The AI Project I Botched and What It Cost Me

<i>I once sold an AI voice agent to a Maitland plumbing company that failed spectacularly—costing me $4,500 and the client 60 missed calls a day. Here’s the ugly truth about what I learned.</i>

I remember the exact moment I knew I’d screwed up. It was a Tuesday afternoon in August, and I was sitting in a conference room at a plumbing company in Maitland, Florida. The owner, a guy named Rick who’d built his business from a single truck to a fleet of twelve, was staring at his phone. His face was red. “Sixty missed calls yesterday,” he said. “Sixty. My receptionist quit last week because customers were yelling at her about the AI that didn’t work.”

That AI was my fault. I’d sold him on a voice agent that would handle after-hours calls, book appointments, and send estimates. I’d promised it’d save him $4,500 a month in overtime pay. Instead, it cost him customers, and it cost me a reputation I’d spent years building. This is the story of that project—what I did wrong, what it cost us both, and the unglamorous lessons I now carry into every engagement in Central Florida.

The Pitch That Felt Too Good

Rick came to me through a referral. His business was growing fast—new construction, repipes, emergency clogs—and his two-person office was drowning. He needed a way to handle calls after 5 p.m. and on weekends without hiring a third person. I’d just read about a new voice AI platform that could understand natural language, book appointments, and even answer technical questions like “How much does it cost to snake a main line?” I was excited. I demoed it for Rick, and he was sold in twenty minutes.

I quoted him $3,000 for setup and $500 a month for the software. He said yes on the spot. I felt like a genius. I’d actually helped a local business owner solve a real problem with AI—no buzzwords, no hype, just a tool that worked. Except it didn’t. Not for his business, anyway.

The First Signs of Trouble

The voice agent went live on a Monday. By Wednesday, Rick’s receptionist, Maria, was calling me in a panic. “The AI keeps asking callers to repeat their address three times,” she said. “And when someone says ’emergency,’ it says, ‘I’m sorry, I didn’t catch that.’” I logged into the dashboard and saw the transcripts. The AI was failing on the exact use cases Rick needed most: urgent calls and customers with heavy accents.

I’d tested the AI with my own voice—clear, standard American English, no background noise. But Rick’s customers were a mix: Spanish speakers, Haitian Creole speakers, folks calling from construction sites with jackhammers in the background. The AI couldn’t handle it. It’d hang up on people or, worse, book appointments for the wrong day. One customer showed up on a Saturday for a job that was scheduled for Monday. Rick had to pay a weekend call-out fee to his plumber to make it right.

I tried to fix it. Spent three weeks tweaking the model, adding custom phrases, adjusting the confidence thresholds. Nothing worked. The AI was brittle. Couldn’t handle the chaos of a real plumbing business.

The Real Cost

Let me be honest about the numbers. Rick paid me $3,000 for setup and two months of software fees ($1,000). I refunded the software fees and ate the setup cost—so I lost $3,000 out of pocket. But that was the small part. The bigger hit? Those 60 missed calls per day. Rick’s average job was $350. If even 10% of those missed calls turned into jobs, that’s $2,100 a day in lost revenue. Over two months, that’s over $80,000. He also lost Maria, his receptionist, who quit because she was tired of apologizing to angry customers. Hiring and training a replacement cost him another $2,000.

For me, the cost was worse. I lost a client and got a reputation as the guy who sells broken AI. Two other referrals dried up. I spent months rebuilding trust with local business owners. I learned that selling AI without understanding the messy reality of a small business isn’t just a mistake—it’s a betrayal.

“I learned that selling AI without understanding the messy reality of a small business isn’t just a mistake—it’s a betrayal.”

What I Should Have Done Differently

Looking back, I made three specific mistakes that I now avoid like the plague. First, I didn’t do a proper AI readiness assessment. I assumed that because the voice agent worked in a demo, it’d work in the wild. Didn’t check Rick’s call data: call volume, peak hours, common questions, accent diversity, background noise. I didn’t ask to record a week of calls and analyze them. If I had, I’d have seen that 40% of his callers spoke Spanish as a first language, and another 15% had heavy Southern accents that the AI couldn’t parse.

Second, I rushed the implementation. Set up the AI in a weekend and pushed it live on Monday. Didn’t run a parallel test where the AI handled calls while Maria listened in. Didn’t have a fallback plan for when things went wrong. I should’ve run a two-week pilot with a human backup, then gradually increased the AI’s responsibility. That’s what I do now with every voice agent implementation in Central Florida: start small, measure everything, and have a human escape hatch.

Third, I overpromised. I told Rick the AI would “save him $4,500 a month.” That was a guess, not a calculation. Didn’t account for the cost of errors, the time needed to train the AI, or the fact that some customers just hate talking to machines. Now I’m brutally conservative with projections. I tell clients: “This tool might save you 10 hours a week, but it might also annoy 5% of your callers. Let’s test it and see.”

The Unglamorous Lessons That Stick

That botched project taught me more than any successful one. Here are the lessons I now apply to every engagement, whether it’s a Microsoft 365 Copilot rollout in Lake Mary or a custom chatbot for a Winter Park law firm.

Lesson 1: Start with the messiest customer. Don’t test your AI with the perfect user. Test it with the person who mumbles, the one who calls from a noisy truck, the one who switches between English and Spanish mid-sentence. If it works for them, it’ll work for everyone.

Lesson 2: Measure the cost of failure. Before you deploy any AI, ask the client: “What happens if this breaks? How much does a missed call cost? How much does a wrong appointment cost?” Then build a safety net. For Rick, a simple fallback to a human answering service would’ve cost $200 a month but saved him thousands.

Lesson 3: Be the bad news bearer. I should’ve told Rick upfront: “This AI isn’t ready for your full call volume. It’ll make mistakes. We need to start small and fix it together.” Instead, I sold him a fantasy. Now I lead with limitations. I say things like, “This tool’s good at booking appointments but terrible at handling angry customers. Here’s how we’ll manage that.”

Lesson 4: AI is a tool, not a replacement. Rick wanted to replace his receptionist. But his receptionist did more than answer phones—she smoothed over upset customers, remembered names, and knew which plumber was best at unclogging a specific type of toilet. AI can’t do that. It can only handle the routine stuff. The best AI projects augment people, not replace them.

How I Do It Now

These days, when a Central Florida business owner asks me about AI, I start with a conversation, not a demo. I ask about their biggest frustrations, their worst customer interactions, their most repetitive tasks. Then I suggest a small, low-risk experiment. For example, a property management company in Clermont wanted an AI to handle maintenance requests. Instead of building a full system, we started with a simple chatbot that only handled one type of request: “My AC is broken.” We tested it for two weeks, measured how many times it got the address wrong, and then expanded. It worked. They saved 12 hours a week, and no one got angry.

For a medical practice in Oviedo, we used a fractional AI officer approach—I spent two hours a week with their office manager, teaching her how to use AI tools for scheduling and billing. No big software purchase, no fancy demo. Just practical, hands-on help. That project cost them $1,000 a month and saved them $3,000 in overtime. No drama, no failure.

I also make sure every client understands the basics. I send them to the AI glossary on my site so they know the difference between a language model and a voice agent. I want them to be informed buyers, not dazzled marks.

The Hardest Truth

Here’s what I still wrestle with: I’m not sure AI is right for every business. Sometimes the best solution is a second receptionist, a better phone system, or just training your staff better. I’ve learned to say no to projects that don’t fit. I turned down a landscaping company in Apopka because their calls were too chaotic—different languages, different job types, constant changes. They needed a human dispatcher, not a bot. I referred them to a staffing agency. That hurt my wallet, but it saved my reputation.

The Maitland plumbing project still stings. I think about Rick sometimes. Last I heard, he hired a new receptionist and added a call-back feature. He’s doing fine without AI. And I’m doing fine because I learned to be honest about what this technology can and cannot do.

If you’re a business owner in Central Florida thinking about AI, I’ll tell you the same thing I tell everyone: It’s not magic. It’s a tool that can save you time and money if you use it right. But it can also cost you customers and credibility if you rush it. Start small, measure everything, and have a backup plan. And if you want to talk about whether AI actually makes sense for your business, reach out. I’ll give you an honest answer, even if it’s not the one you want to hear.

"I learned that selling AI without understanding the messy reality of a small business is not just a mistake—it's a betrayal."

Frequently asked questions

What was the biggest mistake in the botched AI project?

The biggest mistake was not doing a proper AI readiness assessment before implementation. I didn't analyze call data, accent diversity, or background noise, so the AI failed on the exact use cases the client needed most.

How much did the failed AI project cost?

I lost $3,000 out of pocket from refunds and setup costs. The client lost an estimated $80,000 in missed revenue over two months, plus the cost of hiring a new receptionist.

What is an AI readiness assessment?

It's a process where I analyze a business's data—call volumes, customer demographics, common issues—to determine whether AI can actually help and what risks exist. It prevents the kind of failure I experienced.

How do you test AI before full deployment now?

I run a two-week pilot with a human backup, measure error rates, and only expand if the AI performs well. I also test with the messiest customers first.

Is AI always the right solution for small businesses?

No. Sometimes a second receptionist or better training is more effective. I've learned to say no to projects that don't fit and refer clients to other solutions.

What should a business owner ask before buying AI?

Ask: 'What happens if this breaks? How much does a mistake cost? Can we start small and test it?' If the vendor can't answer those clearly, walk away.