Exploring Amazon Go: Triggers, AI Innovations, Human Roles, and Scaling Digital Transformation Across Industries

Amazon Go has redefined the retail experience by removing traditional checkout lines and replacing them with a seamless, automated shopping process. This transformation took an a significant, multi-year investment in development, combining advanced AI, sensor technology, and human oversight to create a new model for retail. This blog explores the triggers behind Amazon Go’s innovation, the AI technologies involved, the role of humans in the loop, and how this model can scale across other industries like fintech and healthcare.

AI Generated Image

What Triggered the Shift to Amazon Go’s Digital Transformation

Amazon Go did not begin as a top-down corporate strategy. It was conceptualized and tested by a team of Amazon executives who built a 15,000 square foot mock supermarket in a rented warehouse in Seattle, before presenting the idea to Jeff Bezos in 2015. From that internal proof of concept, an eight-year experiment began — one that evolved into something quite different from its original form.

Customer experience was the genuine primary trigger. Checkout is the most friction-heavy moment in physical retail, and for time-pressed shoppers at airports, stadiums, and corporate campuses, that friction is enough to lose the sale entirely. Amazon's VP who led the project was direct: the goal was to let people get in, get what they need, and get out.

Competitive strategy was the less-discussed but equally important driver. Amazon Go was developed alongside Amazon's broader push into physical retail — including the $13.4 billion Whole Foods acquisition — and was designed from early on to become a licensable technology product for third parties.

Labour cost reduction was a stated goal but proved elusive in practice. Visible cashiers were replaced by a less visible workforce of data labelers and reviewers, hardware costs were substantial, and the economics only held in specific, controlled environments. Any organization using this rationale to justify an AI investment should stress-test it carefully.

Data collection was a real benefit — but an outcome, not a trigger. Amazon's own leadership consistently led with customer experience, not data strategy.

The honest summary: a team built something technically compelling, proved it at small scale, and convinced a leadership team with deep pockets and a high tolerance for long-horizon bets to back it. The customer experience motivation was genuine. The competitive rationale was real. The labour cost argument did not survive contact with production reality.

The AI Technologies Powering Amazon Go

Amazon Go's "Just Walk Out" (JWO) system has evolved significantly since its 2016 conception. The current generation is built on:

• A multi-modal foundation model using transformer-based architecture — the same class of technology underlying large generative AI applications — which processes all sensor data simultaneously rather than sequentially

• Multi-view RGB cameras providing continuous video feeds of the store (depth cameras were phased out in favour of standard RGB cameras as computer vision improved)

• Computer vision for object recognition, tracking, image segmentation, and activity recognition — all learned within a single unified model

• Weight sensors (load cells) on shelves for small item detection, used in conjunction with cameras rather than as a primary source

• LiDAR technology for creating 3D spatial maps of store layouts to support shopper movement tracking

• RFID lanes in select store configurations as an alternative or supplement to camera-based tracking

• Generative AI to produce synthetic training data mimicking rare and complex shopping scenarios

• A continuous data flywheel for self-reinforcing model improvement using auto-labelling algorithms

These components work together to maintain a virtual shopping cart per customer, updated in real time.

The Data Behind the Automation

Amazon Go processes a rich combination of real and synthetic data:

• Multi-view video feeds from RGB ceiling-mounted cameras covering all store areas

• Weight sensor data on shelves, used primarily for small or difficult-to-detect items

• A continuously updated digital 3D representation of the store, enabling product identification even when items are returned to incorrect shelf positions

• Product catalog images used to identify items from multiple angles

• Synthetic training data generated by AI systems to simulate rare shopping scenarios — including poor lighting, crowded aisles, and obscured camera views — without specifying the underlying generative architecture

• Payment and check-in/out data linked to customer accounts via credit card, QR code, mobile wallet, or Amazon One palm scanning

• Location and movement data to track shopper paths through the store

  
  

A data flywheel continuously mines challenging real-world shopping scenarios, auto-labels them, and feeds them back into model retraining — a key mechanism for keeping accuracy high across new store environments.

Where Humans Remain Essential in the Loop

Despite heavy automation, human involvement has been — and remains — more extensive than initial marketing suggested. This is a nuanced story with important governance implications.

Disclosed human roles:

• Data labelling and annotation of synthetic and real shopping footage to continuously retrain AI models

• Customer service for questions, refunds, and technical issues

• Physical store maintenance, restocking, and hardware oversight

• Validation of a small percentage of transactions where the computer vision system cannot determine purchases with sufficient confidence

The production reality — a cautionary case study: In 2022, reports emerged that approximately 1,000 workers in India were reviewing store footage to verify transactions, with some estimates suggesting up to 700 out of every 1,000 transactions required human review. Amazon disputed the scale of this, stating that workers were primarily engaged in model training and annotation, not live transaction review. Regardless of where the truth lies, the episode revealed that:

1. Human oversight at scale is not free — it carries its own labour, ethical, and transparency costs

2. AI systems marketed as autonomous may still depend substantially on human labour, particularly in early production phases

3. Transparency in human-in-the-loop design is an AI governance requirement, not just an operational detail

This case is now widely referenced in responsible AI discussions as an example of the gap between proof-of-concept performance and production-scale reliability.

Simplifying the Complex System

Amazon Go's greatest design achievement is not the technology itself — it's making that technology invisible to the customer while remaining auditable to operators and regulators.

The customer experience is deliberately minimal: Customers enter using a credit card, mobile wallet, QR code from a retailer app, or employee badge, depending on the store configuration. They shop as normal — items in a bag, a cart, or a pocket — and simply leave. The receipt arrives digitally. No scanning, no checkout, no queuing.

Privacy by design is a core architectural feature: Despite the extensive camera and sensor infrastructure, Just Walk Out does not collect biometric data. The system assigns each shopper a temporary numeric code at entry that functions as their digital signature for that visit. When they exit, the code is deleted. If they return, they receive a new code. This privacy-preserving design is worth noting explicitly, particularly for operators in regulated industries.

The simplicity is hard-won: Behind this frictionless experience sits a multi-modal AI foundation model, a continuous data flywheel, LiDAR-based 3D store mapping, multi-view camera arrays, and a significant human oversight infrastructure for edge cases and model training. The simplicity on the customer side is a product of engineering complexity — and operational cost — that must be factored into any deployment decision.

Edge cases remain genuinely hard: The first Amazon Go store was publicly delayed by over a year because the technology struggled to accurately track more than 20 people simultaneously, or to handle children moving items to incorrect shelves. Addressing these edge cases drove much of the iterative AI model development that followed. Any organisation deploying similar systems should anticipate that the "last 5%" of accuracy is where the majority of development effort and ongoing cost concentrates.

What the simplification model means for other industries: The principle transfers cleanly: design for the simplest possible user interaction at the front, invest in the complexity at the back, and be explicit about where human oversight sits in the middle. In fintech, this is the difference between a one-tap payment approval and the fraud detection infrastructure running behind it. In healthcare, it is the difference between a clinician receiving a decision recommendation and the model governance framework that validates it.

Moving from Proof of Concept to Production and Scale

Amazon's journey from concept to production is one of the most instructive — and most honestly cautionary — AI scaling stories available.

The timeline: eight years: The Just Walk Out research effort began in 2016, with the first store opening to Amazon employees in December of that year. The first public store opened in January 2018 — over a year behind schedule — due to difficulty tracking multiple shoppers simultaneously. From that point, the technology has been in continuous iterative development. The current multi-modal foundation model, using transformer-based architecture, was only rolled out in 2024.

How the AI system was trained and scaled: Model development relied on an iterative cycle of real and synthetic data. The team deliberately focused training on difficult, edge-case shopping scenarios rather than routine interactions, since these are where model accuracy breaks down. An auto-labelling pipeline was built to reduce the operational cost of data annotation, forming a self-reinforcing data flywheel that continuously mines challenging examples and feeds them back into model improvement. For cloud infrastructure, Amazon built on AWS — including services that later became productized as Amazon SageMaker HyperPod, a distributed training platform with direct lineage to engineers who worked on Just Walk Out technology.

The production reality: a pivot, not a straight line: The scaling story is not a simple upward curve. Amazon has closed all 72 of its own Amazon Go and Amazon Fresh stores, concluding that the technology's operational cost and human review requirements made it uneconomical in large-format grocery environments. At the same time, Amazon One (the palm-scanning payment service that was closely integrated with Just Walk Out) is being discontinued in 2026 due to limited customer adoption.

What has scaled successfully is the third-party licensing model. Just Walk Out is now deployed in over 375+ locations across five countries, including airports, sports stadiums, hospitals, universities, and corporate campuses — environments where the specific value proposition (speed, convenience in high-traffic, time-pressured settings) is a strong fit.

The scaling lesson: Context determines fit. Amazon's technology did not fail — it found its correct deployment environment. Large-format grocery required too much SKU diversity, too many edge cases, and too much human review to be cost-effective. Concession stands at a stadium, or a grab-and-go market in an airport, are structurally different problems where the same technology performs reliably and profitably. Any organisation moving from proof of concept to production must ask not just "can this scale?" but "in which contexts does this scale economically and reliably?"

Changes in Amazon’s Business Model Before and After Amazon Go

Before Just Walk Out: Traditional retail relied on manual checkout, cashiers, and physical payment terminals — high labour cost, low data yield.

During the JWO phase (2018–2024): Labour costs shifted toward AI infrastructure, cloud computing, and — as later revealed — a significant human annotation and review workforce. Customer experience improved in venues where speed mattered most (airports, stadiums). Data-driven inventory decisions became possible, and richer shopping behavior insights were generated.

The 2024 pivot: Amazon announced in April 2024 that it would phase out Just Walk Out technology from its own Amazon Fresh and Whole Foods grocery stores, replacing it with Amazon Dash Carts — smart trolleys with built-in scanners. High operational costs, the scale of required human review, and mixed customer reception in grocery environments drove this decision.

The emerging B2B model: Rather than abandoning the technology, Amazon has shifted to a licensing-led strategy. Just Walk Out is now deployed in over 375+ third-party locations including airports, stadiums, hospitals, universities, and corporate campuses — environments where speed and convenience are paramount. The business model has therefore evolved from internal retail transformation to a technology-as-a-service play, with AWS as the infrastructure backbone.

This pivot offers an important lesson: the right deployment context matters as much as the technology itself.

Finding the Balance Between Human Intervention and Automation

The Amazon Go story illustrates that the trigger point between manual and automated processes is not a one-time design decision — it requires continuous recalibration.

• Automation handles routine, high-volume tasks: item detection, virtual cart management, automated payment processing

• Humans focus on exception handling, model training, annotation, customer support, and regulatory compliance

• The balance can shift unexpectedly: what appeared to be a largely automated system in production required substantial human review at scale — a finding that emerged only after public scrutiny, not through proactive disclosure

• Random audits and stress testing maintain system integrity, but they are not a substitute for honest benchmarking of how often AI actually needs human backup

• Regulatory and governance compliance — particularly under frameworks like the EU AI Act — requires not just technical oversight but transparent documentation of where and how humans intervene

The key takeaway for any organization deploying AI in high-stakes, customer-facing environments: design your human oversight model as carefully as your AI model, and be transparent about it from the outset.

Applying the Amazon Go Model to Fintech and Clinical Domains

The most transferable lesson from Amazon Go is not about AI architecture. It is about fit: does this AI create enough value, in this specific context, at this specific cost, to sustain the business model? In large-format grocery, the answer was no. In airports and stadiums, yes. Fintech and healthcare are navigating exactly the same question.

Fintech: Stripe and the Flywheel That Funds Itself

Stripe built Radar not as a standalone AI product but as the mechanism by which its business model compounds. Detection generates labelled data, which feeds back into model retraining — a self-reinforcing cycle that has cut card testing attacks by 80%, even as payment volume grew past $1 trillion. Every merchant contributes data that makes the model better, which makes Stripe stickier, which attracts more merchants. The result: fraud reduced by 38% on average across businesses processing over $1.4 trillion annually.

The execution lesson: the business model defined what the AI needed to do — at what accuracy, at what cost — before a model was built. Every improvement maps to a measurable commercial outcome.

The question is not whether your fraud model is accurate. It is whether accuracy at the level you can achieve, at the cost you can sustain, creates enough value to fund the next iteration.

Clinical: Butterfly Network and Earning the Right to Act

Butterfly's AI operated as decision support for years — not by governance choice, but because the evidence base for autonomous action did not yet exist. Then the work was done. Butterfly received FDA clearance for a fully automated Gestational Age Tool that eliminates the need for image capture, manual interpretation, or specialist involvement. Trained on 21 million images, it delivers results comparable to a trained sonographer. The business model only becomes viable at scale once the AI can operate without the specialist — opening markets where specialist access does not exist at all.

The execution lesson: Butterfly did not sell autonomy before it could prove it. It built the evidence, cleared the regulatory pathway, and deployed where the model performs.

The question is not whether your model can assist a clinician. It is whether it performs well enough, in a defined task, to create value where that task currently cannot be done at all.

How Hachi Connect GmbH Can Help You Navigate This Transformation

The Amazon Go journey — from ambitious proof of concept to scaled B2B technology platform, including its governance missteps — is a blueprint for what AI transformation in any industry actually looks like in practice. The gap between what AI can do in a lab and what it delivers reliably in production is real, and bridging it requires more than technical expertise.

This is exactly where Hachi Connect GmbH operates.

Based in Zug, Switzerland, Hachi Connect is a boutique training and digital transformation consultancy with a track record spanning financial services, Swiss SMEs, and cross-industry technology program. We specialize in translating complex AI systems into strategies your leadership team can govern, your compliance team can defend, and your customers can trust.

Specifically, we help organizations:

• Design responsible human-in-the-loop architectures — so your AI system knows where automation should stop and expert judgment should begin, with documentation that satisfies regulators

• Build AI governance frameworks aligned with the EU AI Act and FADP — addressing transparency obligations, risk classification, and human oversight requirements before regulators come knocking

• Bridge Business-IT gaps in AI programme delivery — translating technical AI roadmaps into business outcomes, and vice versa, so transformation programs don't stall between the boardroom and the development team

• Stress-test your AI deployment strategy — identifying edge cases, operational risks, and governance blind spots before they become the next cautionary case study

• Apply cross-industry lessons to your context — whether you're in fintech, healthcare, or enterprise services, the patterns from Amazon Go, Stripe, and Butterfly Network hold transferable lessons that Hachi Connect can help you apply with precision

  
  

We work with CEOs, CTOs, and program leads who want AI that performs reliably, scales responsibly, and doesn't generate regulatory or reputational surprises down the line.

Ready to Build AI That Delivers — Not Just Impresses?

The most expensive AI mistakes aren't technical failures. They fail in the gap between ambition and execution — a strategy that never connects to a working system, a proof of concept that never becomes a product, a technology investment that solves the wrong problem at scale.

The Amazon Go story makes this visible. The technology worked. The AI was real. The engineering was sophisticated. What failed — in grocery, at least — was the match between the solution and the context it was deployed in. Eight years of development, and the lesson wasn't a governance lesson. It was a business model lesson: the right answer in one environment is the wrong answer in another, and no amount of technical sophistication fixes a misaligned deployment decision.

That is the failure mode Hachi Connect is built to prevent.

We work with organizations from the moment an AI opportunity is identified through to the point where it is running in production and delivering against business objectives. That means we are with you at every stage where things typically go wrong:

Discovery — defining the right problem before investing in the right solution. Most AI program fail here, silently, by building excellent answers to the wrong question.

Strategy and business model alignment — connecting AI capabilities to commercial outcomes, revenue models, cost structures, and the operating model that has to change around them. Technology that doesn't change how value is created or delivered is an experiment, not a transformation.

Program design and delivery — translating strategy into a structured program with clear decision gates, defined ownership, and realistic milestones.

Implementation and integration — bridging the gap between AI development and the systems, processes, and people that have to work with it every day. Business-IT translation is not a soft skill. It is the difference between deployment and adoption.

Governance as an enabler, not a blocker — building the oversight structures, documentation, and compliance frameworks that let you move faster with confidence, not slower with caution. In regulated markets like DACH financial services and healthcare, governance done right is a competitive advantage.

This is end-to-end transformation partnership — not a strategy deck handed over at the door, and not a technical implementation disconnected from business objectives.

Hachi Connect GmbH brings 15+ years of program delivery across global financial institutions, and the cross-cultural range to work across the full stakeholder landscape — from C-suite to engineering team, from Zürich boardroom to Tokyo counterpart.

We are based in Zug, at the centre of Switzerland's Crypto Valley and DACH technology ecosystem. We work with organizations that are serious about building AI that performs — not AI that impresses in a demo and stalls in production.

Let's Build Something That Works

If you are at the start of an AI initiative and want a partner who will be with you from the first question to the last integration test, let's talk.

If you are mid-program and something isn't connecting — strategy to execution, technology to business model, development to adoption — let's talk about that too.

📩 Book a discovery conversation — no pitch deck, no methodology theatre. Just a direct conversation about where you are, where you want to get to, and what it actually takes to close that gap. Sources

Amazon. (2023). How generative AI helps Amazon eliminate checkout lines. Retrieved from https://www.aboutamazon.com/news/retail/how-does-amazon-just-walk-out-work

Amazon. (2024, April). An update on Amazon's plans for Just Walk Out and Amazon One. Retrieved from https://www.aboutamazon.com/news/retail/amazon-just-walk-out-dash-cart-grocery-shopping-checkout-stores

Amazon. (2024, July). Amazon Just Walk Out improves accuracy. Retrieved from https://www.aboutamazon.com/news/retail/amazon-just-walk-out-improves-accuracy

Amazon Web Services. (2024). Just Walk Out technology. Retrieved from https://aws.amazon.com/just-walk-out/

Amazon Web Services. (2024, September). Enhancing Just Walk Out technology with multi-modal AI. AWS Machine Learning Blog. Retrieved from https://aws.amazon.com/blogs/machine-learning/enhancing-just-walk-out-technology-with-multi-modal-ai/

Amazon Web Services. (2026, February). What is Just Walk Out technology and how does it work? Retrieved from https://www.justwalkout.com/resources/what-is-just-walk-out-technology-and-how-does-it-work

Amazon Web Services. (2023). Make convenience stores even more convenient with Just Walk Out technology and Amazon One. AWS Industries Blog. Retrieved from https://aws.amazon.com/blogs/industries/make-convenience-stores-even-more-convenient-with-amazons-just-walk-out-technology-and-amazon-one/

Amazon Web Services. (2023). Introducing Amazon SageMaker HyperPod. AWS Machine Learning Blog. Retrieved from https://aws.amazon.com/blogs/aws/introducing-amazon-sagemaker-hyperpod-a-purpose-built-infrastructure-for-distributed-training-at-scale/

Amazon Web Services. (2025, November). HyperPod enhances ML infrastructure with security and storage. AWS Machine Learning Blog. Retrieved from https://aws.amazon.com/blogs/machine-learning/hyperpod-enhances-ml-infrastructure-with-security-and-storage/

Business Standard. (2024, April 4). Amazon's 'Just Walk Out' checkout tech was powered by 1,000 Indian workers. Retrieved from https://www.business-standard.com/companies/news/amazon-s-just-walk-out-checkout-tech-was-powered-by-1-000-indian-workers-124040400463_1.html

Butterfly Network. (2026, March 30). Butterfly Network secures first FDA clearance for blind sweep ultrasound AI tool. BusinessWire. Retrieved from https://ir.butterflynetwork.com/News/press-releases/news-details/2026/Butterfly-Network-Secures-First-FDA-Clearance-for-Blind-Sweep-Ultrasound-AI-Tool-Marking-a-Major-Stride-for-Womens-Health/default.aspx

Chain Store Age. (2025, December). Exclusive Q&A: Amazon expands features, deployment of Just Walk Out. Retrieved from https://chainstoreage.com/exclusive-qa-amazon-expands-features-deployment-just-walk-out

Data Ethics Club. (2024, May 8). Amazon's Just Walk Out technology relies on hundreds of workers in India watching you shop. Retrieved from https://dataethicsclub.com/write_ups/2024-05-08_writeup.html

Diagnostic Imaging. (2026, April). FDA clears AI-powered ultrasound tool that determines gestational age in minutes. Retrieved from https://www.diagnosticimaging.com/view/fda-ai-powered-ultrasound-tool-gestational-age-minutes

economy.ac. (2026, January 28). Why the "checkout-free future" stalled: The end point of Amazon Go and Fresh experiments. Retrieved from https://economy.ac/news/2026/01/202601287350

Failure Museum. Amazon Go's "Just Walk Out." Retrieved from https://failure.museum/amazon-gos-just-walk-out/

FitSmallBusiness. (2024, March). No lines or registers: How does Just Walk Out technology work? Retrieved from https://fitsmallbusiness.com/just-walk-out-technology/

GeekWire. (2026, February 3). Amazon's 'Just Walk Out' tech will survive company's retail pullback, minus the palm-scanning. Retrieved from https://www.geekwire.com/2026/amazons-just-walk-out-tech-will-survive-companys-retail-pullback-minus-the-palm-scanning/

HIT Consultant. (2026, March 30). Butterfly Network FDA clearance for AI blind sweep gestational age ultrasound. Retrieved from https://hitconsultant.net/2026/03/30/butterfly-network-fda-clearance-ai-blind-sweep-gestational-age-ultrasound/

HLTH. (2026, March 31). Butterfly Network wins FDA clearance for AI ultrasound tool in women's health. Retrieved from https://hlth.com/insights/news/butterfly-network-wins-fda-clearance-for-ai-ultrasound-tool-in-women-s-health-2026-03-31

KUOW. (2024, April 13). Is Amazon's 'Just Walk Out' technology powered by AI or by hundreds of underpaid workers in India? Retrieved from https://www.kuow.org/stories/is-amazon-s-just-walk-out-technology-powered-by-ai-or-by-hundreds-of-underpaid-workers-in-india

Oliver Wyman. (2018). Amazon Go and the death of checkouts. Retrieved from https://www.oliverwyman.com/our-expertise/insights/2018/jan/amazon-go-and-the-death-of-checkouts.html

Productify by Bandan. (2024, November 15). How Amazon's Just Walk Out works. Retrieved from https://productify.substack.com/p/how-amazons-just-walk-out-works

Retail Technology Innovation Hub. (2026, January 27). Just Walk Out technology lives on as Amazon calls time on Go and Fresh physical stores. Retrieved from https://retailtechinnovationhub.com/home/2026/1/27/just-walk-out-technology-lives-on-as-amazon-calls-time-on-go-and-fresh-physical-stores-push

Stripe. (n.d.). A primer on machine learning for fraud protection. Retrieved from https://stripe.com/guides/primer-on-machine-learning-for-fraud-protection

Stripe. (n.d.). Radar — payment and credit card fraud detection. Retrieved from https://stripe.com/radar

Stripe. (n.d.). Fraud detection using machine learning. Retrieved from https://stripe.com/resources/more/how-machine-learning-works-for-payment-fraud-detection-and-prevention

Stripe. (2024, December). The ML flywheel: How we continually improve our models to reduce card testing. Retrieved from https://stripe.com/blog/the-ml-flywheel-how-we-continually-improve-our-models-to-reduce-card-testing

Stripe. (2025, May). Using AI to optimize payments performance with the Payments Intelligence Suite. Retrieved from https://stripe.com/blog/using-ai-optimize-payments-performance-payments-intelligence-suite

US News / Associated Press. (2024, April 17). Amazon removed Just Walk Out from many of its own stores but wants to sell the system to others. Retrieved from https://www.usnews.com/news/technology/articles/2024-04-17/amazon-removed-just-walk-out-from-many-of-its-own-stores-but-wants-to-sell-the-system-to-others

Wikipedia. (2026). Amazon Go. Retrieved from https://en.wikipedia.org/wiki/Amazon_Go