People of Packaging Podcast

There are genuine nuggets buried in here - $140B market research spend, $150/run vs. $15k/run economics, 4-8 week to sub-10-minute turnaround, 90%+ correlation to human preferences - but they're heavily diluted by an airport anecdote, a mid-episode podcast ad read, flexo-printing analogies, and Seattle weather banter. The actual demo results are walked through too superficially to deliver real learning.

“$140 billion are spent on this every single year. And I'm sure a good chunk of it on package testing. But it's only reserved for the biggest brands, right? Each packaging study is about 15k if you do it properly.”

“we're trying to model the decision making. We're not trying to replicate what people are going to say because then if you show it anything which uh, that data has not seen, this will fail.”

The security inversion - AI testing is safer than human panels because concepts don't leak to competitors - is a genuinely counterintuitive point. The core premise of AI digital twins for iterative packaging testing is a fresh application in this niche, but the underlying ideas (LLM fine-tuning on domain data, MaxDiff validation) are not novel at a conceptual level, and no truly contrarian or first-principles claims are made.

“We are far more safer than running a packaging of concept test with real humans because that can get leaked here.”

“we've heard of stories where people ran a concept test with real human and got leaked, uh, because some competitor was obviously taking the concept test as part of the panel.”

Ankit Dhawan has legitimately relevant credentials - led product for Amazon's global marketing team launching countries remotely, built early AI voice systems for AWS, seven-plus years in AI before it was mainstream - and is a practitioner who identified a real problem and built a product to solve it. The limitation is that this is fundamentally a startup founder pitching his own SaaS, and the conversation never gets beyond that promotional frame.

“I used to lead product for the global marketing team at Amazon. So through that we used to launch new countries and without ever stepping foot into them.”

“I let AI for AWS essentially Amazon build some of the first artificial voices we built.”

The episode delivers concrete, usable numbers - market size, per-study cost benchmarks, subscription pricing tiers, runs per year, turnaround time, two-week audience build, 90%+ MaxDiff correlation - which is better than most podcast episodes in this space. However, the live demo results are narrated vaguely ('they need to double down,' 'not doing badly') rather than yielding specific, actionable findings, and no case study with real outcome data is presented.

“it starts at $18,000, but the sweet spot most brands are paying us are $27,000, which includes your package testing, concept testing. And they can run 180 per year. So it comes out to be about $150 per run”

“we are fine tuning that specific model for this category. Everything which we're doing, the underlying data is real human behavior in this specific category, which is coming from the qualitative interviews from public social signals, from purchase data”

The host asks a few substantive questions - security and data leakage, the commercial model, how buyer personas are built - but never pushes on the validity of AI-simulated consumer behavior, what the failure modes are, how they prove causation vs. correlation, or who the real competitors are. A mid-interview podcast ad read and a multi-minute India airport story consume meaningful airtime that could have probed the actual claims.

“How do I know that all these iterations are staying confined and aren't leaking out into these large language models”

“If you care about the future of packaging, and I do, you do too. You're listening to a packaging podcast, then you need more.”