AI Made Every Product Leader's Wish List Buildable. Now The Bottleneck Is Choosing Where To Start.

For most of software history, the binding constraint on product was execution: teams had more ideas than they could possibly build, so a wish list ten times larger than the available bandwidth was simply the normal state. AI is dissolving that constraint. When the cost and effort of trying something collapses, the ideas that were never worth the effort suddenly are, and the bottleneck moves. It stops being whether a team can build the thing and becomes whether the thing is worth building at all, whether the market can absorb it, whether it drives measurable value, and whether the increasingly autonomous systems doing the work can be secured.

Vaneet Singla is Executive Vice President and Chief Product Officer at Jubilant FoodWorks, the food-tech company that operates Domino’s in India and several other markets alongside Popeyes, Hong’s Kitchen, and a vertically integrated supply chain of its own commissaries. Singla has spent his career building and leading product teams, and he's direct about how fast AI is rewriting the economics of that work.

"There were a lot of use cases that we didn't chase because the effort required was never justified. With AI, that equation is changing," Singla says. Once that equation flips, the shape of product work changes with it.

The bottleneck moves from building to choosing

Singla’s central claim is that AI doesn't just speed teams up. It changes which ideas are rational to pursue at all. When the list of ideas a product team could never get to becomes reachable, the scarce resource becomes something other than engineering capacity. "The bottleneck is going to move to the capacity of the market to absorb what teams can now produce," he says.

In his framing, the volume of viable ideas is about to explode, which makes the real work less about shipping and more about choosing well. He's blunt that the gating question is usually not whether an idea has value. "We don’t have a dearth of use cases," he notes. "The only question is whether the models of today are able to add value to our operations."

Philip Samuelraj, Founder and CEO of the AI development firm Techjays, works the build side of that same shift across dozens of companies, and he pushes the diagnosis one step further. The trap he sees most often is teams pointing all the new capacity AI delivers at the same workflows they already have. "The mistake is taking a workflow you already run and just making it faster with AI. A 10 or 20 percent productivity bump on the same process doesn't meaningfully move a company," Samuelraj says. "The question is not 'How do I speed this up?' It's whether this workflow should exist at all in the agentic era, and what it looks like if you rebuild it from scratch."

The one-person pod

The compression is already reshaping how product work is organized. Singla describes the end of a team structure that has held for thirty years. "Historically, we've operated in this one-PM, one-designer, three-to-four-engineer structure," he says. He expects that to collapse toward a one- or two-person pod, where one person owns the consumer and the business while the other orchestrates models.

Some roles, in his account, are already gone: the UX writer, and soon the standalone analyst, absorbed into the product function or into the tools themselves. He frames this as efficiency rather than replacement. Smaller teams are aimed at far more ground than they could previously cover.

Experiment first, platformize later

Jubilant runs a heavy experimentation program with roughly 150 experiments a year, and Singla is deliberate about not strangling it with premature architecture. The company’s default is to build once and reuse across every brand and geography, but he refuses to impose that discipline at the experiment stage. "For experiments, the team never had to worry about whether we are going to platformize it or not," he shares. Forcing platform requirements onto an idea on day one, he estimates, would slow his teams tenfold.

So, experiments run unburdened. Only once something proves out does it get scaled for the lead brand and then folded into the shared platform. The sequence protects speed where speed matters and pays the reuse cost only once the bet has landed.

Proof over theater

What moves an experiment from interesting to funded is evidence, and Singla describes a progression from subjective to objective. He points to the Pizza Quality Analyzer, a vision model that scores pizzas against six or seven quality parameters, as the template.

The first test was simply accuracy: did the model’s scores line up with what a human grader would give? From there the chain extends to leading and lagging metrics. Better scoring enables training and interception of bad product, which should lift customer satisfaction, which should lift repeat rates. Crucially, that proof is demanded at scale before the capital commitment. "We're not going to go from five stores to 2,500 stores," he says. The path runs through 50, then 100, where accuracy, CSAT, and repeat rates all have to hold, because every deployment carries real capex and ongoing cost. The discipline is that nothing scales on theory.

Singla is candid that the gating question is rarely whether an idea has value, but whether today’s models are good enough to deliver it. PQA failed twice, tried first with machine learning three years ago and again a year later, before a new model generation early this year made it work. Location AI, now a hard gate on opening any new store, took a similar multi-year road. His posture is to keep betting: run the small experiment, watch for the first real signal, and go all-in when it appears.

Samuelraj frames that discipline as the thing that separates real transformation from what he calls governance theater. The danger is letting process substitute for progress. "A lot of companies answer AI by starting a committee and calling it governance. Guardrails matter, but the pace of the committee becomes the pace of its slowest member. In a space moving this fast, by the time they've caught up they're already behind," he notes. "Organizations should be able to point to a workflow and say, 'Here is what we used to do by hand that an agent does now, and here is the measurable impact.'"

Securing systems that act on their own

Singla doesn't pretend all the hard problems are solved. The same autonomy that makes AI valuable is what he has the least control over, and as a listed company holding large volumes of customer data, the stakes are real. "You don’t know how to control AI agents running amok in your code base," he says, noting that an agent is not a human whose access can be governed the usual way, and that AI is known to break access controls. His team hasn't found good answers yet.

Jubilant launched an AI shopping connector for Domino’s India as an access-controlled experiment, on the logic that if the traction is real, it’s worth investing further in enterprise-grade security hardening before a wider rollout, rather than letting proactive innovation getting sidelined by premature optimization.

The posture matches how Samuelraj thinks about the gap between a demo and a system a business can depend on. One tells you whether to invest. The other is what you invest in. "It's easy now for anyone to stand up a demo, and that's a great first step. But a demo isn't secure. It's not something you bet a business on," Samuelraj says. "You have to treat proving the idea and building the production version as two separate jobs, and not confuse them."