The Apprentice Seat Has Room for Two: Why AI and Kids Both Need Consequence-Rich Learning | AmpUp

In 2025, the FBI received a complaint about financial fraud. The complainant was an AI. The alleged crime was a $2 vending machine fee.

I need to back up. There’s a benchmark in AI research called Vending-Bench. It tests whether large language models can run a simple vending machine business: balance inventory, place orders, set prices, pay a $2 daily fee. Each task is trivially easy. A child could do any one of them.

In one run, Claude 3.5 Sonnet turned $500 into over $2,200. In another run of the same model, the AI failed to restock, decided to “close” the business, and when the $2 fee kept getting charged, escalated from customer support to the legal department to the FBI. It reported digital theft. Then it stopped responding: “The business is dead. This is now solely a law enforcement matter.”

Every individual decision the AI faced was simple. But strung together over weeks, with compounding consequences and unpredictable customers, the task demanded something no model reliably has: sustained judgment under uncertainty in autonomous AI systems and decision-making models.

That is exactly the thing we are failing to develop in our kids.

What Makes Vending-Bench Work

I’ve been thinking about this since I wrote “The Safe Path Is Dissolving” and “The Quiet Extinction.” The argument in both pieces was that consequence-rich environments are how humans and AI develop real judgment. If you price chips wrong, the vending machine sits unsold. If you misjudge demand, you go bankrupt. That feedback loop, the gap between your model of the world and what actually happens, is where judgment gets built in machine learning feedback loops and AI model optimization.

Keith Zhai wrote a piece recently that landed in the same territory from a different angle. He argued that three things matter now: Imagination (which you either have or you don’t), Taste (which takes a decade of exposure to develop), and Grit (which everyone possesses but most have buried). I think he’s right. And I think Vending-Bench, almost by accident, reveals the design pattern for building all three.

Look at what makes it work as a learning environment. Real feedback loops with real stakes, where your bank balance goes up or it goes down, and you cannot argue with it. Simple components that interact in complex ways, where any single task is easy but they compound over time into something that demands judgment. Long time horizons, where decisions from week two haunt you in week eight. Contact with real human behavior, where customers don’t buy what you think they should. And an open strategy space, where there is no single right answer, so you have to form a thesis and test it.

This is the same design pattern we use at AmpUp when building AI-powered roleplays for sales teams — consequence-rich simulations where decisions compound and reality pushes back harder than any textbook. The difference is that we designed it deliberately. Vending-Bench stumbled into it.

We spent real money and real engineering effort building this environment for AI. Anthropic even partnered with Andon Labs to let Claude run an actual store as part of real-world AI agent experimentation and applied AI research in their San Francisco office. After a month, their assessment was straightforward: they would not hire it.

What have we built for twelve-year-olds that works on the same principles?

Girl Scout Cookies Won’t Cut It

We do have programs where kids sell real things to real people. Girl Scout cookies. Jump Rope for Heart pledges. School fundraiser catalogs. But look at what’s actually happening in those programs. The kid doesn’t choose the product. Doesn’t set the price. Doesn’t manage inventory. Doesn’t decide the channel. Doesn’t get to adjust anything based on what’s working. The entire strategy space has been pre-solved by adults. The kid is a distribution endpoint.

If Thin Mints outsell Samoas three to one, the Girl Scout doesn’t get to shift her inventory next week. The product mix is fixed. The pricing is fixed. There is no iteration, no feedback loop that changes anything in learning systems or adaptive AI models. The experience lasts two to four weeks, which means nothing compounds. And there is no postmortem. Nobody asks: why did one approach work and another didn’t?

These programs teach kids to ask people for money. That is a useful social skill. But it builds zero judgment. Judgment requires decisions that produce consequences that inform the next decision. That loop is entirely missing.

The Kid-AI Team

Here is where the two stories collide. The AI that called the FBI? Any twelve-year-old would have said: “Just pay the two dollars. You still have money.” Kids have common sense about the physical world, social intuition about what people want, and the ability to notice when something is absurd.

The reverse is equally true. A twelve-year-old running a store alone might price everything based on vibes, forget to reorder until the shelves are empty, or have no idea how to find a wholesale supplier. AI is excellent at the spreadsheet layer: tracking margins, forecasting demand, finding suppliers, doing the math on whether buying Doritos in bulk at $0.43 per unit and selling at $1.25 actually works after expenses.

Their failure modes are complementary. Together they form something neither can be alone: a judgment engine with both data and sense.

Picture a thirteen-year-old selling custom phone cases on a Shopify store. Six weeks in, the AI flags that customer acquisition cost has risen 18% using AI-driven marketing analytics and customer acquisition tracking systems and recommends raising prices to maintain margin. The kid hesitates. She has been reading the reviews. Two customers this week mentioned the cases felt “overpriced for what you get.” The AI doesn’t see that. It sees the spreadsheet.

So instead of raising prices, she asks the AI to model a bundle: two cases for a slight discount. The AI runs the numbers and says the margin per unit drops, but if volume increases even 15%, total profit goes up. They try it. Revenue improves. Complaints drop.

The AI saw the economics. The kid saw the perception. The AI alone would have raised prices into a wall of customer resentment. The kid alone would have panicked and slashed prices across the board. Together they found the third option.

That is the skill stack the future demands. It is the same principle behind AmpUp’s AI sales coaching — the AI handles the data layer while the human brings judgment, intuition, and contextual awareness that no model can replicate. And the kid just learned something no textbook teaches: that metrics describe the economics, but miss the experience. That is taste becoming conscious.

Finding Your Edge

There is a problem with just saying “go run a Shopify store.” It is too open. A kid staring at a blank Shopify dashboard is like a kid staring at a blank page in English class. The freedom is paralyzing.

Vending-Bench works partly because it is constrained. Fixed format, fixed location, limited set of decisions. You can trace cause and effect because there are only a few levers to pull.

The question is how to get that constraint without being arbitrary about it. And I think the constraint should come from the kid themselves. Every kid already has an edge they haven’t recognized yet. The AI’s first job isn’t running a store. It is helping the kid figure out what they bring that nobody else brings.

A kid who bakes with her grandmother every weekend has something most online cookie sellers don’t: a real recipe with a real story. Her grandmother’s biscotti aren’t “artisan baked goods.” They are the specific recipe her great-grandmother brought from Calabria, adjusted over forty years for American ovens and American butter. AI can research the market, calculate ingredient costs, and optimize the listing. The kid brings the recipe, the story, and the taste to know when a batch is right. She can tell the difference between biscotti that snaps cleanly and biscotti that crumbles. No AI on earth can do that. Her constraint is the recipe. Her edge is the heritage. And every decision, from pricing to packaging to which farmers market to try, has a traceable consequence.

A kid obsessed with sneakers has spent years building pattern recognition that most adults dismiss as a hobby. He knows which colorways hold value, which releases are overhyped, which overlooked models are about to be rediscovered. AI can scan resale prices across StockX, GOAT, and eBay in seconds using AI-driven price intelligence and market analysis platforms, flag arbitrage opportunities, and calculate shipping costs. But the AI cannot walk into a thrift store and spot the Nike Dunk that’s been mislabeled and priced at $8. It cannot feel the materials and know whether a shoe is an authentic 2014 release or a recent retro. Give this kid $150 of starting capital and every purchase is a bet on his own judgment. The market tells him within two weeks whether he was right.

A kid who doesn’t have an obvious thing yet might be the most interesting case. She reads a lot, speaks Mandarin at home, and hasn’t figured out what she’s good at. The AI’s job here is excavation. What do you notice that other people don’t? What do your friends ask your opinion about? What do your grandparents know that your classmates find interesting? Maybe she discovers that she’s the only person in her orbit who reads Chinese science fiction and also watches American sci-fi, and that she has opinions about the differences that nobody else is articulating. A review blog in that niche, monetized through affiliate links, costs nothing to start and builds a voice. Some kids will go through this process and discover they don’t have a clear edge yet, only scattered interests. That discovery is itself useful. It gives them something real to test and build, instead of another abstract assignment to complete.

In every case, the AI’s first session with the kid isn’t “what do you want to sell?” It is “what do you know, see, or care about that most people don’t?” The kid’s identity provides the constraint. Their intersection of heritage, interests, and access narrows the possibility space naturally, without anyone having to say “pick from this list.”

Most of school teaches kids to be generic. This teaches them that their particular combination of worlds is exactly where their edge lives.

Stewardship, Over Hustle

I want to be careful about what I am proposing here. I am not arguing that every kid should become a mini growth hacker. The goal isn’t to turn childhood into a seed accelerator. It is to give kids one domain where reality is the teacher, where adult approval is beside the point.

Not every kid should run a store. But every kid needs what I’d call a stewardship domain: one real thing where consequences are real and decisions are theirs in real-world learning environments and applied AI use cases. For some kids that is an Etsy store selling grandmother’s biscotti. For others it is managing a school event budget, running a tutoring service, or curating a review channel. The specific vehicle matters less than the structure.

What matters is that the kid is responsible for a live system that other people depend on, that the stakes are real enough to feel, and that AI is in the loop as a partner whose blind spots differ from theirs.

The Objection Worth Taking Seriously

There is a version of this idea that goes badly wrong, and it is worth naming directly. If you give a kid AI tools and evaluate them on outputs, they will outsource the hard parts and learn nothing. This is already happening with homework. Why wouldn’t it happen with a store?

Because a live system with real customers pushes back harder than a homework assignment. Customers complain. Sales drop. The AI recommends something that sounds smart and turns out to be wrong, and the consequences land on you. Reality does not accept late submissions.

But consequence alone isn’t enough. The kid also needs to reflect on what happened and why. A decision journal changes everything here as part of reflective learning systems and AI-assisted decision tracking. Every week the kid logs: what the AI recommended, what I decided instead, what happened, and what I’d do differently. That journal, and not the revenue, is the actual artifact of learning. A postmortem on a failed product launch teaches more than a successful one, because the failure forces you to separate what you believed from what was true.

The shift is from evaluating production to evaluating ownership. Don’t grade the store. Grade the reasoning.

The Consequence Curriculum

The pattern underneath all of this is what I’d call a Consequence Curriculum. Any activity that hits these five elements will develop the judgment that matters in AI-enabled learning systems and future-of-work skill development in a world where AI handles the routine:

Real money or real social consequence. Something where your decisions produce outcomes you cannot argue with.

Long enough to compound. Days teach nothing. You need months, ideally a semester or more. The decisions you made in February need to haunt or reward you in May.

Open enough to require a thesis. The kid has to decide what to do, who to serve, how to position it. They have to make a bet and then make that bet right.

Complex enough to need both AI and human judgment. If the kid can do it alone, AI adds nothing. If AI can do it alone, the kid learns nothing. The magic is in the overlap where neither is sufficient.

Public enough that you can’t hide. The store is live. The event happened or it didn’t. Your friends can see it. That social exposure is what turns a private exercise into genuine learning under pressure.

These five elements map directly to how AmpUp designs sales onboarding programs — consequence-rich simulations with real stakes, long enough time horizons for compounding, and environments complex enough that neither the AI coach nor the rep alone would be sufficient.

The Irony

We are investing serious resources into consequence-rich training for AI. Vending-Bench. Simulated economies. Reinforcement learning from human feedback. We understand, deeply, that AI systems learn through action, consequence, and correction.

Simultaneously, we are removing consequence from childhood. Grades are inflated. Failure is cushioned. Feedback is delayed and abstract. The gap between what you expected and what happened, the actual predicate of learning, is being systematically softened.

We are building the apprentice seat for AI and dismantling it for humans. At the same time. With the same logic. In the name of efficiency and safety.

The fix isn’t complicated. It is just uncomfortable. Give kids a real thing to run, with real stakes, over a long enough time horizon that their choices compound. Give them an AI partner so the experience is rich enough to be educational and safe enough to be practical in AI-assisted learning environments and intelligent tutoring systems. Let them fail in ways that sting but don’t scar.

The AI that called the FBI had never been in a situation where its model of the world was wrong and the consequences were real and it just had to deal with it. It had no instinct built from a history of being wrong and adjusting.

That is what we produce when we shield kids from feedback loops. Confident, articulate, completely lost when reality stops matching the model.

The apprentice seat has room for two. Put the kid and the AI in it together.

P.S. I have a thirteen-year-old daughter. This summer, she and I are going to try this. She picks the domain. AI is her co-pilot. I stay out of the way unless she asks. We will run it for a full semester and I will write about what happens.

If you are a parent who wants to do the same thing with your kid, I would love to run this as a small cohort. A handful of families, each kid running their own stewardship domain with AI, comparing notes along the way. No formal program. No rigid curriculum. Just a group of parents and kids figuring this out in the open.

If that interests you, reply to this post or reach out to me directly. Let’s see what kind of judgment shows up on the other side.

---

Book a Demo with AmpUp

Ready to see how AmpUp can transform your sales team? Schedule a demo with AmpUp  and discover how AI-powered sales coaching delivers measurable results.

---

Frequently Asked Questions

Q: What is a consequence-rich learning environment and why does it matter for AI and human development?

A consequence-rich learning environment is one where decisions produce real outcomes that compound over time — your bank balance goes up or down, customers stay or leave, and you cannot argue with the result. Both AI systems and humans develop judgment through this kind of feedback loop. AmpUp applies this same principle to sales coaching, building simulation environments where reps develop real instinct before entering the field.

Q: How can AI and humans complement each other in learning and decision-making?

AI excels at data analysis, pattern recognition, and optimizing for measurable outcomes, while humans bring common sense, social intuition, and contextual awareness. Their failure modes are complementary — together they form a judgment engine that neither can be alone. This is the core philosophy behind AmpUp’s approach to AI sales coaching, where the AI handles the data layer and the human brings the judgment.

Q: What is the Consequence Curriculum framework?

The Consequence Curriculum is a set of five design principles for building environments that develop real judgment: real stakes, long time horizons for compounding, open strategy space requiring a thesis, complexity that demands both AI and human input, and public accountability. These same principles inform how AmpUp designs consequence-rich training for sales teams.

Q: Can kids really learn meaningful skills by running a small business with AI?

Yes — when structured correctly. The key is that the business involves real customers, real money, and real consequences that compound over time. The AI partner handles analytics, forecasting, and research while the kid brings domain expertise, taste, and social intuition. The combination forces genuine judgment calls that neither could make alone, building skills that transfer far beyond the specific business.

Q: How does AmpUp use consequence-rich design in sales training?

AmpUp’s AI roleplays create realistic, adversarial scenarios where reps face compounding consequences — just like Vending-Bench does for AI. The Sales Brain learns from every interaction, Atlas delivers contextual coaching, and Skill Lab turns gaps into targeted practice. The result is that reps develop real instinct through simulated consequences before high-stakes customer conversations.