Building AI Products as a Non-Technical PM: What Actually Works

Everyone's building AI features now. Most of them suck. Not because the models are bad—Claude and GPT-4 are incredible—but because PMs treat AI as a feature, not a design constraint.

I'm not an ML engineer. But I've shipped AI features in enterprise B2B—resume parsing, document verification, workflow automation. Here's what actually works when building AI products as a PM who can't train models.

1. Start with the job-to-be-done, not the capability

Bad framing: "We have GPT-4, what can we build?" Good framing: "Users spend 4 hours/week on resume screening. Can AI reduce that?"

At Keka, we didn't build "AI-powered BGV" because it sounded cool. We built it because manual document verification had a 23% error rate and took 3 days on average. AI was the solution to a measured problem, not a feature in search of one.

2. Design for the 80% case, fallback for the 20%

AI products fail when they promise 100% accuracy. Instead, design for high-confidence cases (80%) and graceful degradation for the rest.

Our approach: AI auto-approves documents with >95% confidence, flags 5-15% for human review, rejects obvious failures. Users trust it because we're transparent about confidence levels.

3. The hardest part isn't the model—it's the training data

You can't build good AI without good data. At Multiplier, we spent 2 months just cleaning and labeling data before building ML features. Unsexy? Yes. Essential? Absolutely.

4. User trust is earned through transparency

People don't trust black boxes. Show your work. For resume parsing, we don't just extract data—we highlight the source text. Users can verify AI decisions instantly. Trust went from 40% to 89%.

5. Build feedback loops from day one

The best AI products get better over time. Build correction interfaces—when users fix AI mistakes, capture that as training data. Our document verification accuracy went from 87% to 96% in 6 months purely from user corrections.

The meta-insight: AI is a material, like steel or plastic. Good PMs understand their materials' properties—strength, flexibility, failure modes. AI's "material properties" are probabilistic, data-hungry, and opaque. Design accordingly.