I don’t talk about this much. In 2019, I built an AI startup. Raised a small seed round. Hired a team. Burned through the money. Shut it down 18 months later.

Here’s what that expensive education taught me.

The Idea

We were building AI-powered competitive intelligence. Scan the web, analyze competitor moves, deliver insights to sales teams.

On paper, it made sense. Sales teams spend hours researching competitors. AI could do it faster.

We built a working product. Some customers even paid.

Then we died.

What Went Wrong

We built AI that was impressive, not useful

Our demos were amazing. The AI could find and synthesize information in seconds. People would say “wow” a lot.

Then they’d ask: “What do I do with this?”

We’d built a cool technology looking for a problem. The insights were interesting but not actionable. Sales teams didn’t change their behavior based on our output.

Lesson: “Impressive” doesn’t mean “valuable.” Users care about outcomes, not technology.

We underestimated data quality

Our AI was only as good as the data it could find. Turns out, a lot of competitive information isn’t on the public web. And what is public is often outdated or wrong.

Garbage in, garbage out. But we charged premium prices.

Lesson: AI amplifies your data. If your data is bad, AI makes it bad faster.

We hired wrong

We loaded up on ML engineers. Brilliant people. Could build anything.

What we needed was people who understood sales teams. What they actually needed. How they actually worked.

We built what engineers thought sales teams needed. Not what they actually needed.

Lesson: Domain expertise beats technical expertise for AI products.

We burned cash on infrastructure

Custom model training. Expensive cloud compute. A data pipeline that could handle 100x our actual usage.

We spent like we were already successful. The infrastructure was beautiful. The customers were few.

Lesson: Use off-the-shelf AI until you have real scale. Custom infrastructure is a luxury.

We ignored the “last mile” problem

Even when our AI produced good insights, getting them into sales workflows was hard. Integration with CRMs was painful. Training users took forever.

The AI was 20% of the value. The other 80% was getting it into people’s actual work.

Lesson: AI implementation is mostly change management, not technology.

What I’d Do Differently

If I built an AI startup today:

Start with the workflow, not the AI. Understand exactly how people work. Where are the gaps? What decisions are hard? Then ask if AI helps.

Build the dumbest possible version first. GPT-4 didn’t exist in 2019. Today, you can test AI concepts with API calls from OpenAI or Anthropic before writing a single line of custom code.

Charge earlier, build later. We built for 6 months before charging anyone. Should have charged after 6 weeks.

Hire customer experts first. One person who deeply understands your users is worth three engineers.

Stay cheap until you have traction. We spent like we had product-market fit. We didn’t. Budget like it might not work.

How This Shapes My Advice

When founders ask me about AI projects, I think about my failure:

• Is the output actionable, or just impressive?
• Is the data good enough?
• Do they understand the users?
• Are they starting cheap or expensive?
• Have they solved the implementation problem?

Most AI projects fail the same ways mine did. The technology is rarely the problem.

Silver Lining

That failure taught me more than any success. I can spot bad AI projects quickly now. I know what questions to ask.

And I’m not seduced by impressive demos anymore. I ask: “What outcome does this create?”

Sometimes the most expensive education is the most valuable.