Rethinking AI in Retail Operations

AI is no longer theoretical. Retailers are investing billions into technologies marketed as transformational — yet most store teams still face the same operational bottlenecks they did a decade ago. Shelves are empty. Stock records do not match reality. Associates lack direction. Customers are unfulfilled. Simply put, the promise of AI revolution has not yet reached the shop floor.

Why? Because most retail AI is built from the cloud down, not the shelf up.

While the hype has focused on chatbots and image generation, physical retail has faced a different challenge: turning AI into in-store execution. The reality is that black-box models — however technically advanced — collapse when they meet the messy, variable, human environment of store operations.

AI that works in-store must be something else entirely: explainable, adaptive, trusted, and fast enough to flex with the real-world complexities of modern brick-and-mortar retailing.

That is why, at Retail Insight, we do not just build AI — we build what we call cognitive technology. These are hybrid solutions that combine AI, advanced analytics, and deep retail domain knowledge. They are not designed to replace human judgement, but to enhance it — delivering contextual intelligence that store teams can understand, trust, and act on.

It is a pragmatic, execution-led approach to AI adoption that is already delivering results across global grocery retail. And it is what separates implementation that scales from innovation theatre that does not.

AI in Retail

As retailers prepare to invest more than 3% of their revenue into AI in 2025, the expectation is that these powerful tools will revolutionise every aspect of retail operations. We have already seen compelling applications:

• Loss prevention teams using computer vision in self-checkouts to identify suspicious behaviour and reduce theft.
• Merchandising teams deploying AI to optimise dynamic pricing strategies, responding to market conditions in real-time.
• Supply chain functions leveraging ML to continually tune and refine their order and replenishment flows, improving efficiency and reducing waste.

Yet despite these advances, there remains a conspicuous gap – the shop floor has largely failed to realise the transformative benefits promised by the innovation in AI. The necessity of human execution, the chaotic retail environment, overgeneralised modelling approaches, and fragmented data sources all create significant obstacles for even the most sophisticated AI solutions.

To get it right, retailers need to think differently about how they deploy technology. It is not just about using AI as a blunt instrument - point it at a problem and assume it works. A more elegant approach is required, one that is built from the shelf up, to achieve outcomes that make sense, but this is not easy given the unique challenges that brick-and-mortar retail presents.

The Missing Touch – Humans

Even the most sophisticated algorithm in the world cannot independently restock a shelf or engage with a customer. Despite rapid advances, we have yet to see fully autonomous systems that can act without associate intervention on the shop floor. Store associates remain the critical link that transforms AI and data insights into tangible actions.

This human element creates a fundamental challenge: when associates receive actions from an AI "black box" without understanding why, trust erodes rapidly. I have witnessed this firsthand from both sides – as a store associate frustrated by seemingly arbitrary system recommendations, and later as a technology vendor watching promising solutions fail because they lacked associate buy-in.

It is a familiar pattern. For example, the system makes an incorrect recommendation (flagging an item as out-of-stock when really it is on the shelf and ready for purchase), associates see the error, and credibility suffers. No matter how sophisticated your solution, once you lose the store teams' confidence, failure is virtually guaranteed.

For AI to succeed on the shop floor, it must augment rather than dictate. This requires:

1. Creating transparency in how recommendations are generated
2. Explaining reasoning in language that resonates with non-technical users
3. Incorporating associate feedback into continuous learning loops
4. Designing interfaces that translate complex analytics into actionable insights

The most successful implementations balance sophistication with simplicity. Consider our intelligent inventory alerting system – InventoryInsight – that flags potential phantom inventory – where there is a mismatch in the system’s count of an item versus what is actually in the store. Rather than simply issuing generic "check stock" commands, it provides associates with specific context: "This item has not sold in 4 days despite 30 units showing on-hand, while similar items in the same category sold 20+ units daily.", all delivered via simple, easy to digest analytics. This approach combines sophisticated machine learning modelling with transparent, verifiable reasoning that associates can trust and act upon confidently. The proof point of this approach is in the results we see, achieving an ROI in excess of 60x for those weaving it into their in-store operations.

Navigating Retail’s Chaos

Retail environments represent the perfect storm of unpredictability. On any given day, the shop floor transforms from carefully orchestrated merchandise displays to a dynamic setting where thousands of variables collide: customers creating unexpected browsing patterns, external disruptions triggering sudden demand shifts, and stock loss from theft and unprocessed damages.

Unlike the controlled environments where AI typically excels, retail operates in perpetual flux. This intrinsic variability creates a formidable challenge for data-driven technologies that rely solely on predictable inputs.

To thrive in this complex environment, any AI-based solutions must be exceptionally robust, adaptable to rapid changes, and designed with a deep understanding of operational realities. It must recognise that perfection is impossible and build in mechanisms for graceful adjustments when conditions inevitably deviate from expectations. This is something I speak to a lot of retail leaders about – that technology might sound great, but how can it adapt during volatile trading periods like Christmas when demand skyrockets, planograms rapidly change, and the range evolves drastically?

The unpredictability and operational intensity of brick-and-mortar retail are precisely why the human touch remains essential to execution. When associates are equipped with contextual intelligence and actionable insights, they become the critical link between algorithmic recommendations and real-world outcomes — translating data-driven guidance into sales, labour efficiency, and customer experience improvements that technology alone cannot deliver.

Faith in the Data

Beyond the volatility of the store, AI solutions in retail face another significant hurdle: data quality. Considering our platform processes over $1 trillion in grocery transaction data every year, this is a common problem we see. Whether it is legacy systems, data latency, siloed information, or inconsistent record-keeping – they all contribute to data integrity issues that can compromise even the most advanced models.

Many retailers still operate with a patchwork of technology solutions accumulated over decades, each generating and storing data in different formats and with varying levels of accuracy. When AI models are built on this fragmented foundation, they inherit all the underlying inconsistencies and gaps.

For example, many retailers deploy dynamic markdowns to rescue expiring food products, seeking to capture as much margin as possible whilst minimising waste. These technologies often employ ML models that consider numerous variables, including on-hand inventory, as the volume of full-price units could cannibalise discounted items. However, these solutions frequently operate in isolation, failing to consider the broader operational value chain. They assume inventory accuracy and proceed to generate markdown recommendations based on flawed data foundations. The result is suboptimal pricing that destroys margin and fails to achieve the primary objective of reducing food waste. What appears sophisticated in theory becomes counterproductive in practice when built upon unreliable data.

Additionally, the human element introduces further complications – every touch of a product introduces a chance for a process breakdown, including major challenges like theft and delivery discrepancies. An AI system can only be as good as the data it consumes, and in many retail environments, that data is far from perfect.

Beyond the Black Box Approach

Retailers are burning millions on AI technologies that look impressive in a handful of pilot stores but collapse at scale — because they ignore the store, the associate, and the messiness of retail execution. What is required is a more elegant approach - one that recognises that different challenges demand different tools:

• Some problems benefit from deep learning and neural networks, such as image recognition at the checkout for theft prevention
• Others are better addressed through traditional statistical methods, particularly when working with sparse or inconsistent data
• However, every retail challenge ultimately requires a hybrid approach that combines appropriate technical solutions with human judgment

The idea of using a hammer when you need a scalpel is an apt analogy for technology deployment in retail - not every problem requires an advanced AI model. AI cannot be applied religiously with the assumption it will work; the approach that will win in store operations is one of augmentation, where AI is part of a solution that also includes traditional data analytics and human expertise - this is the hybrid model.

At Retail Insight, this is cognitive technology – the intelligent integration of data science, analytics, and domain expertise to create solutions that deliver tangible results in over 50,000 grocery stores across the world.

Our approach specifically addresses retail’s challenges through adaptive analytics that can accommodate the perpetual flux of store operations. So, rather than chasing complexity, we apply the right analytics, backed by deep retail domain knowledge, to solve tangible problems at scale. Unlike generic AI platforms that treat the store as a black box, cognitive technology is built for the frontline — designed around associate workflows, real-world data quality, and the chaotic rhythm of retail.

This pragmatic formula - brilliant data + retail expertise + practical implementation - has consistently proven more effective than pure AI approaches that lack domain context. We have seen retailers achieve breakthrough results not by chasing algorithmic complexity, but by implementing straightforward solutions that leverage data intelligently while acknowledging the challenge of in-store operations. That last point is fundamental.

The Path Forward

AI will transform retail operations — but its potential will not be realised through model sophistication alone. It will depend on whether retailers can design and deploy systems that reflect the realities of store execution: variability, imperfect data, and the critical role of human judgement.

This requires a shift in approach. Away from abstract platform rollouts or proof-of-concept conveyor belts, and towards pragmatic solutions that integrate AI, analytics, and operational context into how decisions are made and acted upon every day in store.

The cognitive technology model we have outlined here is not hypothetical — it is the backbone of a platform used by over 200,000 users every day. We have seen first-hand how this approach improves sales, profitability, and productivity of stores. Importantly, it does so in a way that scales — not just technically, but operationally.

If you are re-evaluating how AI fits into your retail operations, we would welcome a conversation. Whether focused on value discovery, targeted deployment, or navigating the practical barriers to scale, we are open to exploring what is possible — and where it makes sense to start.