Customer service: the real question is no longer how much to automate, but where to stop

Companies are seeking to automate their customer relationships more and more. Yet, the automation rate should not be an objective in itself.

As artificial intelligence absorbs simple requests, it concentrates complex situations in human hands. The real challenge is therefore no longer just knowing what technology can automate, but deciding where automation should stop.

This boundary is becoming a strategic question. Automating does not simply mean reducing a workload. As AI absorbs standardized interactions, it modifies the very composition of human work, now more exposed to exception, uncertainty, risk and relationship. The expected skills, organizational models and the way the company creates value are directly transformed.

Customer service thus becomes a particularly revealing laboratory of a much broader transformation of the company. What should humans still do when machines become capable of responding, recommending and now acting?

Automating the simple concentrates the complex

For a long time, contact centers were organized around volume. It was necessary to absorb calls, reduce waiting time, decrease average handling time and improve cost per interaction. The first generations of chatbots extended this logic by automating recurring requests based on predefined scenarios.

Generative AI changes the scale of the phenomenon. A system can now understand a request formulated in natural language, consult several sources of information, synthesize a history, contextualize a response and, when connected to the company’s systems, begin to execute certain actions.

But this evolution produces a less visible consequence. When predictable requests disappear from the human flow, those that remain mechanically become more difficult.

The advisor recovers the exceptions: complex dispute, potential fraud, sensitive financial situation, customer already frustrated after several attempts with an automated assistant.

Automating the simple concentrates the complex.

This transformation directly modifies the profile of necessary skills. We cannot automate standardized tasks while continuing to recruit, train and evaluate employees according to the same models.

The advisor will need more autonomy, analytical capabilities, contextual understanding and relational intelligence. They will also need to know how to exploit AI recommendations, recognize their limits and deviate from them when the situation requires it.

Automation therefore does not only eliminate part of human work. It progressively selects the work that remains for humans. And this work could become more demanding.

AI diffuses expertise. But who will produce new expertise?

The work of Erik Brynjolfsson, Danielle Li and Lindsey Raymond constitutes an important starting point.

By studying 5,172 customer support agents using a generative AI-based assistant, the researchers observe in the published version of their work an average productivity increase of 15%, measured by the number of problems solved per hour. The gains are significantly higher among the least experienced or initially less performing employees, while they are much more limited among the most experienced employees.

The interpretation is particularly interesting for the company. AI seems capable of diffusing part of the practices associated with the best employees to those with less experience. It can thus transform expertise that until now was largely individual into capacity available on a larger scale in the organization.

A field experiment conducted with Alibaba and made public in 2026 nevertheless nuances this observation. Generative assistance improves processing speed and certain quality indicators, with particularly significant gains among initially less performing employees. But the most performing agents benefit less from the tool and may even experience a deterioration in certain indicators.

The managerial consequence deserves to be highlighted: there is probably not a single relevant way to augment all employees with AI.

A novice may need a relatively directive system, suggesting responses and procedures. An expert could on the contrary derive more value from a less prescriptive AI, used to search, verify, challenge a hypothesis or explore an alternative.

AI personalization could therefore not only concern the customer. It could also concern the employee.

But a more fundamental question appears.

If AI accelerates the transmission of existing expertise, how will the company continue to produce new expertise?

Diffusing expertise and producing expertise are two different functions. This question goes far beyond customer service. It potentially concerns engineers, consultants, lawyers, developers, analysts, finance or health professionals.

If junior employees perform fewer and fewer tasks that historically allowed them to acquire the fundamentals of a profession, how will they build the expertise that will make them tomorrow’s experts?

The challenge is therefore no longer just organizing AI adoption. It is also necessary to preserve the conditions for producing human reasoning, experimentation and judgment.

An AI that speaks is not an AI that acts

A second rupture is now visible.

The first generative AI applications in customer service mainly assisted conversation. They proposed a formulation, summarized a file, retrieved information or suggested a procedure.

Agentic systems change the nature of the problem. They can identify the customer’s intention, search for information in several systems, verify certain rules, trigger an action and then confirm its execution.

An AI that speaks is not an AI that acts.

As long as AI advises, humans remain the main decision point. When it acts, the company delegates part of its operational power to it.

The question is then no longer just that of the model’s performance. It becomes that of the delegation scope that the organization accepts to entrust to it.

The work published in 2026 around Nubank’s support systems illustrates this evolution. They describe five production deployments covering notably card delivery, debt management, credit limits, card management and product explanation. The determining point does not only lie in the model used, but in the entire evaluation, control and improvement system surrounding its deployment.

When a system moves from “here is what I recommend” to “I have performed the operation”, the consequences of an error change scale. They can become financial, contractual or operational.

How far can the agent act alone? From what threshold does human validation become mandatory? How to detect an unusual situation? How to trace decisions made? Who intervenes when the system hesitates? Who assumes responsibility in case of error? The transition from chatbot to agent thus transforms customer service into a governance subject.

In Europe, this governance also has a regulatory dimension. Since August 2, 2026, the transparency obligations provided for in Article 50 of the AI Act apply to certain categories of AI systems. For systems intended to interact directly with natural persons, they must notably be informed that they are interacting with an AI, except when this situation is obvious to a reasonably informed, attentive and prudent person. Transparency therefore also becomes an element of the design of certain automated journeys.

Klarna: reducing costs is not optimizing the relationship

The Klarna case particularly illustrates the tension between economic efficiency and value creation.

The Swedish fintech has become one of the symbols of customer service automation. After heavily using AI in a logic of productivity gains and cost reduction, Klarna has progressively rebalanced its strategy. In September 2025, Sebastian Siemiatkowski acknowledged that the company had probably gone too far in this direction and indicated wanting to use AI more to improve services and products.

This movement does not mean that Klarna is abandoning AI. This is precisely what makes the case interesting.

It does not tell the story of automation failure. It shows the limits of a strategy that would measure its success essentially by the reduction of mobilized resources.

An interaction can cost less without producing more value.

Optimizing the processing cost is not equivalent to optimizing the customer relationship.

The automation rate can therefore become a misleading indicator if it is dissociated from resolution quality, trust and created value.

Where to place the boundary? The EntrepreneurIA decision grid

For leaders, these transformations can be translated into a simple decision grid around three configurations: autonomous AI, augmented human, expert human.

This is not a validated scientific taxonomy, but a strategic grid proposed by EntrepreneurIA to help organizations position their interactions.

Autonomous AI is suitable for repetitive, documented requests presenting low risk: retrieving information, verifying a status or explaining a simple procedure.

The augmented human becomes relevant when AI can prepare the decision without having to assume it alone. It searches for information, synthesizes history, proposes a procedure or identifies certain risks. The employee retains judgment and responsibility.

The expert human intervenes when uncertainty, risk or relational value become high: negotiation, dispute, fraud, financial vulnerability, emotional situation or strategic customer.

Three variables then structure the decision: Complexity. Risk. Strategic value.

A low-complexity operation is not necessarily a good candidate for automation if the potential cost of an error is high.

Conversely, a request that AI is technically capable of handling can remain human for commercial or relational reasons.

A customer wishing to renew a major contract, increase their investment or access a premium service may formulate a perfectly automatable request. The company can nevertheless decide that this interaction deserves a human.

The automation boundary does not only depend on what technology can do. It also depends on what the company considers strategic.

 

This grid invites each organization to ask three questions:

• What can we automate?
• Which interactions require an augmented human?
• What do we want to preserve as human interaction?

The customer does not necessarily want to choose between AI and human

The transformation of customer service must also be observed from the other side of the interaction.

The 2025 Customer Service Observatory by Ipsos bva, conducted among 5,000 people in five European countries, reveals a particularly strong preference for humans in France. According to the study, 90% of French respondents prefer to wait longer to be connected with a human advisor rather than a virtual advisor. 72% state that they would feel deceived if a company did not inform them that they are speaking to an AI.

These figures are significant. They do not necessarily mean, however, that consumers reject any interaction with AI.

Research published in 2026 in the Journal of Retailing and Consumer Services brings an interesting nuance. Through four studies, its authors observe that consumers prefer AI more when they communicate objective information, while they favor humans more when the information is subjective.

The question is therefore probably not: do customers prefer humans or AI? It is rather: for which interaction, in which context and with what level of stakes do they prefer one or the other?

There is probably no absolute preference for humans or for AI. There are situations in which one or the other produces more value.

In many cases, the customer mainly wants their request to be understood and resolved quickly, correctly and without unnecessary disruption.

A good experience is first and foremost an experience in which the customer does not suffer from the company’s technological choices.

If an AI correctly and immediately resolves a simple request, human intervention does not necessarily bring additional value. But when the system no longer understands, when risk increases or when the situation becomes sensitive, continuing automation can deteriorate the experience.

A performing AI is therefore not only an AI capable of responding. It is also an AI capable of recognizing the moment when it should no longer respond.

The quality of the transition from machine to human then becomes as important as the capabilities of the machine itself.

Could access to humans become a privilege?

Automation finally opens a more disturbing question. The more numerous automated interactions become, the more access to a truly competent human interlocutor can acquire value.

In 2025, Sebastian Siemiatkowski himself mentioned the possibility that human customer service could become a form of “VIP” experience, in an environment where automated service would be available on a large scale.

This perspective deserves to be taken seriously. In certain sectors, the quality of human interaction already constitutes an important part of the value proposition. This is the case for private banking, luxury, certain health services, B2B consulting or high-end travel.

In an environment where automated conversation becomes abundant, human attention can become rarer. And what becomes rare can gain value.

But this evolution carries a risk: human intervention could progressively be concentrated on certain segments considered strategic or profitable, while others would be primarily directed to automated interfaces.

We could then see a two-speed customer relationship emerge: a rich human experience for some and a mainly automated experience for others.

The challenge would no longer only be knowing whether AI improves or degrades customer service. It would also be necessary to ask for which customers it improves the experience and for which customers it progressively replaces access to humans.

The question becomes economic, strategic, but also ethical.

After twenty years of digitalization of customer service, could access to a competent human become a privilege?

Measure automation or measure relevance?

AI’s new capabilities finally force organizations to review their indicators. Traditional measures remain useful, but they are no longer sufficient.

Three families of KPIs can be distinguished.

• Efficiency: cost per interaction, processing time, waiting time, automation rate.
• Quality: first contact resolution, repeat requests, satisfaction, abandonment, loyalty.
• Human-AI orchestration: transfer rate to an advisor, unnecessary transfers, too-late transfers, human corrections, non-escalated errors and quality of continuity between AI and human.

Measuring only the automation rate amounts to measuring how much work has been entrusted to the machine. This does not say whether this distribution was relevant.

An still exploratory indicator, and undoubtedly difficult to construct robustly, nevertheless deserves to be studied: how many situations should AI have transferred to a human but did not? Because the major risk may not only lie in what AI cannot do. It lies in its inability to recognize that it cannot do it.

Conclusion: the company must decide where automation stops

Artificial intelligence creates a new automation capability. But it does not alone determine how this capability should be used.

Recent work shows that it would be reductive to simply oppose replacement and augmentation. AI can greatly improve the performance of some employees and much less that of others. It can handle certain requests alone, prepare human decision in others and now begin to act directly on the company’s systems.

The central question therefore becomes that of the distribution of roles between AI and humans. This distribution must take into account the complexity, risk and strategic value of each interaction.

Value creation will therefore not depend on the highest automation rate, but on the organization’s ability to position AI in the right place, strengthen human skills and preserve human intervention when it produces more value than automation.

The question is no longer knowing how many interactions a company can automate. It is determining which ones it absolutely should not automate.

Sources and references

Brynjolfsson, E., Li, D. & Raymond, L. R. (2025), “Generative AI at Work”, The Quarterly Journal of Economics, vol. 140, no 2, pp. 889-942. DOI: 10.1093/qje/qjae044.

Ni, X., Wang, Y., Feng, T., Lu, L. X., Wang, Y. & Zhou, C. (2026), Generative AI in Action: Field Experimental Evidence from Alibaba’s Customer Service Operations, arXiv:2603.29888. Preprint.

Gupta, A., Rossell, K., Alcobaça, E., Pacheco, J. C. L., de Lima, C. B., Tang, S., Rabachini, L. P., Moneda, L., Fei, H., Silva, D. & Ramanath, R. (2026), Building Customer Support AI Agents at 100M-User Scale: An Evaluation-Driven Framework.

Ding, Z., Zhang, Y., Sun, J., Goh, M. & Yang, Z. (2026), “Harmonizing Human Touch and AI Precision in Customer Service”, Journal of Service Research, vol. 29, no 3. DOI: 10.1177/10946705251384692.

Dai, X., Zhang, L., Huang, Z. & Wang, L. (2026), “AI or Human: How the Type of Information to Be Disclosed Alters Customer Service Agent Preferences”, Journal of Retailing and Consumer Services, vol. 89, Part B, article 104621. DOI: 10.1016/j.jretconser.2025.104621.

Ipsos bva / Élu Service Client de l’Année (2025), Customer Services Observatory 2025: humans (still) at the heart of customer relationships in the age of AI, European study among 5,000 people.

European Commission (2026), Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act, July 20, 2026.

Mukherjee, S. & Wang, E., Reuters (2025), “Sweden’s Klarna shifts AI focus from cost cuts to growth”, September 10, 2025.

Davis, D.-M., TechCrunch (2025), “Klarna CEO says company will use humans to offer VIP customer service”, June 4, 2025.