I Gave an AI Agent My Debit Card for a Week. Here's What It Bought.
Measuring whether an AI agent can autonomously research, select, and purchase products with real money while operating within predefined spending constraints.
The question
Can an AI agent be trusted to make purchasing decisions with real money?
Agentic commerce is a new model of e-commerce in which AI agents don’t just recommend products. They can research, compare, negotiate, and even complete purchases on your behalf within predefined permissions and constraints. Rather than manually browsing websites and checking out yourself, you delegate a goal to an AI agent, which reasons through the task, selects an option, and executes the purchase while you remain in control through configurable approval and spending guardrails.
Why this matters
I became fascinated by the story of Jesse Gent, a former startup CEO, who landed a feature in The Cut about using five AI agents to run her household. What locked me into her story was the fact that she let one of her AI agents make purchases on her behalf. In the moment, I couldn’t imagine letting what is essentially a text-predicting entity, such as a language model, decide when, where, and what to purchase on my behalf. As a human, I barely trust myself sometimes with that responsibility.
Why did I feel so apprehensive about letting AI make a purchase on my behalf? The obvious conclusion was that its non-deterministic nature makes its actions far too unpredictable. I couldn’t imagine calling AMEX to dispute a charge made by my AI agent. I’m sure that would be a first! But what if the outcome could be more controlled? Not to mention, what logic could trigger an AI agent to make a purchase?
So, I figured this was best tested with a custom solution created with my two hands and a debit card loaded with a dollar amount that I wasn’t too afraid to lose.
How to follow this experiment
This experiment hasn’t began yet. When it does, updates will appear on the timeline as they happen — not on a schedule. Subscribe to the Pretty Wired Labs YouTube channel to be notified, or come back to the Experiment Log to check status.