Why AI in Customer Experience Goes Beyond Chatbots

For years, the ambition for AI in customer service has been simple: reduce contact with the customer as much as possible.

Give customers a chatbot, automate repetitive questions and deflect as many conversations as possible before they reach an agent. There is nothing wrong with this method, a customer asking where their parcel is probably doesn’t need to speak to a person to get their answer. But this is only one part of the economics of customer service.  

What happens if the query does reach an agent? How quickly can they understand the problem? How many systems do they need to search? How consistently are processes followed? Once the conversation ends, how do you know whether it was handled well?

These questions become particularly important in ecommerce, where operational complexity tends to increase at the same time as customer contact. This is where the more interesting application of AI begins. The opportunity isn’t simply to put AI between the customer and customer service. It’s to use it across the whole operation.

The chatbot isn’t the strategy

The first generation of customer-service automation was largely designed around deflection. That created a fairly obvious incentive: maximise the number of conversations that never reach an agent.

But containment and resolution aren’t necessarily the same thing. A bot that prevents a customer reaching an agent can produce a good automation metric while creating a poor customer experience. Even an effective virtual assistant only addresses the first stage of the journey.

Once a complex query reaches a person, many traditional problems of contact-centre operations remain. Agents need to establish context, they search knowledge bases and processes, move between systems, make judgment calls. Team leaders provide support, conversations are summarised and quality teams review a sample of interactions after the event.

AI can affect all of those things. That’s why we’ve approached the problem in three layers: before, during and after the customer connects with an agent.

Before the connection: resolve what doesn’t need a human

The first job of automation should be straightforward: don’t make customers wait for answers that technology can provide immediately. This is a situation in which a virtual assistant can handle customer queries, including order tracking, returns and other transactional requests.  

Alongside that sits AI-powered routing and customer profile building. When a conversation needs an agent instead of a virtual assistant, the aim isn’t simply to put it into the next available queue. The system can establish what it is that the customer needs and provide the context of the conversation before the agent takes over.  

This matters. Good automation shouldn’t make the customer choose between a bot and a human. It should make the transition between the two almost invisible.  

The impact of this layer is significant. Between July and December 2025, the proportion of live-chat customer queries without an agent intervention increased from around 22% to more than 50%. This happened during the period when retail contact centres are usually under the most pressure due to peak retail season.

However, it is what happens next that we’re most interested in.

During the connection: remove the search, not the human

Anyone who has spent time around a large contact centre knows how much of an agent's job isn’t actually talking to customers. It’s searching: which policy applies to this order? What is the process for this country? What happened earlier in the customer journey? Is this customer eligible for a refund? Where is the latest version of that procedure?

The information exists, but finding it quickly can be an issue.

The second layer of AI therefore works alongside the agent rather than replacing them. Orbit Co-Pilot provides real-time process guidance and recommended responses as the conversations are happening. Virtual team-leader capabilities and AI agent assistance give agents access to the information they would otherwise have to find manually.

This means that the agent remains responsible for the conversation, but the technology removes much of the search around it. Since introducing these capabilities, average handle time has fallen by around two minutes per interaction.  

At contact-centre scale, two minutes changes the economics considerably. It also changes the job.

Agents can spend less time navigating systems and more time understanding and resolving the customer's problem.

After the connection: the quality problem nobody talks about

There is another constraint in traditional contact centres that automation fundamentally changes. This is quality assurance.  

Historically, QA has been a sampling exercise. A team leader or quality team can only manually review a fraction of the thousands of conversations happening across an operation. That means coaching, compliance and process improvement are all based on a small proportion of the actual customer experience.

This is where AI steps in. Conversations can be automatically summarised, scored and fed into AI-supported quality coaching and reporting.  

Instead of asking what happened in a sample of interactions, we can score 100% of conversations, creating a different kind of feedback loop where patterns become visible earlier. Training needs can be identified across teams rather than as a result of individual examples that aren’t at all representative of the organisation as a whole. Strong (and weak) performance is easier to recognise, and customer conversations become an operational dataset that informs how processes themselves need to change.  

This post-connection layer is just an important to us as the chatbot that customers see.

Peak exposes the different between automation and infrastructure

Peak season is where this model becomes particularly interesting.  

Traditionally, the response to the rising contact volume during peak trading periods is largely linear. More orders create more queries, more queries require more agents, more agents require more recruitment, training and management capacity.

During Peak season, all these pressures come at once.  

Black Friday doesn’t just test warehouse throughput, it tests how quickly a business can respond when thousands of customers simultaneously want to know where an order is, whether an address can be changes or what happens if something needs to be returned.  

Historically, retailers have dealt with that by adding temporary capacity, but this introduces its own problems. New agents need to learn processes quickly, while those processes are under the most strain. This means consistency becomes harder just when contact volumes are highest.  

AI can change the operating model here. If virtual assistants absorb a larger proportion of requests, fewer agents are needed in the first instance. And with the guidance that AI can provide during the remaining conversations, agents don’t need to memorise every exception to be productive. If every interaction can then be analysed afterwards, quality doesn’t have to become less visible simply because volume has increased.  

Peak stops being purely a workforce-scaling problem, it becomes a systems-design problem and that’s a much more consequential use of AI than adding another chatbot to a  website.

The test isn’t automation. It's outcomes.

There is an understandable concern that increasing automation comes at the expense of customer experience. The data gives us a useful way to test that assumption.

As our live-chat containment increased from roughly 22% in July 2025 to more than 50% in December, customer satisfaction remained around the 90% mark. At the same time, average handle time reduced by approximately two minutes and 100% of conversations became available for automated quality scoring.

But there is another result worth paying attention to.

Employee Net Promoter Score across our contact centres increased from 74 to 83 year on year. That matters because the debate about AI in customer service is often framed as technology versus people. We think that's the wrong framing.

Some of the best opportunities for AI are in the work agents shouldn't have to do in the first place: searching fragmented information, repeatedly summarising conversations, navigating processes and performing routine administration. Removing that work doesn't remove the need for people, it gives people more time for the interactions where judgement, empathy and problem-solving actually matter.

The next phase of CX won’t be containment alone

AI will resolve more customer queries without human intervention, that's inevitable but containment alone is a fairly limited ambition. The bigger opportunity is to rethink the architecture around a customer conversation: what can be resolved before it begins, what information an agent has while it's happening, and what an organisation can learn once it's finished.

For fulfilment operations, that becomes especially important as volumes rise during Peak periods.  The use of AI to break the traditional relationship between growth, contact volume and operational complexity is where businesses will see the most success.

That's the shift we're interested in at THG Fulfil: not AI instead of people, but AI is use across the operation, so people can be better at the parts of customer experience that still need them.