Philip Hall, Managing Director Europe, Rithum, on why clean, machine-readable product data is the foundation for AI-driven demand forecasting, hyper-local demand sensing and stronger retail performance.
For years, retail IT leaders have built forecasting infrastructure on the familiar foundations of historical sales cycles, seasonal models and trend data. These systems seem outdated now but they were a necessity that were rational responses to the available data.
But the operating environment is now unrecognisable, with supply chains more volatile than ever. Disruption, changes to tariff rates or social media-driven demand can instantly render a quarterly model outdated. This isn’t news. CIOs are intimately aware that averages won’t cut it, so they’re now left with the question: how quickly can we build the data infrastructure that AI-driven demand intelligence needs to work?
The data layer limiting the AI
Conversations about AI in retail tend to fixate on pricing engines, personalised recommendations and optimised fulfilment. Ultimately, they’re all outcome-based. What they consistently underweight is the integrity and structure of the underlying product data those systems depend on.
This is where many Digital Transformation initiatives stall. Working across hundreds of retailers and brands at scale at Rithum, the bottleneck is rarely the algorithm. It’s the data the algorithm is operating on, often built from incomplete attributes and product descriptions that vary inconsistently across channels and systems.
AI applied to poor data doesn’t produce intelligent outputs, it produces high-confidence errors at speed. In other words, quick slop.
The foundational question CIOs need to answer before any predictive capability can be realised is whether their product data is machine-readable. Are SKUs enriched with the attributes, classifications and metadata that allow a model to reason about them with precision? The honest answer tends to reveal significant gaps, in which forecasting value completely disappears.
What demand sensing looks like when the architecture is right
When the foundational data is clean and correct, the shift is palpable. With AI, forecasting is transformed from retrospective and dated to continuous, multi-signal systems that integrate live data from channels, returns and fulfilment costs. It’s at this point that systems are able to predict not just volume, but margin.
That distinction is significant. Volume forecasting tells you what will sell. Margin forecasting tells you whether selling it is profitable, accounting for origin point, regional return risk and the relative cost of a stockout versus excess inventory.
Aggregated averages collapse these variables into a single number. AI-driven demand sensing keeps them separate and actionable.
Hyper-local demand signals extend this further. Rather than treating a geographic region as a uniform market, machine learning models can identify demand variation at individual stores and route inventory and advertising spend accordingly. A product trending in a specific postal area based on local seasonality, competitor activity and recent search behaviour tells a completely different story than the national average. Capturing and acting on that difference is where performance is made or lost.
Returns as a diagnostic signal
One of the most consistently underused data sources in retail technology is returns. Operationally, they’re treated as a cost to contain. Analytically, they’re a precise signal about where product information is failing customers.
Research from Rithum found that 18% of consumers identify improving the accuracy of product details – pricing, availability and specifications – as the single most important improvement retailers could make to the online shopping experience. That’s a data quality issue that lives squarely in the CIO’s domain.
When AI cross-references return patterns with product attributes and customer feedback, systemic failure points become visible and addressable. The fixes are rarely complex. Revised titles, additional attributes and updated imagery. But at scale, reducing return rates by even a percentage point or two translates into meaningful margin recovery across thousands of SKUs.
The feedback loop that compounds value
The strategic value of AI in retail commerce isn’t any single capability in isolation, but the closed loop that connects demand prediction, merchandising, media spend and fulfilment into a continuously learning system. When these functions operate from a shared, trusted data foundation, every decision informs the next.
Retailers who invest in that foundation – clean, consistent, enriched product data – drive compounding gains on every AI investment that follows. Those who don’t will find that their AI initiatives produce increasingly confident answers to the wrong questions.
