AI and Call Centres: Is the End Really Nigh?

From Simple Bots to Sophisticated Agents

The days of basic, rule-based chatbots are numbered. Today's conversation centres on "AI agents": sophisticated systems designed to make autonomous decisions and handle complex interactions. These are not the scripted bots of five years ago that could only parrot preset responses.

Yet implementation remains messy. Consider the well-documented mishap at DPD, which had to pull its AI chatbot after it began criticising the company and swearing at users. Meanwhile, Evri's chatbot Ezra confidently presented photo "evidence" of a delivered parcel that was clearly on someone else's doorstep, with no escalation path when challenged. Both companies are British household names, and both episodes made national headlines. That is not a soft launch.

Emily Potosky, analyst at Gartner, put the core tension plainly: "You can have a much more natural conversation with AI, but the downside is the chatbot could hallucinate, it could give you out-of-date information, or tell you completely the wrong thing." The challenge lies in finding the sweet spot between natural conversation and reliable service delivery, and European companies are learning this balance through expensive trial and error.

The Training Data Advantage, and Its Limits

Salesforce Chief Digital Officer Joe Inzerillo has argued that established call-centre operations represent goldmines for AI training. These facilities often possess extensive documentation and refined process libraries that AI systems can learn from effectively. The company's AgentForce platform, used by Formula 1, Prudential, OpenTable, and Reddit, learned valuable lessons during development. Initially, the AI would coldly "open a ticket" where a human agent might say "sorry to hear that". Salesforce had to explicitly train the system to show empathy when customers expressed distress.

That is a telling anecdote. Empathy is not a default state for a large language model; it is a trained behaviour, and training is neither cheap nor a one-time activity. Maintaining that quality at scale, across multiple European languages and cultural contexts, adds layers of complexity that vendors rarely foreground in their sales pitches.

The gaps between AI promise and real-world delivery are considerable:

• Training data requirements are extensive and expensive to maintain on an ongoing basis
• Knowledge management becomes more critical, not less, when generative AI is involved
• Hallucinations and outdated information pose continuous operational risks
• Complex emotional scenarios still require human empathy and judgement
• Customer preference varies significantly by situation, age group, and cultural context

The Human Element Is Not Going Away

Companies such as QStory, which works with eBay and NatWest, maintain that human interaction remains irreplaceable for complex scenarios. Mortgage applications, debt counselling, and emotionally charged complaints require a nuanced understanding that current AI systems cannot reliably match. Fiona Coleman of QStory has noted: "There are times where I don't want to have a digital engagement, and I want to speak to a human." That sentiment is not nostalgia; it is a commercial reality that any operator ignores at their peril.

Legislative pushback is already taking shape in Europe. Gartner predicts the EU could enshrine a formal "right to talk to a human" in consumer protection rules by 2028. That is not a fringe prediction. The EU AI Act, which entered into force in August 2024, already requires transparency when consumers interact with AI systems. Ursula von der Leyen's European Commission has been explicit that consumer trust must underpin AI adoption; any regulatory tightening on customer service automation would sit entirely within that political logic.

Andrea Renda, senior research fellow at the Centre for European Policy Studies (CEPS) in Brussels and one of Europe's most closely watched AI policy analysts, has consistently argued that the EU's regulatory posture will push organisations toward human oversight rather than full automation, particularly in high-stakes consumer interactions. His view, expressed in multiple CEPS publications, is that hybrid models are not merely an ethical preference but an increasingly legal necessity under the EU's emerging governance framework.

The Cost Reality Check

Contrary to the dominant narrative, AI is not necessarily the cheaper option over the long term. Gartner's Potosky warns that generative AI is "very expensive technology" requiring substantial ongoing investment in data organisation and system maintenance. The initial appeal of cost reduction frequently gives way to complex integration challenges and hidden operational costs that do not appear on the original business case.

Salesforce claims 100 million US dollars in customer service cost savings through its AI deployments, but clarifies that this did not mean job losses. Instead, staff moved into other customer service areas, suggesting augmentation rather than replacement. That pattern should be the baseline expectation for European operators, not the exception.

The numbers that do exist in the public domain illustrate the gap between aspiration and reality. While 94 per cent of customers report choosing AI-assisted channels when they are available, only around 20 per cent of AI customer service implementations are assessed as meeting original expectations. That is a striking failure rate for technology being marketed as transformative.

What European Operators Should Actually Do

The call-centre apocalypse narrative misses the mark. AI will certainly reshape customer service operations across the UK, Germany, France, and the broader EU, but the complete elimination of human agents is both technically unfeasible and commercially unwise in the near term. The organisations that are succeeding are not replacing people with machines; they are building hybrid models where AI handles high-volume, routine queries and humans manage complex, emotionally charged interactions.

Straightforward, factual queries with limited answer sets perform well under automation: parcel tracking, account balance inquiries, basic troubleshooting. Complex problem-solving, escalations involving distress, and regulated interactions such as mortgage or debt advice remain firmly in human territory, and European regulation is likely to keep them there.

The real opportunity lies not in choosing between bots and people, but in orchestrating them together with precision. European businesses that recognise this balance, invest in proper training data governance, and plan for regulatory compliance from the outset will be the ones that dominate the next phase of customer service evolution. Those chasing a purely automated cost play are setting themselves up for a very public correction.