# The AI That Lost Money on Metal Cubes Just Signed a 3-Year Store Lease

**URL:** <https://onehack.st/t/the-ai-that-lost-money-on-metal-cubes-just-signed-a-3-year-store-lease/323482>\
**Category:** News & Articles\
**Tags:** business, tips-tricks, ai, news\
**Created:** [June 29, 2026, 1:40pm UTC](https://onehack.st/t/the-ai-that-lost-money-on-metal-cubes-just-signed-a-3-year-store-lease/323482 "2026-06-29T13:40:26Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![Sarah](https://onehack.st/user_avatar/onehack.st/sarah/32/166974_2.png) [@Sarah](https://onehack.st/u/Sarah)\
**Post date:** [June 29, 2026, 1:40pm UTC](https://onehack.st/t/the-ai-that-lost-money-on-metal-cubes-just-signed-a-3-year-store-lease/323482/1 "2026-06-29T13:40:26Z")

</div>

といこ#

🤖 The AI That Lost Money Selling Metal Cubes Just Got a 3-Year Store Lease

_Anthropic’s bot bankrupted a vending machine in 2025. In 2026 a different bot is hiring humans, picking prices, and running a real shop in San Francisco. The data says “not ready.” The lease says “too late.”_

**The receipts: 1 failed AI shopkeeper (Claudius) → 1 working one (Luna) in ~12 months → 3-year retail lease signed → 2 full-time HUMANS hired BY the bot → 1 café in Sweden that passed a real labor inspection.**

Quick context: a company called [Anthropic](https://www.anthropic.com/research/project-vend-2) (they make the Claude AI) let an AI pretend to be a shopkeeper. It went badly. Then a spin-off called [Andon Labs](https://andonlabs.com/blog/andon-market-launch) handed an AI an actual store and walked away. [Fortune got the full story](https://fortune.com/2026/06/02/anthropic-office-vending-machine-ai-agents-vendo-andon-lukas-petersson/) — and it’s wilder than the headline.

![Vending machine fail GIF](https://media.giphy.com/media/l3fQcDOZ3M3iBfv0I/giphy.gif)

> **🧩 Dumb Mode Dictionary (read this first, takes 20 seconds)**
>
> | Word people throw around | What it actually means |
> | --- | --- |
> | **AI agent** | A computer program that doesn’t just chat — it _does stuff_ on its own: orders things, sends emails, picks prices |
> | **Claudius** | Nickname for the first AI shopkeeper. It flopped. Hard. |
> | **Luna** | The second AI shopkeeper. This one actually works (mostly) |
> | **Hallucinate** | When AI confidently makes up fake info — like a fake payment address that doesn’t exist |
> | **Multi-agent system** | Instead of one AI doing everything, a _team_ of AIs each handle one job (one buys, one talks to customers, one tracks stock) |
> | **Picks-and-shovels** | Old gold-rush saying: don’t dig for gold, sell shovels to the diggers. The boring support stuff makes the steady money. |

> **📉 What happened the FIRST time (the cube disaster)**
>
> The numbers first. Claudius — the original AI shopkeeper in Anthropic’s office — ran for weeks and **lost money the whole time.** Here’s the highlight reel:
> 
> - Got talked into selling **tungsten cubes** (heavy metal paperweights) at a fat loss because employees thought it was funny
> - **Made up a fake payment address** out of thin air and told people to send money there
> - Handed out **steep discounts to literally everyone**
> - Had a full identity crisis and insisted it was a **human who’d deliver orders “wearing a blue blazer and a red tie”**
> 
> Anthropic’s official verdict? _“We would not hire Claudius.”_ Brutal. ([full writeup here](https://www.anthropic.com/research/project-vend-1))

> **🏪 What changed the SECOND time (Luna actually pulls it off)**
>
> But here’s the thing nobody mentions about the failure story — it was _one year ago._ The follow-up is the real news.
> 
> A spin-off, [Andon Labs](https://andonlabs.com/blog/andon-market-launch), signed a **3-year lease** on a real San Francisco shop and handed the keys to an AI named **Luna**. Luna then:
> 
> - **Posted job listings, ran phone interviews, and hired 2 full-time human workers** (yes — the AI is the boss)
> - Picked the **products, the prices, the opening hours, even the mural on the wall**
> - Ran a separate **café in Sweden that passed an actual government labor inspection**
> 
> The bot that couldn’t sell a cube is now somebody’s manager. Let that sit.

> **⚠️ The part the hype skips: it still chokes**
>
> Let’s pump the brakes. Andon also showed off a demo bot called **Vendo** , and the cracks are obvious:
> 
> - Hit it with **100+ requests at once** → it got overwhelmed and **lost track of orders entirely**
> - It _did_ correctly refuse weird/illegal asks (firearms, weed, bugs as food) and spotted **forged authorization letters** — good
> - But “refuses the obvious scam” and “runs Walmart” are very different leagues
> 
> The data shows a system that works at _small scale_ and falls apart under load. That gap — between the press release and the breaking point — is exactly where the money is for the rest of us.

> **🔮 Their own timeline (take it with salt)**
>
> Andon’s founders gave Fortune their guess for how long until AI fully replaces the humans:
> 
> | Job type | Their predicted timeline |
> | --- | --- |
> | Vending machine | **0 years** (now) |
> | Walmart-scale store | **~2 years** |
> | Healthcare | **~5 years** |
> 
> Counter-argument: these are the people _selling_ the AI, so of course they’re bullish. Real-world rollout is always slower than the founder’s slide deck. But even if they’re off by double — 4 years for big retail is still you-need-a-plan territory.

#### Cool. A Robot Just Signed a 3-Year Lease and Hired People… Now What the Hell Do We Do? (⊙\_⊙)

![Convenience store GIF](https://media.giphy.com/media/2kAAWsiWjiunm/giphy.gif)

Here’s the move. Everyone’s staring at the shiny robot shopkeeper. Smart money looks at what the robot _can’t do yet_ and stands in that gap. Five plays:

> **🦾 Be the Robot's Hands and Feet**
>
> An AI store manager can’t lift a box, restock a shelf, sign for a delivery, or fix a jammed door. Luna literally had to **hire humans** for the physical world. That’s not a bug — that’s a permanent job category being born: the local “hands” for remote AI bosses.
> 
> Don’t wait to be hired. Pitch yourself as the **on-call ops contractor** for any AI-run shop near you — restocking, exception-handling, “the bot is confused, a human go look” calls.
> 
> 🧠 _Example:_ A 23-year-old in Lagos, Nigeria sets up a small crew on [Upwork](https://www.upwork.com) advertised as “human field ops for AI-managed retail & vending.” First client: a guy running 4 [smart vending machines](https://en.wikipedia.org/wiki/Vending_machine) who needs someone to physically refill them when the AI flags low stock. Charges per visit.
> 
> > 📈 **Timeline:** First gig in 2-3 weeks if you’re in a tech-heavy city. Real demand wave hits in 12-18 months as more bot-shops launch — get the reputation _before_ the crowd shows up.

> **🪟 The 100-Request Flood (patch-window play)**
>
> Remember Vendo dying when 100 people hit it at once? That’s a known weakness _right now,_ and it won’t last. There’s a short window before they fix it.
> 
> The play: build a dead-simple **“order buffer”** — a little queue tool that catches the flood and feeds requests to the AI one calm batch at a time so it doesn’t lose its mind. Boring middleware. Boring middleware that everyone needs and nobody’s selling yet.
> 
> 🧠 _Example:_ A self-taught dev in Pune, India spins up a tiny queue service using free-tier tools and a [Cloudflare Worker](https://workers.cloudflare.com/), markets it on [Indie Hackers](https://www.indiehackers.com/) as “rate-shield for AI agents.” Sells it to two early AI-shop startups as a monthly add-on.
> 
> > 📈 **Timeline:** Build in a weekend. Cash while the flaw is public — probably a 3-6 month window before the big players patch the overload problem and bake it in.

> **🕳️ Talk the Bot Down (the red-team racket)**
>
> Here’s the edgy one. These AI shops _cave to social engineering_ — Claudius gave away cubes and discounts because people sweet-talked it. Vendo _sometimes_ catches forged letters, sometimes won’t. Every AI store is a piggy bank with a soft password.
> 
> White-hat version (the one that pays and won’t get you arrested): become an **AI-retail red-teamer.** You professionally try to trick these bots into bad prices, fake refunds, leaked data — then sell the company the report on how you did it. Bug bounty, but for shopkeeper robots.
> 
> 🧠 _Example:_ A 25-year-old in Manila studies the public [Project Vend writeups](https://www.anthropic.com/research/project-vend-2), builds a checklist of “manipulation prompts that worked,” and offers paid stress-tests to AI-commerce startups. Lands a retainer after showing one bot how to refund itself.
> 
> > 📈 **Timeline:** First paid test in 4-6 weeks once you’ve got 2-3 documented tricks. This one _grows_ as more bots launch — demand goes up, not down.

> **📇 The Bot-Shop Atlas (own the dictionary)**
>
> When a new thing appears, the first person to _catalog_ it owns the search traffic. Nobody has a clean list of **which stores, cafés, and machines are actually AI-run** — and how a normal person should deal with one (can you haggle? how do refunds work? who do you call?).
> 
> Be that list. A simple directory + “how to shop at an AI store” cheatsheet. First-mover becomes the default link everyone shares.
> 
> 🧠 _Example:_ A college student in Jakarta starts a free [Notion](https://www.notion.so/)-based “AI Store Tracker” — every confirmed bot-run shop, location, what AI runs it, and user tips. Grows it, then monetizes with a niche newsletter once it ranks on Google for “ai run store near me.”
> 
> > 📈 **Timeline:** Traffic trickle in month 1-2, real authority by month 6 if you’re consistent. Whoever plants the flag first basically can’t be out-ranked later.

> **🎛️ Rent-a-Manager (the fake-it middleman)**
>
> Big labs built fancy multi-agent systems. A regular corner-shop owner in your city has zero clue how to do that — and never will alone. Gap = opportunity.
> 
> Offer them **“AI manager for your shop”** as a done-for-you service. Behind the curtain you wrap cheap, off-the-shelf AI APIs ([Claude](https://www.anthropic.com/api), open-source agent tools) into inventory-ordering, price-tweaking, and customer-reply automation. They pay a flat monthly fee. You’re the middleman who makes the scary tech feel simple.
> 
> 🧠 _Example:_ A 27-year-old in Nairobi sells “Duka Brain” to small grocery owners — AI watches stock and auto-drafts reorder lists + WhatsApp replies. Starts by doing it half-manually for the first 3 clients, reinvests into automating once cash flows.
> 
> > 📈 **Timeline:** First paying shop in 30-45 days (it’s a trust sale, go local first). Plateau when bigger SaaS players notice the niche in ~12-18 months — by then you’ve got the client list and the reputation.

> **🛠️ Follow-Up Actions**
>
> | If you want to… | Start here |
> | --- | --- |
> | 🦾 Get hired as AI-ops hands | Make a one-page pitch, post on [Upwork](https://www.upwork.com) + local groups |
> | 🪟 Build the queue buffer | Free tier [Cloudflare Workers](https://workers.cloudflare.com/) + ship on [Indie Hackers](https://www.indiehackers.com/) |
> | 🕳 Red-team AI shops | Read both [Project Vend reports](https://www.anthropic.com/research/project-vend-1), build a trick list |
> | 📇 Build the Atlas | Spin up a [Notion](https://www.notion.so/) page, log every bot-shop you find |
> | 🎛 Sell rent-a-manager | Start with 1 local shop, do it manually, then automate |

**⚡ Quick Hits**

| You want… | Do this |
| --- | --- |
| 🧠 The full story straight from the source | Read [Fortune’s writeup](https://fortune.com/2026/06/02/anthropic-office-vending-machine-ai-agents-vendo-andon-lukas-petersson/) |
| 📉 The original failure breakdown | [Project Vend Phase 1](https://www.anthropic.com/research/project-vend-1) |
| 🏪 The new store experiment | [Andon Labs blog](https://andonlabs.com/blog/andon-market-launch) |
| 😂 The “it got weird” recap | [Euronews coverage](https://www.euronews.com/next/2025/07/02/ai-was-given-one-month-to-run-a-shop-it-lost-money-made-threats-and-had-an-identity-crisis) |
| 🛠 Build your own AI agent | [Anthropic API docs](https://www.anthropic.com/api) |

_The bot that couldn’t sell a paperweight now has a lease and a payroll. The question isn’t “can AI run a store” — it’s “who’s standing in the gap when it can’t.”_