There’s a particular tiredness that shows up in AI programs right about the time the demo stops being cute.
You’ve got a stack that was supposed to make you faster—Foundry, Salesforce, a coding agent, a process bot with a confident name. You can feel the almost. The workflow runs until it doesn’t. The dependency graph looks fine until someone asks for a change. The pilot is "successful" because the team quietly removed the hard part and presented the rest under better lighting. You’re not failing because you lack enthusiasm. You’re exhausted because you’ve been asking a different kind of intelligence to impersonate yesterday’s process, inside systems that still assume a human will absorb every contradiction.
On Invisible Machines, Evan J. Schwartz (Chief Innovation Officer at AMCS Group, and an adjunct building a course he calls AI stewardship) gave that exhaustion a cleaner diagnosis than "we need better prompts."
His split is simple: humans still own the outcome. AI is increasingly in charge of the output. Confuse those two, and you will automate the wrong thing with great confidence.
Schwartz arrives at the point from three directions at once. He lives the vendor-side reality of enterprise software. He sits with customers trying to make AI pay rent. And he teaches the next cohort of builders who will graduate into companies that still haven’t decided whether AI is a sticker on the logo or a redesign of the work.
What he’s teaching is not "become a prompt engineer." It’s a widening at the top of the skill set—an inverse pyramid. You’re a little architect, a little product owner, a little team lead. You stop proving value by personally completing the vertical of doing, and you start proving value by directing: what are the actors, which areas change together, where must dependencies point, what policy constrains the machine before anyone mentions the feature.
Skip that conversation and the model will still deliver. That’s the trap. You’ll get code that satisfies the ask and resists every later change—defensively viable, architecturally brittle, increasingly expensive in tokens as the product grows. Schwartz’s classroom claim is almost the opposite of the vibe-coding romance: first- and second-year developers can accelerate into architectural judgment if you put structure in the box and make them feel the cost of change through tests and maintenance effort.
Robb Wilson presses the organizational rhyme. Teams arrive with fifty-five use cases and an unspoken preference: automate what we already do, exactly as we do it. That feels conservative. It often makes the work harder, because you’re forcing a non-human intelligence to mime a human procedure that was never designed for it. The upside of AI isn’t infinite clerks copying your current motion. It’s the chance to redesign the motion—then staff it with an army that doesn’t sleep and doesn’t need the meeting.
Schwartz’s internet parallel lands because it’s embarrassingly familiar. Early on, companies threw "internet" across the business like a blanket and kept the rubber stamps. Today nobody can imagine operating without the network. AI is in the blanket phase. Lathering it on top of the filing cabinet is not transformation. It’s cosplay with an invoice.
His practical answer is a two-step path that doesn’t pretend humans are tidy.
Step one: don’t fully automate the process as-is. Delegate necessary but low-value work to AI and compress high-value judgment into people. You get a lift that can pay for the tools, calm the board, and build confidence. Humans are messy, lazy, and greedy in the ordinary ways; without a quick win, resistance wins.
Step two: forget the inherited process. Ask how you would have done the whole cycle if AI had been present from the beginning. That’s where the multiples show up. One AMCS customer story makes the asymmetry concrete: bids come in, run the gauntlet, return a ready quote, put a truck on the road that afternoon—while the competitor is still trying to align calendars for the RFP meeting.
Between those steps sits the culture problem Schwartz refuses to soft-pedal. In his stewardship model there are risk-averse, pragmatic, innovator, and early-adopter cultures. You cannot run an innovator adoption program inside a risk-averse company and expect it to behave. Most organizations change when pain arrives—the lost bid, the competitor’s same-day truck, the moment the "safe" delay becomes existential. "There is no change without pain" isn’t a motivational poster. It’s a scheduling note.
Large enterprises add a structural failure mode of their own. Optimize one segment hard enough and you overrun the upstream tissue that feeds it, or drain the data source so fast you create a new bottleneck wearing a success costume. Schwartz’s phrase for the correction: they have to learn to eat themselves from the outside in—customer face first—before they congratulate a local efficiency that broke the organism.
Near the end of the conversation, the human stakes get sharper than the architecture talk.
The people best suited to become AI stewards are often a company’s best producers. They already know how to make the thing. The problem is identity: their sense of self is welded to the output they’re being asked to delegate. Schwartz’s rough cut is brutal and useful—about thirty percent will refuse, not because they can’t make the leap to outcomes, but because they won’t. Education systems that trained cogs should not be shocked when cogs decline the promotion to director.
Robb adds the Dreamliner lesson from another life: essential and non-essential systems on the 787 were deliberately separated. Mix them—treat a cute email draft like a flight-control problem, or drop a mission-critical workload into the same loose apparatus as the non-critical wins—and you don’t get a learning organization. You get a scrap pile and a story that "AI doesn’t work." Schwartz rhymes it back to stewardship: vision without architecture is "take the hill, some of you may die." Structure first is how you keep critical systems from depending on toys, and how you keep a human pedal available when the fancy orchestration goes dark.
His advice to companies about headcount is the line that should travel. If your first move is to cut people after you’ve invested in training them to command AI, you’re not being efficient—you’re donating asymmetric capability to your competitor. Grow the empire with the same headcount. Rightsize last, and only when you’ve actually hit the wall of what the combined human-plus-AI system can become. And say that out loud in a lived vision, not a laminated statement. Break the contract once and people will—as Josh notes—automate quietly for themselves, or refuse to participate in their own extinction.
The wall in Foundry, in Salesforce, in the coding agent that almost ships, is rarely a model-quality problem in the way slide decks pretend. It’s a stewardship problem: outcome confused with output, feature requested before structure, automation confused with reduction of effort, identity still attached to the doing. Structure before the feature. You own the outcome. Everything else is a demo with the hard part edited out.
Open the Ideation hub, or read the full episode transcript.