The AI Edge Podcast

The episode has scattered genuine insights - particularly around audio latency for agentic ads, the brand-guidelines adherence problem, and the composable vs. storing CDP distinction - but large stretches are occupied by news recaps and vague commentary that produce little actionable learning for a B2B operator.

“AI is still really bad at staying on brand. So the number one problem we hear from CMOS is I tried AI in 2025, I'm trying it again in 2026, but no matter which foundation model I use, it's not able to adhere to my brand guidelines”

“the reason I don't believe yet that AI should be doing media buying is because you can't yell at your AI for doing it wrong”

A handful of reframings are genuinely fresh - the 'throat to choke' accountability argument against autonomous media buying, and positioning Databricks as a non-competitor that validates Hightouch's existence - but most of the episode recycles familiar AI-hype-versus-reality narratives and standard composable-data-stack talking points.

“databricks customer lake is a massive threat to CDPs that store data. Because those CDPs should not have existed in the first place”

“you can't yell at your AI for doing it wrong... you could turn your AI, but you also can't fire your AI”

Kashish is a genuine co-founder/operator of a $2.75B-valuation company with named enterprise customers at scale, giving him real practitioner credibility; however, much of the episode functions as a product briefing rather than extracting his deepest hard-won knowledge.

“co founder and co CEO of a company called hitouch. And what we've built is the agentic marketing platform. We work with some of the largest B2C brands, folks like Domino's Pizza, Spotify”

“we've known about databricks building this for about a year”

The episode earns credit for named customers, competitor names, a valuation figure, and the unsourced but concrete 60/40 brand-guidelines stat, but many claims - timelines, accuracy figures, and deal terms - are hand-waved or explicitly deflected, leaving the evidence base thin.

“60% of people don't have their brand guidelines. The 40% of people that do have brand guidelines written. Still, AI is not so good at instruction following and determinism”

“we have to ingest every one of your emails in the past, tag those, look at your CMS and your DAM and build a vector database on top of that”

The hosts occasionally probe interesting distinctions - infrastructure vs. activation, scope boundaries, the Databricks competitive angle - but they largely allow the guest to stay in product-pitch mode, explicitly acknowledge missed questions due to time, and generate no productive disagreement or challenge to any claim.

“I have so many more questions about this potential deal. Um, but we don't have time for all of them”

“Kashish, I wanted to dig into this difference between say an infrastructure layer versus activation, um, layer”