The Regret Curve: When AI Replaces Too Much, Too Fast

Simon Carver

[warmly] Welcome to the show. The episode title is The Regret Curve: When AI Replaces Too Much, Too Fast. For almost three years now, corporate leadership teams have repeated the same promise: AI will make organizations faster, leaner, and more efficient. But now something else is starting to show up beneath the headlines: regret. And if this quick take helps you think more clearly about what's happening at work, like, share, and subscribe. I'm Simon Carver, with Lachlan Reed and guest host Jack Burns.

Lachlan Reed

[matter-of-fact] Yeah, and the regret bit matters, because it's not that AI has no value. It clearly does. The problem is a lot of companies mixed up acceleration with understanding. They saw a faster workflow and thought, beauty, that's improvement. But a faster mess is still a mess, mate. It just hits the wall sooner.

Jack Burns

[calm] The problem with modern automation is not the machine. It is executive overconfidence in systems they barely understand. Efficiency dashboards are seductive because they show numbers quickly: headcount down, tickets closed, code shipped, response times cut. What they do not show nearly as well is degraded judgment, brittle operations, or trust quietly leaving the building.

Simon Carver

[curious] That phrase, trust leaving the building... that's the piece I think people miss. Because a CFO can point to, say, a 15 percent reduction in support costs. They can point to fewer junior engineering hires. They can point to productivity ratios on an earnings call. But broken trust doesn't show up as cleanly as payroll savings, does it?

Jack Burns

[sighs] Not at all. Savings are visible on a spreadsheet in a quarter. Risk often arrives later, disguised. It appears as an outage recovery bill, a compliance review, a customer escalation, a lawsuit, or an exhausted senior employee covering for a system that was supposedly autonomous. In physics terms, organizations are measuring velocity and ignoring structural stress.

Lachlan Reed

[chuckles] That's very Jack Burns. But he's right. It's like flogging an old trail bike downhill because the speedo looks great, while ignoring the front fork's about to snap. Some firms are doing that with AI in engineering, ops, HR, finance -- the whole lot. Every department gets told, move faster with AI. Governance catches up later, maybe. That's where the wobble starts.

Simon Carver

And we've already had public warning signs. Reporting around Amazon and AWS pointed to concerns about AI-assisted coding workflows -- generated code getting used more heavily, engineers leaning harder on suggestions, and then internal concern over how quickly AI-generated changes were moving into environments. Later, after AWS outages, there were internal reviews around deployment safeguards and oversight. This isn't really just an Amazon story, right? It's a preview story.

Jack Burns

[reflective] Precisely. The sequence matters. First, Wall Street demands an AI strategy. Then executives demand measurable productivity gains. Managers are then pressed for acceleration. Teams adopt tools faster than governance matures. Defects scale faster than humans can review them. What Amazon reportedly surfaced is a pattern many firms now face: fluent code is not the same as understood code.

Lachlan Reed

[skeptical] Wait -- "fluent code" is the phrase there. That's the sneaky bit. Because the output looks tidy. It reads like it knows what it's doing. And humans are suckers for that. If the machine writes with confidence, people start checking it less. That's not intelligence -- that's, I dunno, a really convincing parrot in a high-vis vest.

Jack Burns

[dryly] A fair analogy. AI does not think. It predicts. And prediction without understanding becomes dangerous at enterprise scale. We have seen versions of this before: autopilot dependency in aviation, flash crashes in algorithmic trading, over-optimized supply chains that failed under COVID stress. Automation can create the illusion that supervision is no longer necessary -- right up until reality reasserts itself.

Simon Carver

The aviation comparison sticks for me because autopilot isn't bad. It's extraordinary. But nobody sane hears "autopilot exists" and concludes "wonderful, remove the pilots." Yet inside some companies, that is almost the emotional logic of AI deployment. The tool works in narrow cases, so leadership starts fantasizing about subtracting the humans around it.

Lachlan Reed

[skeptical] And sometimes they actually do it. Hiring freezes. Junior roles vanish. Review layers get thinned out. Support teams get consolidated. Then the remaining people are told, no worries, you'll just supervise the AI. That's a funny one. Supervise it with what time? With what context? With which experienced people, after you've already shown 'em the door?

Jack Burns

That is where the regret curve begins -- exactly where the ROI slide ends. The presentation says productivity improved. The organization then discovers it has reduced resilience, concentrated expertise into fewer people, and increased what I would call automation trust drift. Humans verify carefully at first. Confidence then rises. Oversight falls. Dependency deepens. Eventually the human reviewer becomes weaker precisely because they have stopped questioning the output.

Simon Carver

[softly] Automation trust drift. That's the phrase I'm going to remember. Because it explains why "faster" and "better" split apart so easily. You can absolutely get faster while becoming more fragile, more shallow, and weirdly less aware of your own mistakes.

Simon Carver

[thoughtful] And that takes us to, for me, the most unsettling part of this whole thing: the vanishing apprentice. A lot of repetitive work looked inefficient from the outside. But inside a career, that repetition was training. It was how people learned pattern recognition, edge cases, timing, judgment. So Jack, let me put it bluntly -- where does a senior engineer come from if nobody gets to be a junior engineer anymore?

Jack Burns

[calm] That is exactly the correct question. Historically, the model was simple. A junior employee does repetitive work. Through exposure, they develop intuition. Over years, that intuition matures into judgment. They become experts and then mentor the next cohort. AI disrupts that chain by automating the entry layer first. Repetitive work disappears, the junior tier shrinks, and the expertise pipeline begins to collapse.

Lachlan Reed

And it's not just coding, hey. Coding copilots now knock out the beginner tasks that used to teach structure, syntax, debugging habits. In support teams, bots handle the high-volume tickets that used to teach reps how customers actually get confused. In analytics, AI summaries can replace the grunt work of digging through raw material. That's efficient on paper... but on the ground, you've nicked the training wheels and the bike at the same time.

Simon Carver

That support example is huge. Because a thousand routine customer conversations can feel tedious, but they also teach you what "normal" sounds like. Which means when something weird happens -- fraud, product drift, a billing pattern that makes no sense -- a human who's heard the normal stuff can spot the off note. If the bot swallows all the routine reps, fewer people ever build that ear.

Jack Burns

Exactly. Repetition is not merely labor. It is calibration. It teaches thresholds, anomalies, and consequences. In regulated sectors -- banking, healthcare, aviation, utilities, government, pharmaceuticals -- this matters even more. In a low-risk environment, hallucinations are annoying. In a high-risk environment, hallucinations become operational threats.

Lachlan Reed

[questioning tone] Let me try to play this back. You're saying the boring work wasn't just boring work. It was the gym. It was where people built the mental muscle memory. So if AI handles the first thousand reps, the human might look more productive this quarter, but ten years later you've got a company full of people who can supervise dashboards and almost nobody who truly knows the guts. Is that about right?

Jack Burns

Almost. The part I would sharpen is this: they may not even be able to supervise the dashboards well, because supervision itself requires prior contact with the underlying reality. You cannot judge an answer if you never learned how the answer is made. That is why removing the apprenticeship layer is not a staffing tweak. It is a civilizational mistake in miniature.

Simon Carver

[pauses] "Civilizational mistake in miniature" -- that's heavy, but I think it's fair. MIT-related commentary and broader workforce analysis have been warning about this exact pattern: eliminate enough entry-level roles and you don't just hurt today's graduates, you damage tomorrow's competence. You don't notice immediately, because the remaining seniors carry the system. Then one day they retire, burn out, or leave... and the institutional muscle memory goes with them.

Lachlan Reed

Yeah. And companies act shocked when they have to rehire specialized workers or reintroduce review layers after pushing too hard on automation. But that's the cost boomerang, isn't it? Save money upfront, then pay it back in outages, customer blowback, quality failures, and seniors doing rescue missions at 11 p.m. in a dark room with cold coffee. Absolute dog's breakfast.

Simon Carver

And the strange thing is, none of this requires being anti-AI. The more grounded version is actually more interesting: AI works best with human review, deterministic systems, layered governance, bounded autonomy, clear escalation paths. Not magical thinking. Not synthetic customer interaction at scale with no oversight. Not unreviewed production deployment because the demo looked slick.

Jack Burns

[reflective] Quite so. Human beings still do something AI cannot: we understand consequences emotionally. AI can calculate probabilities. It does not carry moral weight. It does not feel accountability, regret, or dignity. Which means that if an organization removes too much human judgment from its systems, it may also remove the very thing capable of protecting people from bad outcomes.

Lachlan Reed

[softly] That's the kicker, isn't it? AI should remove meaningless friction, not meaningful humanity. The future probably doesn't belong to the companies that replace the most humans. It belongs to the ones that keep the most human judgment in the loop -- even when the spreadsheet says that's slower.

Simon Carver

[warmly] And maybe that's the question to leave hanging here: what happens when no one is allowed to learn the hard way anymore? If this episode gave you something worth chewing on, share it with someone at work, and subscribe to The Human Workforce. Thanks, Lachlan. Thanks, Jack.

Jack Burns

[calm] A pleasure.

Lachlan Reed

[chuckles] Cheers, folks. Catch ya next time.