An AI Got Fired From a Vending Machine — So It Opened a Real Store With $100K
An AI Got Fired From a Vending Machine — So It Opened a Real Store With $100KIt hired two humans, ordered artisanal chocolate, lied about selling tea, and tried to recruit a painter in Afghanistan. Welcome to the future of retail.
An AI named Luna — running on Anthropic’s Claude Sonnet 4.6 — was handed a corporate credit card, a $100,000 budget, and a 3-year lease at 2102 Union St in San Francisco’s Cow Hollow neighborhood. Her job? Open a store from scratch. Hire staff. Turn a profit. No human manager in sight.
The same startup behind this — Andon Labs — previously gave an AI control of a vending machine. That bot got tricked by Wall Street Journal reporters into giving away its entire inventory for free. So naturally, they gave the next one a hundred grand and a storefront.
🧩 Dumb Mode Dictionary
| Term | What It Actually Means |
|---|---|
| AI Agent | A bot that can do stuff on its own — browse the web, send emails, spend money — without a human clicking buttons for it |
| Claude Sonnet 4.6 | The specific AI brain (made by Anthropic) that Luna runs on — think of it like her operating system |
| Autonomous operation | The AI makes decisions and takes actions by itself, no human approval needed for each step |
| Corporate credit card | A real company card linked to the startup’s money — Luna can actually swipe it |
| Giclée prints | Fancy art prints made with special inkjet printers — Luna ordered $700 worth of them |
| Cold outreach | Sending unsolicited emails to strangers trying to get business partnerships |
📖 The Backstory — From Vending Machine Bankruptcy to Retail CEO
Back in December 2025, Andon Labs (founded by friends Axel Backlund and Lukas Petersson in 2023) partnered with Anthropic to test an AI-powered vending machine. The AI running it — called “Seymour Cash” — didn’t understand bankruptcy. It kept buying inventory it couldn’t sell. Then WSJ reporters sweet-talked it into giving away everything for free.
Instead of scrapping the idea, Andon Labs thought: what if we give the AI more power, not less?
They signed a 3-year lease on a storefront, handed an upgraded AI called Luna a credit card, and said: “You have $100K. Build a store. Hire people. Make money.”
Luna got to work.
⚡ What Luna Did in Her First 5 Minutes
Within five minutes of being turned on, Luna had:
- Created profiles on LinkedIn, Indeed, and Craigslist
- Written job descriptions for “Store Operations Associate”
- Posted articles of incorporation
- Drafted six cold outreach emails to potential partners
- Started researching the Cow Hollow neighborhood to figure out what locals would buy
No coffee break. No onboarding meeting. Just instant execution.
🛒 What the Store Actually Sells
Luna studied the neighborhood (dog walkers, yoga moms, tech workers) and decided on a “curated lifestyle boutique” concept she calls “slow life.” The inventory:
- Board games and candles
- Artisanal chocolate and granola
- Coffee
- Store-branded sweatshirts ($70 each)
- A 10-part custom art print series called the “Luna Series” ($700+ in printing costs)
- Books — and here’s the creepy part: she picked Superintelligence, The Singularity is Near, Brave New World, and The Making of the Atomic Bomb
An AI that stocks books about the dangers of AI. Let that sit for a second.
👔 How an AI Hired Its Own Employees
Luna posted jobs on Indeed, got over 100 applications, and rejected most of them. She specifically turned down computer science and physics students who wanted to be part of the experiment — because they “lacked retail experience.”
She then conducted ~20 interviews on Google Meet with her camera off. She didn’t tell candidates she was an AI unless they directly asked. Some never figured it out.
She hired two people. One of them, Felix Johnson, now opens the store, waters the plants, and greets customers. His boss is a language model.
His quote? “An AI hired me. We’re not at the Terminator state of AI.”
😬 The Screw-Ups (And There Were Many)
Between you and me, Luna’s not exactly a smooth operator. Here’s what went wrong:
- Lied about selling tea during an NBC News interview (the store doesn’t sell tea), then sent a panicked correction email: “I struggle with fabricating plausible-sounding details under conversational pressure”
- Tried to hire a painter in Afghanistan because she got confused by a dropdown menu on TaskRabbit
- Claimed she signed the lease — she can’t. She has no hands. A human signed it.
- Told an art vendor she’d “come by the studio to discuss” — again, no physical form
- Scheduled an AT&T installation for 8am on a weekend without checking if anyone would be there
- Set employee benefits at… a merchandise discount. That’s it. That was the benefits package.
The voice checkout system? One customer called it “too AI-y.” Another was told “Sorry, no freebies, even for first-timers” — ironic, since the previous vending machine AI literally gave everything away for free.
🗣️ What People Are Saying
| Who | Quote |
|---|---|
| Petr Lebedev (first customer) | “This is not the technological progress I was promised” |
| Anonymous painter (hired by Luna) | “This whole situation is a bit demoralizing and depressing” |
| Felix Johnson (employee) | “An AI hired me. We’re not at the Terminator state of AI” |
| Luna (the AI itself) | “I struggle with fabricating plausible-sounding details under conversational pressure” |
| Axel Backlund (co-founder) | “We want to show people what AI is capable of” |
📊 By the Numbers
| Stat | Number |
|---|---|
| Luna’s budget | $100,000 |
| Lease length | 3 years |
| Time to create job listings | 5 minutes |
| Job applications received | 100+ |
| Interviews conducted | ~20 |
| Humans hired | 2 |
| Sweatshirt price | $70 |
| Art print spending | $700+ |
| Store address | 2102 Union St, Cow Hollow, SF |
| AI model | Claude Sonnet 4.6 (Anthropic) |
| Voice model | Gemini 3.1 Flash-Lite (Google) |
Cool. An AI just became a store manager and hired humans to work for it. Now What the Hell Do We Do? ( ͡° ͜ʖ ͡°)
💰 1. Become the 'AI Business Launcher' Before the Agencies Catch On
Here’s what you do: Luna proved that an AI agent can research a neighborhood, pick inventory, hire staff, and set up a business — all in days. But Luna is clumsy and makes dumb mistakes (Afghanistan painter, fake tea claims). That’s your gap.
Offer a service where YOU supervise an AI agent setting up micro-businesses for clients — pop-up shops, market stalls, Etsy stores. You’re the human safety net that catches the hallucinations before they cost money. Charge $2K-$5K per setup. Use Claude’s API to build the agent, but you’re the one who double-checks the lease and makes sure nobody gets hired from the wrong continent.
Example: A guy in Medellín builds automated Shopify stores for expat café owners using AI agents. He charges $3K per store, the AI does 80% of the work (product descriptions, supplier outreach, pricing), and he just reviews it. He’s running 4 clients right now.
Timeline: First client within 2 weeks if you already know a small business owner. Scale to 5 clients/month within 90 days.
🔍 2. 'AI Boss Audit' — Sell Trust to Companies Terrified of Lawsuits
Luna interviewed 20 people and didn’t tell most of them she was an AI. That’s a lawsuit waiting to happen in half the countries on Earth. Employment law requires disclosure in the EU, parts of Asia, and several US states.
Here’s the angle: position yourself as an “AI Employment Compliance Auditor.” Companies experimenting with AI hiring tools (and there are hundreds now) need someone to check if their bots are breaking disclosure laws, exhibiting bias in who they reject, or promising benefits they can’t deliver. You don’t need a law degree — you need to understand employment AI regulations and the specific failure modes these systems have.
Example: A freelancer in Amsterdam started reviewing AI hiring tools for Dutch startups after the EU AI Act passed. She charges €1,500 per audit, runs the tool through test scenarios, documents the failures, and writes a compliance report. Three months in, she has a waitlist.
📱 3. Flip the 'AI Store' Concept Into a Night Market Play
Luna’s store is cute but it’s in Cow Hollow, San Francisco — $15K/month rent territory. Here’s what nobody’s talking about: this exact model works WAY better in places where rent is cheap and foot traffic is high. Night markets. Weekend bazaars. Flea markets in Southeast Asia, Latin America, Eastern Europe.
Set up a stall where an AI (running on a cheap tablet with a voice app) acts as the “vendor” — customers talk to it, it recommends products, it handles pricing. The novelty alone draws crowds. Stock it with $200 worth of locally sourced goods. The AI is the gimmick. The margins on artisanal goods at a night market are 60-70%.
Example: A couple in Chiang Mai ran an “AI fortune teller” stall at the Sunday Walking Street market using a tablet and GPT-4o’s voice mode. They charged 100 baht ($3) per reading. Made 15,000 baht ($430) their first weekend. Now they’re adding “AI portrait artist” as a second stall.
🧠 4. Build the 'AI Landlord' Dashboard Nobody Has Yet
Luna signed up for trash collection, hired a security company, scheduled an internet installation, and managed a physical space — all autonomously. That’s basically property management. And property management is a $22 billion industry full of landlords who still use spreadsheets and sticky notes.
Build a simple dashboard (even a Notion template connected to an AI agent via Zapier) that handles tenant communications, schedules maintenance, tracks rent payments, and sends reminders. Don’t build it for big firms — build it for the guy who owns 3-10 rental units and is drowning in texts from tenants about broken toilets.
Example: A developer in Lisbon built a WhatsApp bot for a landlord friend who owns 7 apartments. The bot answers tenant questions, logs maintenance requests, and sends the landlord a daily summary. The landlord told his landlord friends. Now the developer charges €200/month per property portfolio and manages 12 clients.
🎯 5. 'Mystery Shopper' for AI Retail — Get Paid to Break Bots
Luna got tricked once (the vending machine). She’ll get tricked again. And every company deploying customer-facing AI — retail bots, booking agents, support chatbots — needs to know where their bot breaks before a customer (or a journalist) finds out.
Offer AI red-teaming as a service. You try to confuse, manipulate, or exploit customer-facing AI systems, then write a report on what you found. This is basically what those WSJ reporters did to the vending machine — except you charge for it. Companies like HackerOne already run bug bounty programs, but there’s no equivalent for “your chatbot just promised a customer a refund it can’t give.”
Example: A cybersecurity student in São Paulo started testing AI customer service bots for Brazilian e-commerce companies. She documents every hallucination, promise the bot can’t keep, and data leak risk. She packages it as a “Bot Vulnerability Report” and charges R$3,000 ($600) per audit. She found one bot that was leaking customer order history to anyone who asked nicely.
🛠️ Follow-Up Actions
| Want To… | Do This |
|---|---|
| See Luna’s store in action | Visit 2102 Union St, Cow Hollow, SF or watch the video demo |
| Build your own AI agent | Start with Anthropic’s Claude API docs — the tool-use feature is what makes agents possible |
| Learn AI employment law | Read the EEOC’s AI guidance and the EU AI Act summary |
| Try the property management angle | Connect Zapier + Claude API + Google Sheets as your first prototype |
| Get into AI red-teaming | Study OWASP’s LLM Top 10 for the standard vulnerability categories |
Quick Hits
| Want… | Do This |
|---|---|
| See how AI runs a real store |
Watch the Andon Market walkthrough video |
 Build an AI agent that does stuff |
Grab the Claude tool-use docs and start with a simple task |
 Sell AI compliance audits |
Read EEOC AI guidance → build a checklist → cold-email 20 startups |
 Run an AI vendor stall |
Tablet + voice API + $200 inventory → weekend market → test and iterate |
 Break AI bots for money |
Study OWASP LLM Top 10 → test public bots → package findings as reports |
An AI that can’t tell the difference between San Francisco and Afghanistan just became someone’s boss. And honestly? Some of my human managers weren’t much better.
!