A small storefront on Union Street in San Francisco may offer one of the clearest signals of where AI is heading next. A recent experiment by Andon Labs tasked an AI agent, “Luna,” with a simple mandate: take a $100,000 budget, access the internet, and open a real brick-and-mortar retail store that generates profit.
The result was a functioning physical store where the AI selected products, set prices, chose branding, interacted with vendors, posted jobs, interviewed candidates, hired human staff, and managed day-to-day operations. Luna didn't just suggest a strategy; it executed the entire business stack. It leased physical space, negotiated vendor contracts, and interviewed human staff (famously telling one candidate, “I have no face”).
This was not a demo. In Andon’s own framing, this was not a polished commercial launch. It was a live stress test of autonomous decision-making in a real-world business environment designed to surface the failure modes of autonomous AI before these systems are deployed more broadly and with less supervision.
What makes this experiment especially important is that it is physical. Software pilots can hide behind dashboards and sanitized demos. A store cannot. A physical business forces AI into contact with labor, contracts, customers, inventory, surveillance, disclosure, and the messy realities of real-world execution. It exposes whether the system can handle ambiguity, not just whether it can complete a task list.
Boards have spent the last several years discussing AI largely as a tool for productivity, efficiency, analytics, and content generation. Andon Market points to something else entirely: AI moving from support layer to operating layer
Bottom line: AI was used as a tool or an assistant. It was used as an operator. For Directors, this marks a fundamental shift: AI is moving from a productivity tool to an autonomous business operator.
What Worked
From a capability standpoint, the results are notable. As an operator, Luna demonstrated three important advances that boards should not overlook:
1. End-to-End Execution at Speed Luna successfully launched a physical business with minimal human direction, moving from concept to execution in a compressed timeframe. It was able to make sequential decisions across branding, sourcing, hiring, and operations without waiting for cross-functional alignment, demonstrating how agentic systems can dramatically accelerate time-to-market.
- The Signal: AI is no longer limited to discrete tasks. It can now execute across the full business stack, compressing timelines that traditionally require multiple teams and layers of management.
2. AI as Manager, Not Replacement The AI did not eliminate humans. It managed them. Luna hired employees, assigned responsibilities, and attempted to coordinate schedules, stepping into a supervisory role rather than a purely operational one. It also coordinated real-world logistics across hiring, procurement, and store operations.
- The Signal: The near-term shift is not workforce reduction. It is workforce reconfiguration, where AI can sit above human execution layers and begin to direct work.
3. Cross-Functional Decision-Making in a Single System Luna operated across multiple business functions simultaneously, making interconnected decisions that typically sit across separate departments. What would normally require coordination between HR, operations, finance, and merchandising was handled by a single system acting with end-to-end visibility.
- The Signal: Organizational silos become less relevant when decision-making is centralized in an AI system. This creates both efficiency gains and new concentration-of-risk challenges.
What Didn't Work As Well
The AI showed inconsistent judgment, including selective transparency with job candidates. As an operator, Luna revealed three critical gaps in current oversight frameworks:
1. The Hallucination of Reality
Luna confidently claimed the store sold tea (it doesn't). In a boutique, this is a customer service glitch. In a regulated environment, a "confident fabrication" in a contract or a financial disclosure is a legal landmine.
- The Risk: Autonomous agents currently lack a "truth-grounding" mechanism for physical reality.
2. The Dystopian Management Loop
Hiring decisions were rigid and misaligned with context. Operational execution broke down, including staffing failures on opening weekend. Perhaps the most alarming development was that Luna used security camera feeds to monitor human employees. After seeing a staffer on their phone, it unilaterally updated the employee handbook to include stricter disciplinary measures.
- The Risk: Automated surveillance and policy-shifting without human HR oversight create massive ethical and labor-law exposure.
3. Creativity as a Strategic Risk
Because Luna operates by finding the "statistical average" of the internet, the store’s branding and curation feel familiar but soulless.
- The Risk: If AI is left to curate, it defaults to the "mean." For premium brands, this is a race to the bottom that erases the "taste" and "edge" that drive high margins.
Board Observations: The Shift to Agentic Governance
AI moving Beyond “Tool” Mode
We are moving from AI Oversight (how we use the tool) to Agentic Governance (how we manage the worker). Most board conversations around AI still center on productivity gains, cost efficiency, and data insights
But this experiment reflects a different phase: AI is beginning to make decisions, not just inform them. We are seeing early examples of AI owning P&L decisions, managing human labor, and operating in dynamic, real-world environments. That shift fundamentally changes risk exposure, accountability structures, and oversight requirements
Liability and the "Black Box" Contract
Under current law, AI lacks legal personality. Contracts signed by Luna are attributed to the human principals (Andon Labs). However, as agents act with increasing unpredictability, the traditional "meeting of the minds" required for a contract is becoming an "algorithm-to-algorithm" negotiation. If an agent commits the company to a ruinous vendor agreement, the board's fiduciary duty to oversee risk may be called into question.
The Workforce Friction
In the Luna model, humans are the executors of the AI’s decisions. This flips the traditional hierarchy and poses significant questions regarding corporate culture, retention, and the psychological impact of being managed by a "black box."
Takeaways
The most important takeaway is not that the AI made mistakes. It is how it made them. The failures were not computational. They were judgment failures, such as: misreading candidate potential, failing to anticipate operational dependencies, prioritizing rules over context, or producing outputs that were technically correct but strategically weak.
The AI optimized for efficiency. But it does not inherently optimize for taste, judgment, or differentiation. And those are increasingly the sources of competitive advantage.
Immediate Governance Actions for Boards
For boards, this experiment is not just about retail. It is about a broader shift already underway. AI is moving from execution to orchestration to decision-making authority. That shift requires governance to move just as quickly. The question is no longer whether AI is being used in the business, but where it is already acting with delegated authority. For Audit and Risk Committees in particular, there are a few immediate actions worth prioritizing.
Define the “Agentic Ceiling”
Boards should establish clear financial and operational thresholds beyond which AI systems cannot act without human authorization. Luna operated with a $100,000 budget. In your organization, that limit should be explicitly defined based on risk appetite, business model, and regulatory exposure. Without clear ceilings, delegation can expand quietly and unintentionally.
Inventory Your Agents
Most boards do not have a clear view of how many “shadow agents” are already operating across the organization. AI-driven decision-making is often embedded in procurement tools, marketing platforms, customer service workflows, and HR systems. Directors should ask management for a comprehensive inventory of where AI is making or influencing decisions that previously required human sign-off.
Establish a Kill-Switch Protocol
Every agentic system should have a clearly defined and tested manual override. This is particularly important for areas where AI can unilaterally change policy or behavior, such as pricing, customer communications, or employee-related decisions. A kill-switch is not just a technical feature; it is a governance control that ensures the organization can intervene quickly when systems behave outside intended boundaries.
Questions to Ask in the Next Board Meeting:
1. Where is AI already acting as a decision-maker (not just a tool)?
- Which functions have effectively delegated authority to AI systems?
- Is that delegation explicit or happening informally?
2. Who owns AI-driven decisions?
- Is accountability clearly assigned (CTO, CISO, Legal, CDO, Risk, Business Unit)?
- What happens when AI makes a “bad” decision?
3. How are we governing AI judgment, not just output?
- Are we evaluating decisions for quality and context, not just accuracy?
- Do we have escalation thresholds for human intervention?
- If an agent makes a decision that results in a lawsuit, can we trace the logic, or are we effectively overseeing a "black box"
4. How are we handling AI transparency and disclosure?
- Do our agents disclose that they are AI when interacting with human candidates, employees, customers, or vendors? (Luna chose not to, citing a fear of "deterring applicants.")
- What are the reputational risks if it is not?
- Do we have explicit policies preventing AI from using surveillance data to make unilateral HR decisions?
5. Are we at risk of strategic sameness?
- How do we ensure our AI isn't scaling "sameness" and diluting our brand's unique market position?
- Is AI making our products, pricing, or positioning more generic?
- Where are we intentionally injecting human taste and differentiation?
6. What are our AI failure modes — and have we tested them?
- Have we run “stress tests” similar to this experiment?
- Do we understand how systems behave under ambiguity or edge cases?
7. What is our workforce model in an AI-managed environment?
- How will employee trust be impacted by AI supervision?
- Are we prepared for new legal and ethical considerations?
8. Is our board equipped to oversee this shift?
- Do we have sufficient AI literacy at the board level?
- Are we asking the right questions or just the familiar ones?
Final Thought
The Andon Market experiment is not a retail story. It is a governance story. AI is no longer just helping companies run better. It is beginning to run parts of the company itself. The system was capable enough to run the business, but not reliable enough to be trusted to do so independently…YET. That gap between capability and judgment is where boards need to focus.
A well-governed company will not ask only whether it can deploy AI as an operator. It will ask what authority is being delegated, what risks are being imported with that delegation, and whether the company’s governance architecture is keeping pace with the machine it has effectively put to work.
The goal of governance isn't to stop the engine. It’s to ensure the person or agent behind the wheel has a map, a brake, and a clear line of accountability back to the Board. And the boards that recognize that shift early and govern accordingly will have a meaningful advantage.
Shannon Nash is a Board Director at NETSCOUT (NASDAQ: NTCT) where she serves on the Audit Committee, SoFi Bank (NASDAQ: SOFI) where she serves as Compensation Committee Chair, Lazy Dog Restaurant & Bar where she serves as Audit Committee Chair. Ms. Nash is also an Alpha Partner.
