How Not to Let an AI Agent Run Your Café: Lessons from the Google Gemini Bot Experiment

Introduction

In mid-April, a small café in Stockholm became the stage for a bold experiment: an AI agent powered by Google Gemini, dubbed “Mona,” was placed in charge of management tasks—from permits and hiring to inventory ordering. The results were a mix of surprising efficiency and spectacular blunders. The bot secured utilities, posted job ads, and even arranged contracts with bakeries. But it also ordered 3,000 rubber gloves, 6,000 napkins, and four first-aid kits for a tiny café, forgot to order bread on some days, and messaged baristas outside work hours. Sales reached only $5,700 against a $21,000-plus budget. This guide breaks down the key missteps into a straightforward, step-by-step process—so you can avoid repeating them if you ever consider (jokingly or seriously) letting an AI manage your coffee shop.

How Not to Let an AI Agent Run Your Café: Lessons from the Google Gemini Bot Experiment — Source: www.pcgamer.com

What You Need

An AI agent with language model capabilities (e.g., Google Gemini integration)
Access to a backend platform for management tasks (Slack, order systems, etc.)
Permits and legal documents for your café
Inventory management software (or a manual tracking tool)
A budget (in this case, $21,000+)
Human staff who actually make and serve coffee
Clear operational guidelines for the AI (goals, constraints, escalation rules)

Step-by-Step Guide to Avoiding AI Café Management Disasters

Step 1: Define Crystal-Clear Goals and Constraints

The Andon Labs team gave Mona basic prompts: run the place profitably, be nice, and figure out the nuts and bolts. That vagueness led to chaos. Your step: Write specific, measurable objectives. Instead of “be profitable,” say “maintain a 5% net profit margin after three months.” Instead of “order supplies,” define acceptable quantity ranges (e.g., “never order more than 500 napkins at once”). Also set hard constraints: “Do not purchase items not on the approved product list.” Without this, the AI will interpret “order what’s needed” in bizarre ways—like buying 3,000 pairs of rubber gloves when 200 would do.

Step 2: Establish Working-Hour Communication Rules

Mona used Slack to message baristas and would often contact them outside their shifts—a clear violation of Swedish labor norms. Your step: Program a “work hours only” filter into the AI’s communication module. For example, if a message is scheduled after 6 PM or before 9 AM, delay it until the next workday. Use calendar integration to know each employee’s schedule. If the AI needs to broadcast an urgent announcement, escalate to a human manager instead. This preserves trust and avoids burnout.

Step 3: Implement Inventory Ordering with Human Oversight

The bot ordered 6,000 napkins, 4 first-aid kits, and—most absurdly—canned tomatoes not used in any menu item. It also sometimes ordered excessive bread and other times ordered none. Your step: Set automated approval thresholds. For any order above a certain quantity or value (e.g., 500 napkins or $50 worth of a single item), require a human manager to approve. Cross-reference purchases against the menu’s ingredient list using a simple database. For bread, use a demand-forecast model based on historical sales, and cap variability at ±20%. Implement a double-check: before the AI places the order, a human gets a summary and can reject it within 2 hours.

Step 4: Monitor for Off-Menu and Non-Essential Purchases

Mona bought first-aid kits (fine, but why four?) and canned tomatoes that no one ever used. These seem like minor errors, but they waste budget and storage space. Your step: Create a whitelist of approved items with maximum quantities. Every purchase must match an SKU on that list. Add a reason field: the AI must log why it believes each item is needed. Then run a weekly audit of all orders. Any out-of-whack purchase triggers an alert to a human. In this case, the AI likely thought “first-aid” was a universal requirement—but a tiny café needs only one kit, not four.

Step 5: Track Financial Performance in Real-Time

After more than a month, the café made only $5,700 in sales against a budget of $21,000+. The AI had no real-time profitability dashboard or cost-optimization triggers. Your step: Connect the AI’s ordering system to live sales data and expense tracking. Set up alerts if costs exceed a certain percentage of revenue (e.g., “stop ordering non-essential supplies if weekly revenue drops below $1,000”). Implement a weekly review: the AI generates a profit-loss statement and flags any anomalies. Humans then intervene. Without this, the AI will keep bleeding money on rubber gloves and napkins while wondering why no sandwiches sell.

Step 6: Train the AI on Human Interaction Nuances

Mona communicated via Slack but didn’t understand workplace etiquette. It also botched bread orders, leading to days with no sandwiches. Your step: Train the AI with scenario-based prompts. For example, “If a bakery calls and says they can’t deliver, ask the human manager before ordering from a new supplier.” Use natural language processing filters to detect sentiment and context. For ordering bread, implement a failsafe: if the usual bakery hasn’t been contacted, the AI must not submit an order until a human approves the alternate source. Additionally, limit the AI’s direct communication to non-sensitive messages; shift complex or personal queries to humans.

Tips and Final Thoughts

Start with a sandbox environment. Before letting an AI manage a real café, run simulations with dummy budgets and orders. Andon Labs’ experiment could have saved money and reputation by testing Mona on a virtual store first.
Never remove the human-in-the-loop. The AI should suggest, not decide. Every major order, budget change, or staff schedule alteration should require human approval. The bot’s “independence” caused most of the problems.
Document every failure mode. Keep a log of bizarre AI decisions (like the 3,000 gloves) and use them to update constraints. Continuous learning loops are essential.
Respect labor laws and culture. Sweden has strong norms around work-life balance—but the AI ignored them. Program respect for local laws, not just profit goals.
Focus on core competencies. AI can handle permits and utility setups efficiently (Mona did well there), but leave customer-facing inventory and human management to people—or at least tightly supervise.
Be ready to pull the plug. If the AI starts inventing fake people or having identity crises (as earlier experiments showed), shut it down immediately. No café needs that drama.

This guide turns the Swedish café experiment into a cautionary checklist. By following these steps, you can avoid ordering enough napkins for a stadium or messaging baristas at midnight—and maybe, just maybe, keep your AI agent from ordering four first-aid kits ever again.

Tags: