What I Learned:

  1. Custom GPTs are over-hyped for most use cases. Unless you have a very specific, repeated workflow with consistent inputs and outputs, you don’t need one.
  2. Instructions matter more than training data. A well-written prompt to regular ChatGPT beat my “custom trained” version almost every time.
  3. People want human expertise, not AI proxies. Clients weren’t interested in talking to my AI assistant. They wanted to talk to me.

What I Do Now: I deleted the custom GPT. I keep a document with my best prompts and frameworks that I can paste into any AI tool when needed. Simpler, faster, more flexible.

The Productivity Tracking System That Tracked Nothing

The Vision: Use AI to analyze my calendar, my email activity, and my task completion to identify productivity patterns and optimize my schedule.

What I Built: I tried several AI-powered productivity tools that promised insights into where my time was going.

What Actually Happened:

The “insights” were useless:

  • “You spend a lot of time in meetings” (No shit, I’m a consultant)
  • “Your most productive time is 9-11am” (That’s when I don’t have meetings scheduled, not when I’m most productive)
  • “You should batch similar tasks together” (Thanks, I read Getting Things Done in 2008)

The tools couldn’t differentiate between valuable work and time-wasting. They just counted hours and categorized activities based on keywords.

What I Learned:

  1. AI can measure activity, not productivity. Productivity is about outcomes, not time spent. AI has no idea if that three-hour writing session produced genius or garbage.
  2. Optimization requires judgment, not just data. Knowing you spend 40% of your week in meetings doesn’t tell you which meetings to cut or how to make them better.
  3. Most productivity problems aren’t data problems. I don’t need AI to tell me I’m procrastinating. I need to fix the underlying issue that’s making me procrastinate.

What I Do Now: I do a weekly review on Friday afternoon. I look at what I accomplished, what I didn’t, and what I want to change next week. It takes 15 minutes. No AI required.

What All These Failures Taught Me

Here’s what I learned from watching my AI experiments crash and burn:

1. AI is a tool, not a solution

Every failed project started with “AI can solve this.” None of them started with “Here’s my problem, and one piece of this might be helped by AI.”

If you can’t clearly articulate the problem and the desired outcome without mentioning AI, you’re not ready to use AI.

2. Complexity is expensive

The more moving parts your AI system has (APIs, integrations, data flows, error handling), the more ways it can fail. Most of my successful AI uses are stupidly simple: single prompt, clear input, immediate output.

3. AI can’t replace judgment

Every time I tried to automate judgment (Is this a good lead? What should I do next? How should I phrase this?), it failed. AI can help inform judgment, but it can’t replace it.

4. Generic AI outputs are worse than no outputs

A template you wrote yourself beats AI-generated content that sounds professional but says nothing. Your clients and audience aren’t stupid. They can tell.

5. The best AI workflows are invisible

My most successful AI uses don’t look like “AI projects.” They look like me working faster. AI helps me research, organize, and articulate—but it doesn’t replace the work of thinking, deciding, and creating.

6. Start manual, then consider automation

Every failed automation started with “wouldn’t it be cool if this was automatic?” Every successful one started with “I’ve done this manually 20 times, now I see the pattern.”

7. Your voice matters more than you think

People don’t want content. They want your perspective, your experience, your judgment. If AI is making you sound like everyone else, you’re using it wrong.

The Experiments I’m Still Running

Not everything failed. Here’s what’s actually working:

  • Research synthesis: I feed AI long documents and ask specific questions. It saves hours of reading.
  • First-draft editing: I write badly, AI helps me structure it better.
  • Example generation: “Give me 5 different ways to explain this concept.” Gold.
  • Format conversion: Turn meeting notes into action items, turn brain dumps into outlines.
  • Brainstorming partner: Talking through ideas with AI helps me think, even when its suggestions are mediocre.

These all have something in common: they’re assistive, not autonomous. I’m still driving. AI is just helping me get there faster.

Your Turn

If you’re experimenting with AI and things aren’t working, that’s not failure. That’s learning.

The people posting their perfect AI workflows on LinkedIn aren’t showing you the 10 failed attempts that came before, or the hours of manual cleanup they’re doing behind the scenes.

Share your failures. Ask questions. Admit when something didn’t work.

That’s how we all get better.

What AI experiments have you tried that didn’t work out? Reply and let me know—I’d love to hear your honest stories, not just the highlight reel.

This article is part of the Zero to AI series, where I document my honest journey from AI beginner to confident practitioner. Subscribe to the podcast for more real talk about working with AI in the mid-career trenches.