Customers can accept AI but still prefer people

Cost savings alone cannot decide where customer-service AI belongs

A customer-service task can be expensive to staff and easy to automate, yet still be a poor choice for AI. Customers may need specialized judgment, an understanding of their circumstances or a response that recognizes the emotional weight of their problem. A savings calculation cannot establish whether removing human involvement will preserve those qualities.

Yet the common deployment question, “what can we automate?”, encourages companies to start with technical capability. They identify work a system can perform, prioritize interactions with high volumes and staffing costs, then calculate the savings. This sequence assigns work before establishing what customers need or where AI and people have different strengths within an interaction.

A better allocation starts with those strengths and remains open to revision. Teams need to decide which interactions AI handles, which require people and how customers move between them. Task fit, clear disclosure, customer trust and access to humans should shape those decisions alongside cost, so deployment choices account for customers who need judgment or doubt that automated service will understand and resolve their issue.

Test human-AI combinations against the strongest alternative

The first allocation decision is whether combining people and AI improves the work. A team at the MIT Center for Collective Intelligence reviewed more than 100 studies of human-AI collaboration, covering 370 measured outcomes across 106 experiments. On average, combined systems failed to outperform the stronger of the human-only and AI-only alternatives.

“This was our most surprising finding,” said Thomas Malone, an MIT Sloan professor and director of the MIT Center for Collective Intelligence. Some combinations performed worse; Malone described the underperformance as “statistically significantly worse.” The results challenge the assumption that putting people and AI together automatically preserves the strengths of both.

For service leaders, this finding changes the benchmark for a proposed workflow. A combination may improve on an existing process while still performing worse than the strongest standalone option. Testing against that stronger alternative shows whether the added coordination and human involvement improve results enough to justify them.

The review also identifies where human participation helps: human-AI teams outperformed AI alone on tasks requiring specialized judgment, context and emotional intelligence. Beating AI alone is a narrower result than beating the best standalone system, but it shows where people can improve an AI-led process. Michelle Vaccaro at MIT describes the difference as AI’s strength in repetitive, data-driven tasks and people’s strength in context and emotional intelligence.

These differences give teams a reason to examine how work is organized when a combination underperforms. Simply adding AI or a person leaves responsibilities unresolved; each contribution needs to match the task’s demands. The research covers collaboration across tasks, so applying it to customer service requires testing the proposed workflow and asking a separate question: whether customers accept it.

Customer acceptance needs its own evidence

Customer surveys show why successful task completion cannot stand in for acceptance. In Metrigy’s Customer Experience Optimization 2025-26 consumer study, 84.7% of participants preferred a human to an AI agent. When assured that their issue would be resolved, 80.1% still preferred a person. That assurance was a survey condition, so the result measures stated preference under that condition rather than reactions to personally experienced AI resolution.

Metrigy has a commercial interest in demand for customer-experience research, which matters when using its findings to guide deployment decisions. Its study also found that 46% of participants were willing to use AI agents for particular purposes. This selective willingness can coexist with a general preference for people because accepting automation for a defined task differs from choosing a preferred form of support overall.

Waiting time also leaves much of the preference for people intact. A 2025 HubSpot/SurveyMonkey survey of roughly 15,000 consumers across seven markets found that 82% preferred human support when waiting times were identical, while about a quarter said they liked AI in customer-service interactions. HubSpot and SurveyMonkey have commercial interests in customer-service and survey tools, respectively. Equal waiting times remove a speed advantage from this comparison while leaving other aspects of the interaction open.

Together, these surveys establish that customers can prefer people while remaining willing to use AI selectively. They measure different conditions and should inform separate deployment questions: which tasks customers accept AI handling, and what makes them seek human help. Service teams should examine whether complexity, consequences and emotional demands change those answers among their own customers.

Reported choices at unnamed brands offer a different view of acceptance. In conversations with customer experience leaders at brands offering a genuine choice between AI and a person, customers were reported to have often selected AI because it was fast, accurate and understood them. These accounts describe behavior in particular services; the surveys describe stated preferences across larger groups. Each answers a different question.

The brand experiences suggest that capable AI can satisfy some customers’ desire for understanding and resolution. The broader preference for people still matters, so companies have a reason to test selective AI use while allowing for customers who will continue to choose human service.

Allocate work according to task demands

Selective use needs an initial allocation that teams can test. Start by identifying whether success depends mainly on repetition, volume and data processing, or on specialized judgment, contextual understanding and emotional intelligence. Then assign responsibility at each stage, including how people and AI work together when a request contains both kinds of demand.

AI’s comparative strengths make repetitive, high-volume and data-driven work reasonable candidates for automation. Pattern detection, forecasting and sorting illustrate these demands, while Metrigy’s accepted transactional uses show how clearly defined tasks appear in customer service: directing someone to the right person, confirming a shipment or order, and scheduling. Each gives the customer a relatively clear outcome to recognize.

A clear outcome makes a task a sensible starting point for evaluation, but customer acceptance and operational performance still need separate tests. A customer may welcome automated scheduling even when an implementation handles it poorly. Conversely, a system may complete the scheduling task successfully while a customer continues to prefer a person.

More consequential interactions require closer attention to how context changes a satisfactory response. A request can contain data-heavy steps while still needing a person to interpret circumstances or exercise judgment. Human involvement should remain where those demands materially affect the outcome, with AI assigned the portions it handles well.

This allocation is a starting hypothesis for the actual customer base. Distinguish transactional requests from complex, consequential or emotional ones, then examine acceptance and results within each group. AI’s comparative strengths and customers’ accepted transactional uses can point toward the same allocation, but each requires its own evidence before responsibility expands.

Disclose AI clearly and design access to people

Once responsibility is allocated, customers need to understand who is handling their request. Under Article 50 of the EU AI Act, people have been required to be informed when they are interacting with AI since August 2, 2026. Customer-service teams therefore need to identify relevant EU interactions involving a chatbot, virtual assistant or automated phone system and make the disclosure clear at the beginning.

Clear timing makes disclosure part of the service flow. A vague notice buried in terms and conditions does not meet the disclosure requirement described by Article 50, so teams need to decide how customers learn that AI is involved before proceeding. The notice belongs where customers encounter the system and can understand its role.

Understanding that role raises a separate question: how can customers reach a person? Article 50 establishes the disclosure obligation, while providing a clear human route is a service-design recommendation for customers who need or prefer human help. Keeping these responsibilities distinct helps teams assign ownership accurately and avoids presenting every desirable service feature as a legal requirement.

The two decisions can work together to support trust. Disclosure helps customers understand the interaction, and a visible route to a person gives them a way to act on their preference. These choices may increase confidence in automated service and willingness to use it later, although trust still depends on what happens during the interaction.

Use four signals to design responsibility and handoffs

Task allocation and interaction design need to be checked against what happens across the service. A practical framework, Four Signals, brings together Friction, Conditional Preference, Transparency and Frontline Knowledge. This is a working synthesis for service design rather than a validated predictive model; it helps teams compare different kinds of evidence before assigning responsibility.

Friction starts with the effort customers already expend across their journey. Map the confusing step, the moment someone abandons the process and the point at which they seek help, then investigate why each occurs. These observations expose problems that predate automation. Transferring a difficult step to a bot can preserve or compound the difficulty unless the underlying cause changes.

Once the difficulty is understood, Conditional Preference adds the customer’s judgment about who should handle it. Establish which tasks the actual customer population accepts AI performing and which prompt a preference for a person. General surveys provide context, while local evidence guides the allocation and prevents willingness to use AI for one purpose from becoming permission to automate every interaction.

Customer preferences are meaningful only when people understand the available choices, which is where Transparency enters the framework. Check whether the AI disclosure and human route described earlier are understandable at the point of use. Even with disclosure, an interaction can erode trust when customers cannot work out the next step or get the help they need.

Frontline Knowledge then tests the designed experience against operational reality. Customer Success, Support and field teams encounter broken automated paths, customer difficulties and failed attempts at resolution. Their knowledge is often unstructured, but it can expose problems before they appear in formal reports. Consulting these teams brings experience into allocation decisions that aggregate measures may conceal.

Comparing the four signals lets teams examine the same service decision from several directions. Friction identifies effort, preference identifies acceptance, transparency tests informed participation, and frontline knowledge exposes operational gaps. When these observations support the same allocation, the decision has several distinct grounds. When they conflict, teams have a specific question to investigate before changing responsibility.

Acting on those findings requires redesigning the full process around the allocation. Specify who owns each stage and each handoff, the transfer of an interaction between AI and a person or between people, so the customer keeps a clear route toward resolution. The same design must establish how customers reach human help and who takes responsibility when an automated path fails.

Ownership therefore extends beyond selecting technology. Licensing an AI model supplies a tool, while the service organization must assign responsibility for outcomes, failed paths and later changes to the allocation. The quality of those operating decisions matters more here than the power of the licensed system. Human participation needs a planned role in maintaining customer relationships, supported by clear ownership.

Revise responsibility as customers and results change

With ownership established, teams can use genuine customer choice to examine where AI has earned confidence. The unnamed-brand experiences illustrate why that choice can reveal acceptance that a general preference survey misses. Let customers experience AI’s speed, accuracy and ability to understand a request while making its role clear and keeping access to a person available.

Customer choice then needs to be considered alongside operational results before AI takes on more responsibility. This proposed method for selective expansion allows for customers who continue to prefer people even when successful resolution and waiting time are addressed. Their preference remains part of the operating decision because technical improvement does not necessarily remove it.

An accountable owner must also be able to reduce AI’s role when the evidence calls for it. That owner needs authority to retain human responsibility, restore it after an unsuccessful transfer or approve a further automated step when customer behavior and service results justify the change. Where trust has yet to develop, human involvement should remain available, and the process must make returning responsibility to people operationally possible.

Recap

The executive decision is not how much customer service AI can automate, but where it should hold responsibility. Cost savings belong in that decision, alongside task performance, customer acceptance and the judgment an interaction requires. None of these measures can substitute for the others.

Before expanding AI’s role, leaders should require evidence that the proposed workflow performs well against the strongest alternative. They should also establish whether customers accept it, whether people can reach human help and who owns resolution when an automated path fails. A successful transaction alone does not answer all those questions.

Treat each allocation as a decision that can be revised. Give an accountable owner the authority to expand AI’s role when results support it and restore human responsibility when they do not. The goal is a service operation that uses AI where it helps, preserves human judgment where it matters and remains answerable to the customer throughout.