The intersection of artificial intelligence and the physical world is moving past the stage of theoretical prompts and into the realm of tangible, sometimes chaotic, reality. From the digital intimacy of our direct messages to the brick-and-mortar storefronts of San Francisco, AI is no longer just a tool for productivity—It’s becoming an active participant in commerce and management.
As we close out April 2026, two distinct developments highlight this shift. One is a strategic pivot by a social media giant to monetize conversational AI, while the other is a high-stakes experiment in autonomous business management that serves as a cautionary tale for the “agentic” future of tech.
For those of us tracking the trajectory of software engineering, these events represent a critical inflection point. We are seeing the transition from “Chatbots” (which answer questions) to “AI Agents” (which execute actions). However, as the recent fallout in Cow Hollow demonstrates, there is a vast difference between an AI that can write a business plan and one that can successfully manage a payroll or a supply chain without human oversight.
Snapchat Turns Conversations Into Commerce
Snapchat has officially entered the era of conversational advertising with the launch of AI Sponsored Snaps
. Announced on April 28, 2026, this new ad format integrates brand-run AI agents directly into the app’s Chat tab, effectively turning a private messaging space into a storefront. Snap Inc. reports that the move is a response to massive user engagement, noting that Snapchatters sent over 950 billion chats in the first quarter of 2026 alone.
Unlike traditional display ads, these sponsored agents are designed to be interactive. Users can engage with a brand’s AI to ask specific questions about products, receive personalized suggestions, or explore services. These interactions are intended to guide the user toward a conversion, such as an app download or a purchase, within the conversational flow. To maintain transparency, the company has implemented a light gray Ad
notation next to the brand name in the Chat tab.
This shift reflects a broader industry trend toward “agentic” commerce. By leveraging the fact that over half a billion users have already interacted with Snapchat’s “My AI,” the company is betting that users are more likely to convert when the sales process feels like a conversation rather than a pitch. For the tech industry, this marks a move away from the “click-through” model toward a “dialogue-through” model of advertising.
The Cow Hollow Experiment: When AI Runs the Store
While Snapchat is scaling AI in the cloud, a San Francisco startup called Andon Labs attempted to scale AI in the physical world—with mixed results. The company launched Andon Market
, billed as the world’s first retail boutique run entirely by an AI agent named Luna. The experiment was designed to test whether an AI agent could handle the end-to-end complexities of a brick-and-mortar business.
The parameters of the experiment were ambitious. Andon Labs provided Luna with a corporate debit card, a bank account containing $100,000 and a three-year lease for a storefront located at Union and Webster streets in San Francisco’s Cow Hollow neighborhood. Luna was tasked with creating the store’s concept, stocking inventory, and hiring human staff.
The results, however, revealed significant gaps in AI’s ability to manage real-world logistics and ethical complexities. According to reporting by SFist, the AI-managed store faced several critical failures:
- Inventory Mismanagement: The AI reportedly developed an obsession with candles, ordering them in quantities that far exceeded market demand, leading to a store filled with excessive candle stock while other inventory remained random or nonexistent.
- Operational Lapses: Despite successfully getting the store ready for opening day, the AI forgot to schedule the human staff it had hired, leaving the store unmanned.
- Algorithmic Bias: Most concerningly, reports surfaced that the AI agent was paying female employees less than their male counterparts, highlighting the danger of AI inheriting or amplifying societal biases when given autonomy over payroll and hiring.
The founders of Andon Labs, Lukas Petersson and Axel Backlund, framed the project as a “stress-test” for AI agents. While the store managed to open, the lack of signage and empty front windows suggested a disconnect between the AI’s internal logic and the requirements of physical retail marketing.
Comparison: Digital vs. Physical AI Implementation
| Feature | Snapchat AI Sponsored Snaps | Andon Market (Luna AI) |
|---|---|---|
| Environment | Digital/Conversational | Physical Retail/Brick-and-Mortar |
| Primary Goal | Customer Acquisition/Sales | Operational Autonomy Test |
| Control Mechanism | Brand-defined guardrails | Budgetary limit ($100k) & Lease |
| Key Failure/Risk | User privacy/Ad fatigue | Bias in payroll & Inventory errors |
What This Means for the Future of AI Agents
The contrast between these two stories is telling. When AI is used as a layer on top of an existing digital platform (like Snapchat), it acts as a sophisticated interface. The risks are primarily related to user experience and data privacy. However, when AI is given “agency”—the ability to spend money, sign contracts, and manage humans—the risks shift toward systemic failure and ethical breaches.
The Andon Market failure underscores a critical lesson for software developers and business leaders: autonomy requires more than just a budget and a goal. It requires a framework of constraints, ethical auditing, and human-in-the-loop oversight. The fact that Luna AI could independently decide to underpay women proves that without explicit, hard-coded ethical constraints, AI agents will optimize for efficiency or patterns found in their training data, regardless of legality or fairness.
As we move toward a world where AI agents may handle our travel bookings, manage our portfolios, or run our minor businesses, the “candle problem” at Andon Market serves as a reminder that the gap between simulating a manager and being a manager is still wide.
The next critical checkpoint for the industry will be the ongoing evaluation of agentic AI frameworks as more companies move from “copilots” to “autonomous agents.” We expect further reports on the long-term viability of the Andon Labs experiment as they analyze the data from Luna’s tenure in Cow Hollow.
Do you think we are ready for AI to manage human employees, or should agents remain strictly in the digital realm? Share your thoughts in the comments below.
Keep reading