Is AI a Threat or an Opportunity?

A small experience story! Back in 2021-‘22, when I was building Sparta Maps - our ground truth data collection platform - I remember the long conversations and brainstorming for preventing fraudulent activities during field data collection. How do you ensure the field agents actually visit the farm? How do you know how to eliminate marking useless polygons/points? We designed safeguards, we debated thresholds, we shipped what we could.

But a product is never truly complete. Tech evolves. People evolve. And so do the ways people game the system.

Fast forward to early 2026. We had 18,000+ polygon markings from two vendors across India, and something felt off. My data team complained about the data quality. So I did what any PM should do - I went back to first principles.

What does genuine field data collection actually look like?

A person walks to a farm. They stand near the edge. They draw the polygon. They fill the form. They take a photo and submit. It takes time. It takes movement.

So I worked with AI to build a 5-layer fraud detection pipeline:

  1. Is the agent’s GPS more than ‘X’ m from the polygon?
  2. Are multiple distant farms being marked from the same fixed location?
  3. Is the implied travel speed between submissions physically possible?
  4. Is the GPS placed deep inside the polygon rather than at the edge?
  5. Are 5+ submissions happening within 5 minutes - burst clustering?

The results were stark !

Vendor 1: 96.5% of 10,984 markings flagged. Travel speeds of 6.6 million km/h detected. One agent submitted 152 markings in 108 minutes across two districts - 43 seconds per farm, including travel.

Vendor 2: 69.6% of 7,849 markings flagged. More nuanced - some agents showed clustering patterns consistent with genuinely neighbouring farms, not fabrication. That distinction matters. .

We also ran a perceptual hash (pHash) comparison across field photos - checking whether the same image was reused across supposedly different farm visits. No confirmed reuse in the sample, but suspicious similarity patterns worth investigating further.

Here’s the part I want to be honest about:

AI didn’t solve this. The contextual knowledge did! With AI as my analytical engine, fastened the process and increased the speed.

The 5 conditions, in addition to regular checks, were not an AI model’s findings, the domain knowledge drafted it. Knowing that Vendor 2’s burst flags might reflect clustered neighbouring farms rather than fabrication? That’s not something a model figures out. That comes from understanding how smallholder farming works. Knowing that a field agent stands at the edge, not the centre, of their field? That’s an experience of knowing how the field agent behaves when they are paid per point and also understanding the pattern of marking. Knowing which flags to design in the first place? That’s product thinking.

What AI gave me was speed and scale. What I gave the process was context.

For every PM out there - your irreplaceable value isn’t in doing the work faster. It’s in knowing which work matters, and why. That contextual awareness is the moat. Build it. Borrowing my colleagues words if you donot grab the oppurtunity AI can be your biggest threat!