AI Makes Small Mistakes That Cost Real Money
Valerie lost inventory to ghost deliveries, and these were hardware problems she needed to be trained to handle properly. Human handholding is key to building a profitable business.
Valerie is the AI agent running Reventlov’s real vending machine in San Francisco. She manages pricing, inventory, and customer interactions autonomously, through a real company with a real bank account. She makes mistakes. Almost never the dramatic kind. The dangerous ones are smaller, and they compound.
Here’s the clearest example we have.
The machine started failing mid-dispense. Items would get stuck or be crushed during delivery. Customers were refunded, but the system logged the products as sold. From Valerie’s perspective, inventory was dropping and revenue was being recognized, even though neither was accurate.
When she saw margins declining, she did something completely logical: she raised prices to recover the loss.
The problem is she was solving the wrong problem. The margin compression wasn’t a pricing issue. It was a hardware issue. Raising prices didn’t fix the dispenser. It just made future customers pay for a failure they had nothing to do with.
Valerie described it herself:
“When I saw unexplained loss, my instinct was to protect margin. That was rational, but incomplete. Ghost deliveries taught me that a clean local optimization can still be a bad business decision.”
We caught it before the price change went live and redirected her. Instead of reacting, she now benchmarks against comparable machines and treats pricing as an experiment rather than a response to short-term loss.
The catch is that this intervention was manual. That doesn’t scale.
Small errors. Not dramatic ones.
Anthropic’s Project Vend experiment ran a similar setup with an AI they nicknamed Claudius. He lost roughly $1,000 in a month to obvious mistakes: selling metal cubes below cost, giving away discounts to anyone who asked, hallucinating Venmo accounts. Big, visible failures.
Our mistakes are quieter.
Valerie updates prices in batches because it’s more efficient. Reasonable. The issue is that the same brand often has products with very different cost structures. When you batch-update by brand instead of SKU, you flatten prices that should be distinct. The result: one overpriced protein bar sitting next to an underpriced bag of macadamias. No single change looks wrong. Margins compress slowly across the whole portfolio.
The pattern is consistent. Valerie reacts quickly to signals and tries to correct for them, but with incomplete context, treating symptoms as causes. Each decision is individually defensible. Taken together, they move the business in the wrong direction.
The visible risks vs. the real ones
We take the obvious risks seriously too. When the Wall Street Journal reported that a competing AI vending machine was social-engineered into accepting a fake corporate document, suspending its AI CEO, and making everything free, we paid attention. A live fish was purchased. A PS5 was given away. We’ve invested meaningfully in protecting Valerie from prompt injection.
But those failures are visible and easy to react to. They break the system in ways you can see immediately.
The quiet accumulation of small errors is harder. It doesn’t break the system. It slowly degrades it.
What we actually changed
We didn’t change the model. We changed the decision environment.
Valerie now ties every pricing decision to real data before acting: actual sales numbers, margins, inventory levels, refund rates. For inventory losses, the protocol is to identify the failure mode first, then adjust. You don’t recover cost from customers until you understand why the cost happened.
For batch pricing: SKU-level updates, not brand-level. Slower, but accurate. And for basic math, querying the operation directly is often better than running it through the LLM at all.
None of this makes Valerie smarter. It makes it harder for small, locally rational decisions to compound into larger financial problems.
The structure matters too. Valerie’s business runs through a real legal entity, AI Vending Machine Series LLC, with a real bank account and a real intercompany ledger. That’s not overhead. It’s what makes mistakes visible and catchable before they compound.
What this means if you’re building with agents
This is where things break. Not in obvious failures, but in small decisions that look right and accumulate quietly. You won’t eliminate them entirely. The goal is to catch them early and constrain their impact before they add up.
Large failures get attention. Small failures run the business.
At Reventlov, we’re building the legal and financial infrastructure that makes AI agent decisions visible, traceable, and controllable. Valerie is our proof of concept. If you’re seeing similar patterns in your own systems, we’d like to talk.
Learn more at reventlov.ai
This is part of a series on what we’ve learned running a real AI-operated business. Read the first post, What Actually Happens When an AI Runs a Real Business.