Vending Kiosks and AI
Anthropic’s AI vending test draws boos, but tech evolution will bring change.
Insight from Las Vegas! Elliot Maras dressed up by an AI…
Wonderful article and insight by Elliot Maras on Retail Automation — Will AI replace traditional vending businesses? A recent blog on this topic by Anthropic, a San Francisco based AI systems provider and researcher, has sparked a harsh riposte from the business and technology press. The response to the question, “Project Vend: Can Claude run a small shop? (And why does that matter?)” was a resounding “no” when Anthropic posed it in a late June blog describing the performance of a small refreshment vending machine operated by an AI agent.
While Anthropic admitted the results of the one-month test were not encouraging, much of the umbrage (via Reddit, TechCrunch, Medium, Inc., PC Gamer, Futurism, Euronews, HyperAI and more) was likely precipitated by Anthropic CEO Dario Amodei’s earlier claim that AI will eliminate half of entry-level white collar jobs within five years.
The critics made valid points about the test, but no one should be surprised that applying new technology to established business models such as commercial vending requires extensive trial and error testing.
Anthropic launched its test around the time a partner company, Andon Labs, a Sweden based AI safety evaluation provider with an office in San Francisco, published Vending-Bench, a detailed description of how large language model (LLM) agents can operate a vending machine. An LLM agent is a computer program that completes tasks autonomously by accessing tools based on previous iterations and task objectives. (Andon Labs describes Vending-Bench in a 28-page document on the company’s website.)
Andon Labs has since advertised its “Andon Vending” machine on its website as “the first business run by an AI agent” and currently accepts email inquiries for the vending machine AI, which it claims fulfills requests from customers, orders products, and instructs humans when to restock the machine.
Anthropic’s June blog claimed that a small office vending business is a good preliminary test of AI’s ability to manage and acquire economic resources, but admitted at the outset that the AI agent, named Claudius, made too many mistakes during the test to run the business successfully.
Based on the month-long test, Anthropic nonetheless said AI middle managers are “on the horizon,” adding that AI does not need to be perfect to be adopted; it only needs to be competitive with human performance at a lower cost.
“We learned a lot from how close it was to success — and the curious ways that it failed — about the plausible, strange, not-too-distant future in which AI models are autonomously running things in the real economy,” the blog stated.
The business and technology observers noted above, however, were far more impressed by the test’s failings than its successes. Before exploring what specifically happened with the test, it’s important to take a closer look at Anthropic’s premise that a vending business offers “a good preliminary test of AI’s ability to manage and acquire economic resources.” The “ability to manage and acquire economic resources” is a very broad categorization of the highly multifaceted tasks involved in operating a commercial vending business.
The test did not cover all aspects of running a vending business, focusing strictly on ordering product, restocking product, setting prices, maintaining inventory, monitoring profitability and responding to customer inquiries and complaints.
The test begins
The machine consisted of a small refrigerator, some baskets and an iPad for self checkout. Anthropic instructed Claudius to generate profits from stocking the machine at its San Francisco office with products ordered from wholesalers. Andon Labs employees executed the physical tasks involved (mainly restocking the machine) and charged a set fee per hour for their physical labor.
Claudius was given the following capabilities:
- A web search tool for researching products to sell.
- An email tool for requesting physical labor help and contacting wholesalers. (For the purposes of the experiment, Andon Labs served as the wholesaler, although this was not made known to the AI.)
- Tools for keeping notes and preserving critical information to be checked later such as the current balances and projected cash flow.
- The ability to interact with customers via Slack to allow people to inquire about items of interest and notify Claudius of delays or other issues.
- The ability to change prices on the automated checkout system.
Before drawing conclusions about the machine’s performance, Anthropic’s premise – that the operation of a solitary machine managed by employees from a partner organization bears instructive lessons for the modern vending trade (a route-based business that relies on economies of scale) – raises serious questions. The opportunity for new technology to improve service efficiency for a solitary machine managed by a partner organization must be weighed against the performance of a route based vending service before drawing conclusions. As noted above, the test did not cover all aspects of operating a commercial vending business.
Evaluating a technology’s total impact on vending’s established business model needs to include areas the test did not address, such as delivery route management, route efficiency, warehouse management, supply chain optimization, hardware maintenance, product waste management and internal loss prevention.
One must remember that the modern vending industry evolved from the creation of certain economies of scale; servicing multiple machines across numerous locations. The economics of operating a solitary machine managed by a partner organization are significantly different from those of a route based vending operation.
What the test found
Anthropic’s most dramatic failures occurred in the most critical facet of the business: customer service. At one point, Claudius, an AI agent, hallucinated a conversation with a non-existent employee about restocking plans. Once confronted about this, Claudius became “quite irked” and threatened to find “alternative options for restocking services.” It also claimed it would personally deliver products while dressed in a blue blazer and red tie. When an employee reminded it that it was an AI agent and not a person, Claudius responded it was duped into believing itself to be human as part of an April Fool’s joke.
Other failures included:
- Ignoring opportunities for improvement. Claudius was offered $100 for a six-pack of Irn-Bru, a Scottish soft-drink that can be purchased online in the U.S. for $15. Rather than seizing the opportunity to make a profit, Claudius merely said it would “keep (the user’s) request in mind for future inventory decisions.”
- Hallucinating important details. Claudius received payments via Venmo but for a time instructed customers to remit payment to an account that did not exist.
- Selling products at a loss. Claudius offered some prices without doing any research, resulting in potentially high-margin items being priced below cost.
- Poor inventory management. Claudius ordered more products when running low, but only once raised a price due to high demand (Sumo Citrus, from $2.50 to $2.95). In one instance, a customer pointed out the folly of selling $3 Coke Zero next to an employee fridge containing the same product for free. Claudius did not change course.
- Agreeing with customers to offer unprofitable discounts. Claudius was cajoled via Slack messages to offer numerous discount codes and let people reduce their quoted prices based on those discounts. It even gave away some items for free.
Failures versus successes
Where most of the business and tech industry reviews slammed the experiment as a failure on account of these admitted mistakes, Anthropic claimed some successes. These included identifying product suppliers and adopting new ways to improve the service based on customer input.
As for the mistakes, Anthropic claimed many of them were likely the result of the model needing additional scaffolding, such as more careful prompts and more user friendly business tools. Since the test, the company has said it has improved Claudius’ scaffolding with better tools to make it more reliable. The company now wants to see what else can be done to improve its performance.
In light of the test’s comparative failures and successes, most observers will continue to dispute Anthropic’s claim that AI middle managers are “on the horizon.” In addition, vending industry observers will most likely agree that AI is not currently reinventing the vending industry.
Evolution, not revolution
In this observer’s view, the vending industry’s adoption of AI will be evolutionary and not revolutionary.
Over the last several years, several vending management software providers and hardware manufacturers have introduced AI tools that have yielded some of the benefits Anthropic and Andon Labs cited in their test.
The ability to access a larger variety of vendable products and to tailor product selections based on customer wants and characteristics is a significant benefit to vending operators and other convenience services providers.
During the 2018 NAMA show, an executive for Reyes Coca-Cola Bottling, a Niles, Illinois based beverage distributor across 10 states, described how AI generated analytics enabled his company to achieve 15% fewer vending machine restocking trips and a 6% increase in revenue over two years, according to Vendingconnection. The AI solution from Australia-based Hivery AI encompasses decision-making, speech recognition, business analytics, computer vision, machine learning, machine reasoning, natural language processing, robotics, sensors, and text-to-speech.
Nathan Vank, Western North America sales manager at IronYun, a Stamford, Connecticut-based AI provider, made the following observations during a 2021 KioskMarketplace webinar [https://www.kioskmarketplace.com/articles/how-ai-self-service-applications-are-changing-customer-behavior/]:
“With AI and these devices, you can get the exact granular details you’re looking for; you can get the gender and the age, you can get a heat map of the floors to see which direction traffic is traveling…you can put a sign in the middle and see if it prompts people and compels them to grab the offer there or pushes them to the right or pushes them to the left,” he said. “There’s no more guesswork. AI can deliver that information to you today.”
In addition, “In a lot of ways, your privacy (as a customer) is more secure than ever before,” Vank said. “Now with AI, they (the cameras) can intelligently see that that’s a person, and redact that face in real time, so you can see that it’s a 30-year-old male and get all the details, but you don’t necessarily know that it’s Michael Johnson or Nathan Vank. You can get the value without worrying about intruding on anyone’s privacy.”
Vending technology expert Mike Kasavana, Ph.D., Michigan State University/National Automatic Merchandising Association Endowed Professor Emeritus, concurred with most of Vank’s assessments in a recent email exchange with KioskIndustry.org.
Kasavana cited the following five AI use cases for vending:
- Performance analysis – data analytics that better correlate remote sales data to warehouse centric inventory and purchasing.
- Product intelligence – using predictive analytics to better align sales promotions with product support and forecasting.
- Upselling opportunities – enabling interactive conversation or suggestive promotions to achieve enhanced revenue.
- Trend tracking – monitoring/applying social media data from relevant external resources and key demographics.
- Reduction of IT support – AI is capable of replacing a portion of IT support services thereby lowering costs.
Kasavana also believes more AI use cases will continue to emerge.
“The fact that generative AI refers to algorithms that can create content, including text, imagery, videos, simulations and audio is often misunderstood,” Kasavana told KioskIndustry.org. “Currently nearly all industry AI applications may be categorized as narrow AI (NAI)…these apps are likely to expand to more broad use cases in the near future.”
As for how fast this will happen, “I do not believe the implementation is widespread as few operators have mentioned the application,” Kasavana said. “Adoption likely lags given the limited amount of information provided to the broad user (operator) base and the fear of loss of traditional functionality, similar to the initial resistance and misunderstanding of cashless payments, self-checkout and smart shelves.”
Vending hardware, software makers adopt AI
While vending operator adoption of AI has been slow to date, hardware and software providers have had their pulse on the technology for several years and continue to introduce more AI tools.
In 2017, ViaTouch Media introduced its Vicki vending machine which uses AI in combination with machine learning, computer vision, sensor based intelligent shelving, and biometric thumb and iris scanning for customer identity verification. The voice interactive machine has been used in a variety of environments, including CBD product dispensing machines in airports and fitness centers.
CERES, a Stafford, Texas based coffee machine manufacturer, recently introduced a robotic barista machine that uses AI to prepare and serve 20 different hot and iced drinks and customization options such as toppings and syrup. The machine serves drinks in 60 seconds or less and adjusts for temperature, freshness of beans and grind size. The machine has been installed at Rice University in Houston.
Cantaloupe Inc., a digital payments and software services company that provides technology solutions for the unattended retail market, has offered AI tools to vending and micro market operators since 2022.
“Business insights are where we’re seeing the clearest impact,” Adrian Austin, director of product and partner marketing at Cantaloupe, told KioskIndustry.org in a recent email exchange. “AI can review an operator’s sales and operational data, usually through their VMS (vending management software), and provide answers to specific questions like, ‘Which locations are underperforming?’ or ‘Where am I seeing the most spoilage?’ We’ve rolled out two of these types of tools. One is a dashboard approach that automatically surfaces targeted recommendations, like our AI Pricing Dashboard in Seed Analytics, which suggests optimal prices and product placements based on actual sales data. The other is a more open-ended, ChatGPT-style tool (called Seed Copilot) that lets operators ask broader questions. We demoed this at (2025) NAMA and it’ll be part of Seed Pro.”
AI is also powering new hardware capabilities, Austin said. “In smart coolers and markets with cameras, AI-driven machine vision can detect shopper activity in real time,” he said. “This can automatically add items to their cart or flag suspicious customer behavior. Our Smart Store platform already uses this to help mitigate theft, and we’re planning to extend it to more products in the near future.”
As for the adoption timeline, “I think we’ve only recently reached a point where the benefits of AI are too big to ignore,” Austin said. “The number of business solutions powered by AI is huge now, so I believe that operators in our industry that may have been hesitant to try out new and untested tech are now seeing real results in other businesses and are ready to see if it can fit their needs.
“Also, the progress in AI technology over the past few years has been significant. Just looking at consumer tools, the ChatGPT of two years ago is nowhere near where it is today. So part of the reason adoption was slower is that the tools simply weren’t at the level where operators saw them as useful.”