Chatbots are now dominant. To make them better, get your data in order

These big steps forward in conversational technology might not be obviously apparent if you had spent the last few weeks, as I have, playing around with dozens of UK retail customer service portals.

There are some notable exceptions and many strong experiences, but a lot of interactions with customer service chat involve being misunderstood, receiving lists of irrelevant help articles, or quickly being redirected to a live agent if queries get a little bit complex.

A recent study from Nextiva found 90% of companies in the US, UK and Canada admitted that poor handovers from bots to humans cause frustration for their consumers. Last year, the Institute of Customer Service (ICS) reported an eight-year low in its UK Customer Satisfaction Index, citing chatbot doom loops as among the reasons for the dip.

It is quite clear that generative AI offers steps forward in not just talking back to consumers, but in simply understanding what they are contacting your brand about in the first place.

Some retailers are already seeing possibilities for it enhancing the customer experience through chat-based product advice, recommendations and even transactions that take place wholly within a chat window.

Yet it is not a magic wand. Vendors say that successful chat experiences typically require work on the back-end to get right. That might be building a more exhaustive help database from which generative AI assistants can draw on to guide their answers, or, in those more advanced cases, connecting up databases to add additional functionalities to the chat experience.

“We talk about it…being the chatbot,” says ICS chief executive Jo Causon, “It’s actually for me the data that is going to be the critical thing – how we collect data, how we structure that data, and how we use that data in order to be able to deliver a better customer experience.”

Chatbots are almost everywhere

The majority (57%) of a sample of 100 leading UK retailers now use chatbots as a first-wave customer service option, according to data gathered by Retail Week. A further 18% use live chat, making chat among the dominant customer service modes in the UK, with a combined total of 75% just behind email and webforms (79%), but ahead of phone (70%).

There are many reasons why it has become so attractive to businesses. Firstly, it offers the ability to triage customers away from customer service agents. Once a human takes the reins of a chat, they can communicate with multiple consumers at the same time, whereas a phone call generally demands the full attention of a worker.

That means 30% of retailers no longer use phone calls as a customer service option at all. It tends to be value or ecommerce-oriented businesses that have become first to make this change. Most chat experiences will redirect to a human if the query cannot be resolved, but a tiny minority of brands will suggest writing an email if the chatbot does not have an answer for them.

Generative AI arrives in CX

Generative AI is clearly making its way into the customer service experience. Users of Zara’s website, for example, get to use a chatbot that can respond flexibly to queries, even if it is simply to say that it cannot help. The brand has a PDF disclaimer underneath the interaction saying that the brand does its best, but “cannot guarantee that answers are always accurate or complete”.

One of the technology operators helping brands with using generative AI is Zendesk, which has worked with UK retailers including Next and Lush. The latter reported a 60% first-contact resolution rate by integrating a custom AI assistant called Marvin.

“We believe that companies can automate up to 80% of their interactions through chatbots,” says Zendesk chief technology officer, EMEA, Matthias Goehler, adding that 30 to 40% redirection is a very good number for brands today.

Zendesk and other service providers such as IBM have recently experimented with a pay-per-resolution model, where a brand only pays once the consumer has been triaged away from a customer service operative.

Hallucinations and business case

Brands typically need to start by building their knowledge base. That is the first step in avoiding the problems that many experience when using AI models such as Chat-GPT, DeepSeek, or Claude – hallucinations.

IBM senior partner, consumer products and retail industries leader, EMEA, Elaine Parr pushes back on my use of the word “chatbot” to describe chat-based customer service interactions. “It’s really interesting that you went for chatbots, whereas I would be talking about assistants and agents” she says, adding that customer service chat has come a long way from rules-based interactions that often resulted in doom loops in the past.

“We could not put anything customer-facing that hallucinated,” she says, “and part of the challenge I think we’re facing in tech is the difference between consumer-grade generative AI and enterprise generative AI.”

Smaller language models that are specific to a use case such as answering customer queries may not require the processing power of large language models (LLMs) yet, but could have similar performance in understanding text in the activities they have been designed for. Walmart has flagged the development of these and domain-specific large language models as one of their expected AI trends for 2025.

For companies that do want to use LLMs, one of the protections they can put in place is so-called retrieval augmented generation. This is when an AI is told to stick to a knowledge base, whether a dataset or an information-rich troubleshooting page when answering customer queries.

This is not easy – a recent BBC study showed the difficulty that AI had in reproducing news stories even when told to stick tightly to the text – but, combined with the right disclaimers and warnings, some retailers are finding it provides guardrails enough for deployment.

“What’s exciting is how gen AI can help organisations be more proactive in their entire customer experience. With gen AI, organisations can better understand and identify nuanced customer needs and direct their customers to help – often to human representatives – also empowered with AI,” says Accenture managing director, Mark Farbrace.

The evolution of chat-based shopping

When chat-based experiences get really interesting, additional functionality is added beyond the ability to simply troubleshoot and triage customers. This is not necessarily a new thing (consumers started ordering pizza via Domino’s on Facebook Messenger in 2016). It is not uncommon for retailers to process returns entirely via chatbots.

In 2022, Walmart rolled out a Text-to-Shop feature that used natural language processing to allow consumers to text what they wanted to the Walmart app and have that added to their basket. The retailer has developed its own proprietary AI, which it is using to power a shopping assistant offering product advice.

Among the 100 leading retailers I looked at in the UK, Amazon and Zalando were the only ones with AI tools that had been hooked up to product databases to offer recommendations. The Zalando Assistant, now live in 25 markets, has been used by over 1.5 million customers so far, to field complex queries such as “what coat would work well for a Berlin winter?”.

Zalando reports that queries are about three times the length of search bar queries, suggesting heightened engagement, with conversations typically going back and forth between customers and agents around four times.

“It is not uncommon for retailers to process returns entirely via chatbots”

There is some evidence that consumers are already doing a little bit of this themselves. Adobe reported a fast-growing trend of using generative AI to source ideas for Christmas presents in its latest festive spending report.

What really unites these forward-looking businesses is not just that they are pushing forward with generative AI, but they have done the hard graft of unifying databases with their chat portals.

LLMs are a big step forward, but many functionalities such as order processing could have been added in the past using natural language processing and the correct safeguards to ensure that customers are getting what they want.

What comes next?

The advances keep on coming. In February, OpenAI announced the launch of its Operator agent which can complete tasks such as filling in forms and ordering groceries. The big hype this year is around Agentic AI and whether it will be able to autonomously start automating processes across databases and interactions.

As a result, many retailers will keep asking at what point is the technology ready to deploy. However, the right approach might be to look at your back-end to ensure that it is able to take the best advantage of the fast-evolving technology as it develops – AI cannot yet do this for you (though it may be able to help).

“This knowledge base needs to be good, it needs to be comprehensive, it needs to be up to date, and you need to maintain it,” says Zendesk’s Goehler, “Otherwise it will just create an answer which you don’t necessarily want.”