An AI agent designed to shop online couldn't complete a simple purchase of an 8-pack of Pears soap. This experiment by a MediaNama editor exposes how far agentic commerce is from seamless execution, even for standardized products. For startups building autonomous shopping bots, the complexity of basic e-commerce flows remains a significant hurdle.
How We Got Here
The experiment stems from the author's January piece on agentic negotiations and a looming August 1 hackathon. This personal test aimed to understand shopping complexity before tackling negotiation bots.
The Numbers
- Saki, the AI agent, was instructed to purchase an 8-pack of Pears Pure & Gentle bathing bars.
- The agent runs DeepSeek-V4-Flash and accessed login credentials via Bitwarden's command-line interface.
- The author provided Saki with credit card details, expiry, CVV, a mobile number, and delivery address.
- The transaction value of Rs 400-500 was chosen to test a real purchase with limited financial risk.
- The experiment was a precursor to building an AI negotiation bot for a hackathon scheduled for August 1.
What Happens Next
🇮🇳 Why This Matters for India
For product managers and engineers building agentic commerce solutions in Bangalore or Pune, this experiment highlights the current gulf between AI capability and real-world e-commerce friction.
The Take
The industry focus on AI's negotiation or discovery prowess misses the point: fundamental, seemingly simple e-commerce steps like logging in or applying offers are still opaque to agents. This is where real-world AI applications get stuck, not on the flashy stuff.
Source:
MediaNama ↗