SAI Groups
SAI One – Our Most Advanced
Visual Platform
Powered by cutting-edge visual language models, SAI One understands store activity through video and context, enabling intelligent monitoring, operational insights, and real-time decision support across the entire retail environment.
Brief Overview of the Initiative
“Head office writes the playbook. This technology shows how the game is really being played.”
Developed by SAI , the platform analyses live video using advanced video-language AI. Instead of relying on narrow rule-based detection models, it builds a contextual picture of how stores actually operate — how customers move through the space, how products are handled, and how day-to-day store processes unfold.
From that understanding, the system can surface meaningful operational signals such as shelf availability gaps, out-of-stock items, planogram compliance issues, queue formation, customer engagement with promotional displays, and safety hazards like spills or blocked exits.
Importantly, the system does not bombard colleagues with alerts. Its role is to prioritise insight, helping store managers, central operations teams, and support functions understand what genuinely requires attention.
The impact is already visible. Stock count accuracy has improved by an average of 50% in deployed stores, and a number of previously high-risk stores have moved into the lowest risk quartile, reflecting greater operational stability across the estate.
For customers, this marks a fundamental shift: transforming CCTV from a passive security tool into an active operational intelligence platform.
Thinking, Insight, and Opportunity
“The biggest problem in modern stores isn’t lack of data — it’s too many systems shouting for attention.”
Retail stores are highly complex operating environments. Colleagues must balance customer service, stock replenishment, checkout operations, safety responsibilities and loss prevention — all while managing a growing number of digital systems generating their own notifications.
Online order fulfilment platforms, delivery systems, queue monitoring tools and security technologies all compete for staff attention. The result is familiar across the industry: alert fatigue.
At the same time, retailers generate enormous volumes of visual data through CCTV infrastructure, yet historically that data has provided little operational value. Footage is usually reviewed only after an incident and rarely contributes to day-to-day decision making.
Across Retailers, the issue is consistent: most operational challenges — whether related to availability, service, safety or shrink — are visible in the physical activity of the store itself.
The difficulty is that these patterns are too varied and too contextual for traditional rule-based computer vision systems to interpret reliably.
Combining all the experience gathered by SAI in the past 6 years with our cutting edge AI models, offered a new path forward. By combining computer vision with the reasoning capabilities of our own proprietary large language models, VLMs can interpret complex visual scenes and describe them in operational terms — without needing separate AI models for every specific task.
The same technology can therefore understand store activity holistically and adapt to new operational questions without retraining.
That opened the door to something genuinely different: a platform that reduces operational noise, improves availability insight, supports safer stores, and helps close the long-standing gap between head-office policy and the reality of day-to-day store operations.
Strategy and Solutions
“Many retail AI systems detect incidents. This platform understands how a store is operating.”
The strategy was simple but deliberate: build store intelligence, not another alert-generating technology. Rather than assembling a collection of separate computer vision models, SAI built the platform around a single video language model capable of interpreting store activity as a whole.
Context-Aware Store Intelligence
The system observes interactions between customers, products, shelving and checkout areas, building a continuously updated picture of how the store is functioning throughout the day. Because the VLM understands context, the same technology can identify shelf gaps, misplaced products, queue congestion, safety hazards, and merchandising compliance — all without requiring separate systems.
Increasing Signal While Reducing Operational Noise
Instead of generating a constant stream of alerts, the platform surfaces only meaningful operational insight. Store teams receive clear context when intervention is genuinely required, allowing colleagues to stay focused on customers rather than notifications.
Closing the Policy-to-Store Gap
Retail organisations routinely issue operational policies covering merchandising standards, queue management, safety procedures and store processes. Yet once these policies are issued, retailers often have limited visibility into how they are executed in practice. The VLM platform provides continuous operational awareness across the estate. Leaders can see how standards are being applied across hundreds of stores, identify friction in store workflows, and quickly adjust policies that prove impractical for store teams.
Privacy by Design
The platform deliberately avoids facial recognition and biometric identification. Instead, it analyses interactions between products, fixtures and store processes — supporting operational goals while aligning with emerging regulatory expectations around AI in public environments. A modular architecture allows the platform to operate at the edge, in the cloud or in hybrid configurations. This flexibility was essential for scaling efficiently across Iceland’s diverse store estate without requiring major capital investment in new hardware.
Objectives and, Results
“The goal was simple: help stores run better, not just watch them more closely.”
The initiative targeted measurable improvements across five areas:
- • operational visibility across the estate
- • stock accuracy and product availability
- • staff safety and situational awareness
- • reduction of alert fatigue for store teams
- • visibility into how operational policies are executed in practice
- • Labour model optimization
Since launch, the platform has delivered measurable results.
Stock count accuracy has improved by 50% across deployed stores, supporting better replenishment planning and improving on-shelf availability for customers.
46 stores have moved from high-risk status into the lowest risk quartile, reflecting stronger operational control and reduced exposure to shrink and compliance risk.
Store teams report greater confidence in managing busy environments. With clearer insight into what is happening across the shop floor, managers can act proactively rather than reactively — a meaningful shift in day-to-day operations.
Time spent reviewing CCTV footage manually has also been reduced significantly, allowing operational and loss prevention teams to focus on proactive store management rather than retrospective investigation.
Because the VLM interprets store activity broadly, benefits extend across multiple operational areas simultaneously — including queue management, safety hazard detection and merchandising compliance.
Together these results demonstrate that the initiative has achieved its core aim: strengthening operational resilience across Iceland’s store network while simplifying the working environment for the teams who run those stores.
Future Plans
The long-term goal is simple: allow retailers to ask their stores questions — and receive answers.”
The current platform represents the foundation of a broader operational intelligence capability. Future development will expand what the system understands, how it integrates with retail systems and how quickly insight reaches the people who need it.
Planned enhancements include:
- • deeper merchandising and promotional analytics
- • predictive operational risk modelling
- • expanded safety and workplace monitoring
- • automated store process reporting
- • deeper integration with POS, workforce management and inventory systems
As video language models continue to advance, the interaction model will evolve. Retail teams will increasingly be able to query store activity directly using natural language — asking questions such as:
“Where are queues building right now?”
“Which promotions are attracting the most customer attention this week?”
“Which stores need operational attention today?”
The long-term vision is a unified store intelligence layer providing real-time situational awareness across Iceland’s entire estate — a shift from systems that merely monitor stores to systems that genuinely understand how they operate.
