What Building AI Photography at OYO Taught Me About Useful AI

5 MINS

The most useful AI feature I've ever shipped was also the least flashy: an AI-powered photography solution to automate property go-lives at OYO. It didn't win any design awards. It didn't get a launch tweet. But it cut the time-to-go-live from days to hours and unblocked a real bottleneck across the network.

That feature is the lens I now use for every AI conversation. Useful AI removes a constraint. Everything else is a demo.

The constraint nobody saw

Before the AI work, here's how a property went live at OYO: a photographer travelled to the location, shot the rooms, the photos went through manual review, edits got requested, the photographer went back, and so on. Multi-day cycle, expensive, fragile.

The constraint wasn't the photographer's skill. It was the gap between the photographer's intent and what the system could verify automatically. AI didn't replace the photographer. It just collapsed that gap. Pictures got auto-checked for the things humans were checking inconsistently anyway. The photographer got real-time feedback. The reviewer got a curated batch instead of a manual queue.

That's the pattern. Find the place where humans are doing repetitive verification at scale. That's where AI earns its keep.

Three rules I now apply to any AI feature

After living through that build, I now run every AI feature idea through three filters:

Is the human currently doing this manually, repeatedly, and at scale? If no, you're inventing a problem.

Is the failure mode visible and reversible? If the AI gets it wrong, can the user immediately tell, and can they fix it?

Does the feature reduce a clear time, cost, or quality cost? If you can't put a number on it, you don't have a feature — you have a thesis. Most AI feature ideas in product reviews fail at least one of these. That's fine. Most of them shouldn't be built.

Why "AI for the user" is usually less interesting than "AI for the operator"

Consumer products love to brag about AI features for the *end user* — recommendations, summaries, chat. Some of those work. Many are decoration.

The bigger wins I've seen in my domain are AI features for the operator — the ops manager, the customer-care lead, the merchant onboarder. Those people repeat tasks at high volume, are paid for their time, and have direct accountability for outcomes. AI lands on them harder and more measurably.

If you're building a consumer product right now and you're only thinking about AI on the user side, flip the camera around. The operator side might be where your next real win lives.

The Swiggy and Zomato lessons that still apply

I worked on speedy-reply voice recognition at Swiggy, and on order-relay automation at Zomato. The technology in both cases was nowhere near today's foundation models. But the discipline was the same:

Find the bottleneck the human is bored of.

Reduce its cycle time by 80%.

Leave the human in the loop for the 5% that matters. That's the boring version of AI product. It's also the version that actually ships and stays shipped.

The next thing I want to build

In my current role at Auctane, the same pattern shows up in shipping logistics. There are humans verifying, classifying, routing, escalating. Each of those touchpoints is a candidate for the same playbook.

I'm not chasing the next demo. I'm chasing the next bottleneck. AI is just the tool that finally fits the shape of the problem.

Operations Discipline Is What Separates Real Products from Demos

MVP to MLP: The Real Distance Most Founders Underestimate

Hrishi skipped presentations and built real AI products.

Hrishi Swaroop Bhatnagar was part of the March 2026 cohort at Curious PM, alongside 17 other talented participants.