I write a lot about AI wins. Time to talk about failures.

These three projects cost me $12,000 combined. None of them work today. All of them taught me something.

Failure 1: AI-Powered Lead Scoring ($4,500)

The idea: Feed our CRM data into an AI model. Predict which leads would convert. Focus sales time on high-probability prospects.

The reality: Garbage predictions.

We hired a freelancer to build it. They connected our HubSpot to GPT-4. Created scoring prompts. Built a nice dashboard.

The model scored our best customers as low probability. It scored tire-kickers as high probability. Worse than random.

Why it failed: Not enough data. We had 800 closed deals. Machine learning needs thousands to find patterns. We were asking AI to learn from noise.

What I’d do differently: Start with rule-based scoring. Manual rules based on observable patterns. AI when we have 10x the data.

Failure 2: Automated Content Repurposing ($3,200)

The idea: Write one blog post. AI automatically creates Twitter threads, LinkedIn posts, email snippets, and podcast scripts.

The reality: Content that nobody engaged with.

Built it with Zapier and GPT-4. The automation worked perfectly. Published 50+ pieces of derived content over two months.

Engagement was 80% lower than our manually created content. The AI output was technically correct but soulless. Generic. Forgettable.

Why it failed: Good content has personality. Specific examples. Strong opinions. AI smooths all that away. What’s left is content-shaped filler.

What I’d do differently: Use AI for first drafts only. Human editing for anything public. The extra time is worth the engagement.

Failure 3: Customer Support Triage ($4,300)

The idea: AI reads support tickets. Categorises them. Routes to the right person. Suggests responses.

The reality: Frustrated customers and confused support staff.

The categorisation was 70% accurate. Sounds okay. In practice, 30% wrong means every third customer gets bounced around.

The suggested responses were helpful for simple questions. But simple questions weren’t the problem. Complex questions needed humans anyway.

Why it failed: The 70% success rate created more work than it saved. Staff had to verify every AI decision. Double handling.

What I’d do differently: AI for search and retrieval, not decisions. Let humans find answers faster, not remove humans from the loop.

The Pattern

All three failures share something: I expected AI to replace judgment.

Lead scoring requires understanding context AI couldn’t see. Content needs voice AI can’t replicate. Support triage needs nuance AI misses.

Where AI works for me: augmenting humans. First drafts. Search. Summarisation. Analysis.

Where AI fails for me: autonomous decisions. Creative work. Anything requiring understanding.

The Money Lost

$12,000 in direct costs. Probably another $5,000 in opportunity cost from time spent.

Worth it? Actually, yes.

I tried things. I learned limits. I know where not to invest now.

The startup founders who’ll waste the most money on AI are the ones who never try. They’ll believe the hype. Chase the wrong projects. Learn nothing.

Fail small. Fail fast. Learn the boundaries.