As retailers begin to report measurable conversion uplifts from AI shopping assistants, the industry is moving from experimentation into early evidence of impact. But questions remain over whether these gains are sustainable once pilots scale into complex, multi-market operations. Retail Systems editor Jonathan Easton examines whether conversational commerce is delivering genuine ROI or revealing deeper structural constraints around data, integration, and trust.


The retail industry has spent the past several years speculating about, investing in the development of and starting to implement AI shopping assistants. Now, it is starting to measure them.

The latest example comes from Frasers Group, which launched its “Ask Frasers” tool across its premium fashion platform at the start of April. Upon launch, the company claimed conversion uplifts of up to 25 per cent compared with traditional search journeys. It is a headline-grabbing figure, and one that immediately raises questions about how transformative these tools could become.

Frasers is not alone. Tesco is exploring similar capabilities through its partnership with Adobe, while a growing number of UK and European retailers are piloting conversational interfaces designed to guide shoppers, rather than simply respond to keywords. The shift is subtle but significant: from search and navigation built around retail logic, to interfaces that attempt to interpret customer intent in real time.

What makes this moment different is not the technology itself, but the emergence of measurable commercial impact. Conversion uplift, basket growth, and engagement metrics are beginning to move in ways that retailers can quantify and justify.

Yet the numbers alone do not settle the debate. Early results may reflect novelty, selective user groups, or carefully controlled deployments. The deeper question is whether AI assistants can sustain these gains at scale, across categories, and within the messy reality of retail operations.

In other words, are tools like Ask Frasers the first clear proof that conversational commerce delivers ROI, or simply the most visible example of a trend that is still far from maturity?



The headline numbers – real impact or selective measurement?

If Frasers’ reported uplift signals a breakthrough moment, the industry response is notably cautious. The consensus is not that the numbers are wrong, but that they require far more context than a headline allows.

For retailers already operating AI at scale, the impact is real. Kingfisher offers one of the clearest counterpoints. The group, which operates banners including B&Q and Screwfix, has been embedding AI across its business for several years, developing more than 50 services and deploying assistants such as Hello B&Q and Hello Casto.

According to Mohsen Ghasempour, chief AI officer at Kingfisher, the results are material. “We see significantly higher conversion rates, in some cases more than double, from customers who engage with our AI agents,” he says. “Our recommendation and personalisation engines accounted for around £165 million of group sales in FY25/26.”

That kind of performance suggests the underlying premise holds: guided, conversational journeys can move customers from uncertainty to purchase more effectively than traditional search alone.

However, analysts are quick to stress that not all uplift is created equal. Julian Skelly, managing partner for retail at Publicis Sapient, points to the importance of measurement. “A 25 per cent uplift in conversion from users who engage with the AI assistant is a very different claim to a 25 per cent uplift across the full digital estate,” he says. “Early adopters of any new interface tend to be more motivated shoppers, which naturally flatters the numbers.”

That point is echoed by James Heimers, executive vice president of marketing sciences at agency RAPP, who notes that headline gains are often relative rather than absolute. “In most cases this is a relative uplift on a low baseline, so the absolute gain is more modest than it first appears,” he says.

There is also the question of attribution. Nick Dutton, consultant at Leading Resolutions, argues that gains of this magnitude are rarely driven by a single factor. “The AI assistant was likely deployed alongside other variables, such as targeted discounting,” he says. “Without transparent attribution modelling, it is difficult to credit the tool by itself.”

Taken together, the picture is more nuanced than the initial claim suggests. AI assistants are clearly capable of driving meaningful commercial impact, but the scale of that impact depends heavily on how it is measured, who is using the tool, and what other changes are happening alongside it.


Where AI assistants actually work and where they don’t

The effectiveness of AI shopping assistants is not evenly distributed across retail. Performance is closely tied to the nature of the shopping mission itself.

At Kingfisher, the strongest results are seen in categories where customers need guidance. Its assistants are designed to support more technical, considered purchases, helping shoppers navigate product specifications, compatibility, and project requirements. “They provide tailored advice and support online, which is critical particularly where products can be more technical,” says Ghasempour.

This reflects a broader industry pattern. Publicis Sapient’s Skelly points to problem-solving as the defining factor. “The strongest impact will consistently be in categories where the customer has a problem they can’t easily solve themselves,” he says, highlighting areas such as fit, technical specification, occasion dressing, and gifting.

Consumer behaviour data reinforces that view. Research from Attest indicates that nearly half of shoppers are already using generative AI tools to research purchases, with the highest engagement in high-friction categories such as fashion, beauty, and electronics. In these environments, customers are often uncertain about what to buy, and open to guidance that narrows the field and builds confidence.

In contrast, more habitual retail missions offer a narrower opportunity. Grocery is the most obvious example. Here, speed and familiarity tend to outweigh exploration. As Skelly notes, “the customer broadly knows what they want.” In this context, AI is more likely to add value through efficiency – building baskets, suggesting substitutes, or supporting repeat purchasing – rather than acting as a discovery engine.

Kim Johal, retail specialist at Infobip, makes a similar distinction. “Fashion and lifestyle is the most natural fit for AI assistants because the purchase decision is subjective and guidance dependent,” he says. In grocery and other low-consideration categories, the role becomes more functional, focused on task completion rather than inspiration.

There is also a shift underway in how product discovery itself is structured. Lisa Smith, partner at Kearney, argues that conversational interfaces represent a fundamental change in design philosophy. “They’re closing the gap between what a customer means and what they actually find,” she says. “Traditional navigation was built around the retailer’s logic. Conversational AI is built around the customer’s logic.”

Taken together, these perspectives point to a clear conclusion. AI assistants deliver the greatest impact where they reduce uncertainty and simplify complex decisions. In more routine, repeat-driven journeys, their role is more incremental, improving efficiency rather than fundamentally reshaping behaviour.



Evolution or disruption – what happens to search?

The rise of AI assistants inevitably raises a larger structural question: what happens to the search bar?

At Kingfisher, the current view is pragmatic. AI is being positioned as an additional layer rather than a replacement. “We see AI shopping assistants as an evolution of the existing e-commerce offer, rather than a replacement for it,” says Mohsen Ghasempour. “Traditional search still matters, but AI assistants can add a more personalised, conversational layer.”

That position is widely shared, at least in the short term. Many retailers are integrating assistants alongside existing navigation and search tools, allowing customers to choose how they engage. The underlying mechanics of merchandising, ranking, and product retrieval remain intact.

Yet there are signs that this balance may not hold indefinitely. Adam Davies, director of AI at Forter, describes a growing divergence between two models. The first is the on-site assistant, embedded within the retailer’s own environment. “In many ways, this is an evolution of the existing stack,” he says, giving retailers another way to influence discovery while retaining control over data and the customer relationship.

The second model is more disruptive. Third-party agents, including platforms such as ChatGPT, Gemini, and Perplexity, are beginning to transact on behalf of customers. In this scenario, discovery may happen entirely within the retailer’s own channels. “The customer may never touch the merchant’s domain,” Davies says, shifting the focus from optimising on-site journeys to ensuring products can be found, interpreted, and recommended by external agents.

This introduces a new competitive dynamic. Retailers must not only serve human shoppers, but also machine intermediaries, requiring investment in catalogue structure, metadata, and what is increasingly being described as generative engine optimisation.

Even without that shift, the role of search is already changing. Lisa Smith at Kearney argues that conversational interfaces invert the traditional model. “The search bar was always a blunt instrument,” she says. “AI assistants are closing the gap between what a customer means and what they actually find.” In practice, that means search and navigation may increasingly become fallback options, rather than the primary entry point.

The likely outcome is not immediate displacement, but gradual convergence. As Piyush Patel, chief ecosystem officer at Algolia puts it, “this is less about replacement than convergence,” with conversational interfaces sitting on top of established search and merchandising infrastructure. Over time, however, the balance of power between those layers may shift.



The real constraint – data, not AI

Across every perspective, one theme emerges with unusual consistency: the limiting factor is not the intelligence of the model, but the quality of the data behind it.

For Kingfisher, this has been a central focus. Scaling AI assistants across multiple banners and markets has required sustained investment in platforms, data architecture, and integration. “The biggest barrier to scaling AI assistants, as with most AI services, is having the right foundations in place,” says Mohsen Ghasempour. “This includes high-quality, well-structured data, as well as scalable platforms and governance.”

That view is widely shared. Julian Skelly describes “data plumbing” as “retail’s Achilles’ heel”, arguing that many organisations are still trying to layer AI onto fragmented estates. The result is predictable: pilots perform well in controlled environments, but struggle when exposed to the full complexity of live operations.

A recurring issue is the imbalance between customer and product data. Sarah Arana-Morton, chief executive of Provenance, argues that retailers have focused heavily on understanding shoppers, while neglecting the data needed to describe what they are actually selling. “Retailers have invested heavily in understanding the customer, but far less in understanding the product,” she says. “One incorrect claim can break trust,” particularly in high-consideration categories where shoppers rely on accurate, verifiable information.

At a more operational level, even basic inconsistencies can undermine performance. Alex Moseman, retail and high tech expert at PA Consulting, points to internal disagreements over product taxonomy. “We often see teams debating where a product should sit within the catalogue,” he says. “If teams cannot agree, the AI cannot reliably surface the product.” These issues may appear minor, but in a conversational interface they surface immediately as irrelevant or misleading recommendations.

The integration challenge is equally significant. Product information, inventory, pricing, loyalty, and customer identity are typically spread across multiple systems, often with incompatible data models. Kim Johal, retail specialist at omnichannel marketing platform Infobip notes that “most retail technology stacks were not built for that kind of real-time synchronisation,” with integration work frequently taking two to three times longer than configuring the AI itself.

The result is that AI assistants act as a stress test for the entire retail stack. Weaknesses that were previously hidden within search and navigation become visible the moment a customer asks a direct question and expects a precise answer.

The implication is difficult to avoid. Success in conversational commerce depends less on deploying advanced AI, and more on building the infrastructure that allows it to operate reliably. Without that foundation, performance gains are unlikely to hold.



Trust, control, and customer behaviour

While the technology is advancing quickly, consumer adoption remains conditional.

At Kingfisher, engagement with AI assistants has grown steadily, with more than 60 per cent year-on-year growth in usage. That suggests a clear appetite for more guided, conversational shopping experiences. More broadly, consumer research from Attest indicates that 54 per cent of shoppers would engage with an AI chatbot on a retailer’s website.

However, willingness to try does not automatically translate into sustained use. Trust remains the defining variable.

Lisa Smith at Kearney is direct on this point. “Consumers are sharp,” she says. “They’ll forgive an assistant for not knowing something. They won’t forgive it for confidently getting it wrong.” Inaccurate recommendations, outdated stock information, or generic responses can quickly erode confidence, often after a single interaction.

This fragility is reflected in consumer sentiment data. While many shoppers recognise the potential benefits of AI in improving personalisation and experience, trust in how retailers handle data remains low, with a significant proportion expressing scepticism or outright distrust.

Design plays a critical role in closing that gap. Kim Johal argues that trust is not purely a technology issue, but a question of how and where the assistant is deployed. Retailers seeing stronger engagement are introducing AI gradually, with clear opt-ins, transparent disclosure, and visible paths to human support.

There is also a behavioural shift at play. Conversational interfaces ask customers to relinquish a degree of control, moving from browsing and filtering to accepting guided recommendations. Catherine Frame, director of retail solutions at UiPath, notes that this is “a significant psychological shift”, and one that must be handled carefully.

The assistants that succeed tend to follow a consistent pattern. They are grounded in reliable data, integrated into the journey at the right moment, and transparent about their limitations. When those conditions are met, they can build confidence over time. When they are not, adoption stalls quickly.

The result is a delicate balance. Consumer appetite for conversational commerce is clearly emerging, but it is contingent on consistent, accurate, and trustworthy experiences.



The struggle of scale

If early results point to genuine potential, scaling conversational AI remains the industry’s most persistent challenge.

At Kingfisher, progress has come from treating AI as a core capability rather than a discrete initiative. Assistants such as Hello Casto have been extended across banners including B&Q, supported by the group’s in-house platform, Athena, and integrated into a wider ecosystem spanning marketplaces, data, and retail media. The result is not a standalone tool, but a set of capabilities embedded across the business.

That approach stands in contrast to much of the wider market. Julian Skelly describes a pattern of “pilot purgatory”, where retailers believe they have scaled AI, but in reality, remain confined to limited deployments. The issue is rarely the technology itself. Instead, it is the complexity of moving from controlled environments to live, multi-market operations.

Alex Moseman highlights the gap between pilot and production. “Pilots usually rely on a small, clean slice of data,” he says, while real-world deployment requires integration with live catalogues, real-time stock, and constantly changing conditions. When those systems are not aligned, performance deteriorates quickly.

Organisational factors compound the problem. Nick Dutton points to unclear ownership and fragmented operating models as common barriers. “Stalled programmes often treat AI as a fragmented, departmental experiment, rather than a consistent, centralised operation,” he says. James Heimers echoes this view, noting that siloed teams and disconnected data prevent AI from being embedded into core processes.

Scaling across markets introduces further complexity. Variations in language, regulation, product availability, and customer behaviour all require adaptation. Kim Johal argues that many organisations underestimate this challenge, particularly when moving beyond a single geography or channel.

The programmes that succeed tend to share a consistent set of characteristics: centralised data models, clear ownership, cross-functional coordination, and a willingness to treat AI as part of the operating model rather than an experimental feature.

The broader lesson is straightforward. Building a working assistant is relatively easy. Embedding it into the fabric of a retail organisation – and making it a compelling proposition for customers – is significantly harder.


Conclusion – ROI beyond conversion

Conversion uplift may dominate the headlines, but it captures only part of the value equation.

For retailers such as Kingfisher, the impact of AI assistants extends across the customer journey. Increased engagement, improved product discovery, and more guided interactions all contribute to performance at scale. As Kingfisher’s chief AI officer Mohsen Ghasempour notes, “AI is the tool, not the mission – the key consideration is ensuring any service delivers real, tangible improvements to the customer experience and makes their lives easier.”

In some cases, the benefit lies as much in shaping behaviour over time as in driving immediate transactions. James Heimers argues that focusing solely on conversion risks understating the opportunity. “In many cases the real value will come from increased frequency, loyalty and share of wallet,” he says, particularly in categories where AI supports ongoing engagement rather than one-off purchases.

There are also operational gains that sit outside headline metrics. More effective product discovery can reduce reliance on customer support, while better alignment between intent and outcome can improve satisfaction and reduce returns. These effects are harder to quantify, but materially relevant over time.

Against this sits a more complex cost base. Integration, data normalisation, ongoing optimisation, and governance all require sustained investment. Nick Dutton points out that many of these costs are underestimated, particularly those associated with maintaining and improving the system once it is live. “Hidden costs may tend to lie more within the ongoing operating model,” he says.

Piyush Patel at Algolia reinforces this point, noting that the majority of effort typically sits in data and integration work rather than the AI model itself. Treating the assistant as a front-end layer risks missing the deeper operational commitment required to support it.

The result is a more nuanced ROI equation. Conversion uplift provides an immediate signal, but long-term value depends on how effectively AI is embedded into the broader retail operating model. Retailers that treat it as infrastructure, aligned to measurable commercial outcomes, are more likely to sustain returns. Those that focus narrowly on early uplift risk discovering that the hardest work begins after the pilot ends.

---