As artificial intelligence continues to transform the retail industry, the term “autonomous” is becoming increasingly common. But what does true autonomy actually mean, and how can retailers distinguish genuinely intelligent systems from technologies that still depend heavily on human intervention?
In this edition of MarTech Thoughts, Fernando Dranovsky, VP of Strategic Accounts at Intuitivo, explores the difference between automation and autonomy in physical retail environments. Drawing on real-world deployments of AI-powered retail technology, he examines the challenges of handling unpredictable customer behaviour, reducing operational dependencies and measuring how independently intelligent systems can perform.
From autonomous checkout experiences to the growing importance of operational efficiency, Fernando offers a practical perspective on why the future of retail AI depends not just on what technology can automate, but on how effectively it can operate when human intervention is no longer the default.
What “Autonomous” Means in AI-Powered Retail
Retail AI has no shortage of products described as autonomous. Many stop being autonomous as soon as the environment gets unpredictable — which, in a real point of sale, happens constantly.
Automation follows predefined rules. Autonomous systems need to handle situations where those rules stop being enough. In physical retail, products are moved, returned to the wrong place, picked up in groups and handled by several people at once. A system can automate thousands of known scenarios and still depend on a person whenever it cannot establish what happened.
Automation Solves Predictable Problems
Many retail tasks are easy to automate because the possible outcomes are known in advance. Low inventory can trigger a reorder, and scanning a barcode can trigger a charge.
Retail environments rarely stay that orderly. Someone picks up four products and returns three in a different order. Two shoppers reach into the same shelf at once. A worker grabs six items for colleagues. Writing more rules does not solve these cases. The system must determine what was taken, what came back, and what to charge.
According to the National Retail Federation, retailers reported 93% more shoplifting incidents in 2023 than in 2019, while the dollar loss associated with shoplifting increased 90% over the same period.
That is a useful test for claims of autonomy: what happens when reality no longer matches the conditions the system was designed around?
Autonomy in Production
Traditional self-checkout depends on customers scanning every item correctly. In college and factory micro-markets, operators I work with routinely report shrinkage above 20 percent, much of it straightforward theft. More checkout logic does not solve the underlying problem if the system cannot independently determine what left the shelf.
In July 2024, we installed six Autonomous Points of Purchase (A-POPs) on a community college campus. In the first full month, revenue was 53 percent above the local market benchmark, with no measurable shrinkage and a modeled ten-month payback. We have since expanded across several colleges; our strongest locations generate up to 3.6 times the revenue of the setup they replaced.
At a vehicle manufacturing plant with more than 15,000 workers, over 150 A-POPs handle about 1.2 million transactions a year and roughly $5 million in GMV. Shift changes compress hundreds of transactions into minutes, while products move around the floor and regularly return to the wrong slot. Trying to enumerate every possible condition quickly becomes impractical.
Three Forms of “Fake Autonomy”
The first is human review hidden behind confidence thresholds. When input becomes ambiguous, the transaction goes to a reviewer. Accuracy may improve, but the human dependency remains and the review queue grows with volume.
Reporting around Amazon’s Just Walk Out system illustrates how contested the role of human review can become. The system required roughly 700 human reviews per 1,000 sales in mid-2022. Amazon later disputed the characterization that remote workers generated receipts, saying human associates label and annotate training data rather than watch live transactions. Either way, the debate points to a metric vendors should make explicit: how much human work sits behind an apparently autonomous transaction?
The second uses hardware to reduce the number of situations the software has to interpret: RFID on every product, weight sensors under shelves or a rigidly maintained planogram. These approaches can work, but retailers should distinguish between autonomy in the model and an environment engineered to remove ambiguity.
The third is prediction presented as autonomous decision-making. A replenishment model may forecast demand automatically while a predefined rule determines the action. The model is intelligent, but the decision process remains automated rather than autonomous.
Human intervention becomes an operating cost as deployments grow from a few pilots to hundreds of locations.
Measure Human Intervention
Model accuracy does not show how often the system needs human help. A highly accurate model can still send every uncertain case to a person.
We developed the Intelligent Autonomy Index to measure this more directly. Our fleet currently scores around 0.958 at Level 4, meaning more than 95 percent of transactions complete without human involvement. This is our own framework, and we publish the methodology so others can examine — or adopt it.
The intervention rate is also more useful as a trend than as a single number. If human review is a permanent destination for uncertainty, intervention tends to scale with transaction volume. In our case it is a fallback, and the rate has declined as the models have encountered more production scenarios. A useful question for any vendor is how its intervention rate has changed while transaction volume increased.
Infrastructure determines whether that autonomy is economical. Our pipeline processes more than 30 million images a day. Moving inference to AWS Inferentia increased throughput fivefold while reducing inference cost by 95 percent. Lower compute cost does not make a system autonomous, but it can make autonomy practical at scale.
The Next Frontier Is Operational Autonomy
Running the fleet still creates plenty of manual work. Teams deal with connectivity failures, camera issues, restocking and other day-to-day incidents. If those incidents are handled reactively, headcount continues to grow with the number of locations.
The next step is systems that can diagnose more of their own operational problems: detect camera degradation before accuracy suffers, distinguish a connectivity issue from a firmware failure, attempt routine fixes before generating a ticket and bring new locations online with less on-site support.
Nobody has finished that work, including us. The question is no longer only whether a system can complete a transaction without a person. It is how much of the surrounding operation can run without routine human intervention.
Today, we can automate most transactions without human review. We cannot yet say the same about operating the fleet. That is the problem we are working on next.
About the Author
Fernando Dranovsky is a business executive and entrepreneur with over 15 years of experience in business development, strategic partnerships, and operations management. Since 2021, he has served as VP of Strategic Accounts at Intuitivo, a technology company developing AI-powered solutions for autonomous retail. His expertise includes autonomous retail, AI adoption in physical retail, customer experience, and business growth, with a focus on the commercial applications of emerging technologies.
Connect with Fernando on LinkedIn.
Key Takeaways
Automation is not the same as autonomy: Automated systems follow predefined rules, while autonomous systems must respond to situations that cannot always be anticipated.
Human intervention is a critical measure of autonomy: Model accuracy alone does not reveal how frequently a system depends on people to complete transactions.
Real-world retail environments test AI capabilities: Unpredictable customer behaviour, misplaced products and simultaneous interactions challenge traditional rule-based systems.
Scalability requires operational efficiency: As autonomous retail deployments expand, reducing human review and operational support becomes increasingly important.
The next frontier extends beyond transactions: True operational autonomy will require systems capable of diagnosing and resolving more of their own technical and operational issues.
Continue the Conversation at The MarTech Summit
As AI continues to transform the retail landscape, the conversation is shifting from what technology can automate to how independently and effectively it can operate in real-world environments. As Fernando Dranovsky’s insights demonstrate, achieving true autonomy requires more than intelligent algorithms. It demands practical solutions that can handle unpredictable situations, minimise human intervention and deliver measurable business value.
At The MarTech Summit, we believe these real-world experiences are essential to understanding the future of marketing technology. Our global community brings together marketing, retail, technology and business leaders to exchange practical insights, challenge conventional thinking and explore how AI, data and emerging technologies are reshaping customer experiences and business operations.
Join us at our upcoming summits across Singapore, Berlin, Paris, Stockholm, Madrid, Amsterdam, Manila, Hong Kong, Milan, Dubai, Bangkok, Kuala Lumpur and London, where industry leaders will share real-world case studies and actionable strategies covering AI adoption, retail innovation, customer experience, data-driven marketing and MarTech transformation.
Whether you’re exploring AI-powered customer experiences, looking to improve operational efficiency or seeking to understand the practical applications of emerging technologies, The MarTech Summit offers a space for meaningful conversations, fresh perspectives and ideas that can drive real business impact.
Explore our upcoming events and be part of the global MarTech conversation.
➡ Interested in becoming a partner? Don’t hesitate to get in touch with us at sponsor@themartechsummit.com.
See more MarTech Thoughts interview pieces here!
Stay up to date with our In Person Summit, Session Information, and Agenda Updates by following us on:
Last updated: October 2026
