Run the Numbers

There are pockets of genuine practitioner insight - the semantic layer/blessed queries architecture, the deterministic vs. probabilistic control environment problem, and the token-usage-as-leading-indicator observation - but they're diluted by extended sponsor reads, generic 'just start' platitudes, and setup padding. The ratio of novel ideas to filler is around 50/50.

“We actually have in GitHub a repo of sort of the top like 30 dashboards at the company and behind those dashboards all the queries that are used to pipe”

“We didn't have to monitor the quality of our ERPS processing because it was deterministic. And in a probabilistic world like investing in things like Monitoring and then also like what that looks like in your broader control environment is where companies then need to spend a bunch of time”

The probabilistic-vs-deterministic control environment argument is a genuinely non-obvious reframe that most finance teams haven't worked through. Most other material - build vs. buy, leadership buy-in, data cleanliness first - covers well-worn ground without meaningful contrarian pressure.

“These models are predicting the next most likely token based, a sequence of tokens and they're doing it in a probabilistic manner, which means you don't get the same outputs from the same inputs. And what that means though is a lot of the traditional thinking around user acceptance, testing, post change management, giving you a significant amount of comfort around how an application is going to function. That logic doesn't hold up quite so well.”

“Building M a semantic layer is actually a driver of increased accuracy in a way that organizations don't think about until they've gone a few use cases in”

All three guests are genuine operators or senior practitioners - an SVP FP&A who helped take two companies public, a VP Finance at a public AI-native company, and a PwC partner with a Facebook ads-ML engineering background - not career podcast guests. Solid caliber, though none are true C-suite principals.

“$4 billion run rate, around 8,000 employees globally.”

“I left to go to Facebook where I led ads, ranking, machine learning, engineering teams”

The episode has real company names, specific stack choices (Pigment, Snowflake, GitHub repo of 30 dashboards), and a concrete daily-P&L build story, but it almost entirely lacks hard outcome metrics - no time saved, no error rates, no before/after cycle times - which caps the score meaningfully.

“We have over 30,000 customers and we have something like 50 plus billable SKUs.”

“I sat down one afternoon and I was like, um, I'm. I'm just going to build it. And so it was me plus Clyde, plus a couple hours in the afternoon and was able to actually pull all the data from across the company and across several different tools”

The host deploys some sharp forcing-function questions - the two-year-head-start challenge, the self-service percentage probe, the 90-day action question - and earns credit for following up on RBAC and token visibility. However, he consistently lets vague ROI claims and generic 'just start' answers pass without pushback, and the panel format prevents real depth.

“With how fast AI is moving? How Much. Does having had a two year head start even matter at this point?”

“What percentage of data queries in the company can be self serve realistically versus you need to have a business partner pull it for you in order for it to be correct.”