How Galgo cut manual photo review from 150 hours a month to under 60 flagged cases

The best AI automation I read about this week does not replace a team. It tells the team where to look.

What happened

Zapier named Ignacio Piñeiro of Galgo its 2026 Operations AI Builder of the Year. Galgo is a fintech that lends money for motorcycles in Mexico, and each loan needs a delivery photo of the customer with the vehicle. That photo is its main defense against delivery fraud.

Agencies submit more than 3,000 photos a month. Manual review took roughly three minutes each, about 150 hours a month, and around 8% of evidence was still invalid.

The new flow: when a delivery is logged in Galgo's app, a trigger looks up the deal in HubSpot and pulls the photo. AI by Zapier running GPT-5-mini checks it against strict rules: a real customer and an eligible vehicle together in a real photo. Stock photos and signed documents fail. Invalid photos go to a second model that classifies the reason. Every verdict is logged in Google Sheets and posted to a Slack channel the fraud team watches.

Results reported by Zapier: invalid evidence fell from about 8% to under 2%, and human review dropped to flagged cases only, fewer than 60 photos a month. Piñeiro estimates at least tens of thousands of dollars in avoided first payment defaults.

My take

This is the pattern I push with clients, and it works far beyond fraud:

• Rules first, model second. The prompt encodes a written policy, not a vibe.
• Small model, narrow job. A cheap model gives a yes or no. A second step explains the no.
• Nothing is a black box. Every decision lands in a sheet and a Slack channel, so humans can audit any call.
• Tune on real failures. Piñeiro adjusted the prompt around real misses, like camera watermarks being read as stock photos.

Swap "delivery photo" for "proof of payment", "signed contract" or "lead form" and you have a QA step for most back offices. The goal, in his words, is not reviewing faster. It is making the judgment instantly and consistently so the team only touches exceptions.

I design and build AI automations that do real jobs like lead routing, reporting, data entry, document checks and agent workflows. See what that looks like at romielwillautomate.dev.