AI, discovery, and a big pivot

Leadership wanted an AI model to identify fence damage from a photo and auto-generate a quote. I was skeptical of the full automation, but I didn't say no. I learned how to train the model and we actually ran a successful POC. We got the model to a point where it could accurately identify wood types and basic damage.

But I knew that to actually ship a product, we had to understand the field work better. I started a discovery sprint to map out the risks and assumptions.

Overview of my actions

1. Grow my understanding of how to train specialized AI models
2. Plan a discovery sprint to understand the human problem
3. Share the risks identified with the initial solution
4. Lead a major product pivot

Discovery

We went into the field to shadow adjusters. We learned two key lessons:

• The "Car Office" problem: Adjusters were fast at the physical inspection. They’d walk a yard, take notes on a notepad, and be done in five minutes. The real time was spent afterward: they’d sit in their cars for 20 minutes re-typing their notes into a laptop. They were doing the work twice.
• Claims were more complex than assumed: A real adjustment needs more than just wood type and a photo. It needs ownership, age, and stain type. The AI couldn't see that, and the manual effort to verify those details killed any time the model saved.

The Pivot

I led a small team to pivot. We kept the tech, but shifted the focus to fixing the workflow. We built a mobile app for the field that used a "photo and tag" system. Instead of a notepad, adjusters could tap through notes and categorize damage while walking.

We built in logic to catch follow-up questions on the spot, so they didn't have to walk back to a customer's door because they forgot a detail.