Twenty years ago, Amazon published a paper that would quietly reshape online retail. "Item-to-Item Collaborative Filtering," released in 2003, described the recommendation engine that has since driven an estimated 35% of the company's purchases. At the time, this was experimental technology, the province of a company with exceptional engineering resources and a willingness to invest in uncertain outcomes.
That paper won an IEEE award. More importantly, it established a template that the rest of the industry would spend two decades trying to replicate.
The story of AI in e-commerce is largely the story of technologies pioneered by a handful of large players gradually becoming accessible to everyone else. What Amazon, Alibaba, and a few others built as bespoke competitive advantages are now, increasingly, available as operational tools for mid-market retailers and beyond.
The numbers reflect this shift. According to recent industry research, 89% of retailers are now either actively using AI or running pilot programmes. The global AI-enabled e-commerce market, valued at approximately $7 billion in 2024, is projected to exceed $64 billion by 2034. Perhaps more tellingly, 75% of small and medium-sized businesses report that they are at least experimenting with AI tools, a figure that would have seemed improbable even five years ago.
This is no longer a story about whether AI works. It is a story about implementation.
The limits of off-the-shelf
The democratisation of AI tools has created a paradox. The same technologies that once differentiated Amazon from its competitors are now available, in some form, to almost everyone. Recommendation engines, dynamic pricing systems, customer segmentation, demand forecasting: all can be procured as services or deployed from increasingly sophisticated platforms.
This is, on the surface, good news for smaller players. But accessibility is not the same as effectiveness.
Generic AI solutions are designed to work adequately across a broad range of contexts. They are trained on general datasets, optimised for common use cases, and built to minimise implementation friction. For routine applications, such as a chatbot handling standard enquiries or basic analytics, this is often sufficient.
However, the applications that drive meaningful competitive advantage depend heavily on context specific to each business. A fashion retailer's purchasing patterns differ from those of a pet supplies company. Seasonality, promotional calendars, supplier relationships, customer demographics: all shape what "good" looks like for any given AI application.
Consider demand forecasting. Generic tools exist and can be deployed quickly. But their accuracy depends on factors that vary significantly between businesses: product lifecycle patterns, promotional sensitivity, regional demand variations. A forecast that is 70% accurate may be impressive in one context and commercially useless in another. At New Gradient, our forecasting work accounts for the specific variables, including product categories, advertising patterns, and supplier constraints, that determine whether predictions translate into better inventory decisions or simply technically competent noise.
The same principle applies to customer service. The gap between a chatbot that handles routine queries and one that genuinely understands a business's products and customer expectations is substantial. We have been developing an LLM-based system for a large pet products retailer that draws on their specific product database and years of support history, using retrieval-augmented generation to keep responses accurate and grounded. The goal is not to replace human agents but to handle volume reliably, particularly during seasonal peaks, while freeing the team to focus on queries that require judgement.
Research consistently supports this pattern: companies that develop or significantly customise AI systems outperform those relying purely on generic solutions. Not because internal development is inherently superior, but because it forces engagement with the specific problems worth solving.
From potential to operation
For e-commerce businesses navigating this landscape, several principles hold.
The question is no longer whether to use AI, but where and how. The technology has matured; non-adoption carries its own risks. The strategic question is which applications will generate meaningful returns for your specific business.
Data is typically the limiting factor, not algorithms. The most sophisticated model will underperform if trained on poor quality or unrepresentative data. Investments in data infrastructure often yield better returns than investments in more advanced AI capabilities.
And implementation matters as much as technology selection. The gap between a promising proof of concept and a system running reliably in production is where most projects fail. Clear problem definition, realistic timelines, appropriate resourcing, and the organisational capacity to integrate new tools into existing workflows: these determine outcomes more than the sophistication of the underlying model.
The pattern is consistent: AI capabilities that once required exceptional resources are now widely accessible, but the gap between generic tools and solutions tailored to a specific business's data and operations remains the difference between adequate and effective. At New Gradient, we build bespoke AI systems across industries, with particular depth in e-commerce applications like demand forecasting and customer service automation. If you are exploring where tailored AI could add value to your business, we would welcome the conversation.