Invisible Machines podcast by UX Magazine

The episode surfaces several non-obvious operational points - most notably the distinction between capacity creation and capacity application as separate measurement challenges, the scaling cliff between 15 and 15,000 agents, and the evaluation loop methodology - but loses significant density to prolonged donut analogies, the farm anecdote, and circular philosophical meanderings that eat substantial runtime.

“you not only have to know where the capacity came from, you have to know where that capacity was applied and whether or not that application of that capacity actually was driving greater revenue or greater customer service or whatever”

“if you're building 5 or 10 or 15, sure, you can brute force it, you can fake it, but what happens when you've got 150 agents or 1500 agents or 15,000 agents”

A handful of genuinely fresh framings appear - 'use case zero,' the incremental seeding of human judgment, and the inversion to 'agent in the loop' - but the bulk of the advice (data foundation first, evaluation matters, human oversight needed, AI can make you dumb or smart) is well-worn enterprise AI commentary that circulates widely.

“it's not going to be like one moment in time. We're going to start to seed our judgment over time and it's going to happen in very small increments”

“I call it Use case zero. Because I always feel like if you're going to build an optimus Robot, that use case 0 is that Optimus can build robots, not fold laundry”

Jeff McMillan is a genuine practitioner who built firm-wide AI infrastructure at Morgan Stanley at meaningful scale, now consults and teaches at Columbia - not a thought-leader or career podcaster, and the transcript reflects real operational scar tissue rather than theory.

“first on Wall street as head of firm wide AI at Morgan Stanley. Now through macmillan AI”

“my first project on Wall street, um, was to build out a customer information database. I was in my early 30s, uh, and knew nothing about data. And what's interesting is that problem was critical over 25 years ago”

The episode offers concrete evaluation percentages, a named testing cadence (100 users, 20 questions per week), and the 15,000-agent / 99%+ data quality thresholds, which is more specific than average; however, no named company outcomes, dollar figures, or Morgan Stanley case data are shared, and most examples remain illustrative or hypothetical.

“maybe the first time you run it through the model it's 80% accurate. And then it gets 85 and 90, 93, 94. And then what happens sometimes is...it goes from 89% to 74%”

“I need you to ask 20 questions of the system. And I need 100 people to do that every week on Monday”

The hosts occasionally generate useful friction (the capacity-measurement question, the knowledge management follow-up) but habitually over-talk, complete the guest's sentences, and let analogies run far past their useful life; there is no meaningful pushback or challenge to any of McMillan's claims throughout the 53 minutes.

“Does knowledge management kind of get you there in a way though, if you do it the right way?”

“But companies don't like knowledge management. It's boring. They have to fund something that they don't understand. It's hard to roi, analyze it. Um, is there a way around it?”