AI in Communication-Intensive Roles: What the Research Says

When it comes to AI in communication-intensive roles, the technology question is largely settled. The harder questions are about credibility, customer preference, and where the line between helpful automation and damaging substitution actually lies. This article summarises our research findings on AI in customer service and psychology.

The Task-Based Framework for Thinking About Automation

Economic research on automation has long worked with a task-based model: technology tends to reduce demand for routine, middle-wage jobs while increasing demand for non-routine work at both ends of the wage distribution. In Sweden, this pattern has been statistically significant since at least the 1970s.

When work is broken down by task type — abstract, routine, and communication-related — communication-related tasks stand out clearly. They appear at both wage tails and have been growing in importance. This creates a direct question for AI: if communication-related tasks are both high-value and increasingly prominent, what is actually automatable within them?

The answer requires more than a technical assessment. It requires a framework for thinking about credibility — and that is where a much older body of thought becomes relevant.

The Credibility Problem: Aristotle's Ethos in an AI Context

Aristotle identified three forms of persuasive argument: Logos (facts, statistics, examples), Pathos (emotional resonance), and Ethos (the speaker's credibility and trustworthiness). AI handles Logos well — it retrieves and presents information with precision. Pathos can be approximated through voice design and persona. But Ethos presents a genuine challenge.

How does an AI establish credibility in a communication where the human on the other side knows, or suspects, they are not talking to a person? This connects directly to what philosopher Daniel Dennett called "fake people" — AI systems trained on human thought to such a degree that the distinction becomes meaningless in practice.

Dennett's position is not that such experiments should be prohibited — but that disclosure is essential when people communicate with AI in everyday contexts. The moment a person discovers they have been misled about whether they were talking to a human, the credibility damage is disproportionate to whatever was gained by concealment.

AI as a Therapist: The Evidence

The question of AI in psychology has a longer history than most people realise. Joseph Weizenbaum's Eliza, created in the 1960s, was designed explicitly to demonstrate the limitations of human-chatbot communication. He was startled when it succeeded far beyond his expectations — including with his own colleagues, who began forming what they described as genuine emotional connections with the system.

More recent research, including Per Carlbring's work on AI and psychotherapy, paints a nuanced picture. In one study, AI responses to patients were rated — on average — as more practical, more professional, and even more authentic than human therapist responses. But when participants were told they were communicating with AI, many still preferred human contact, despite the AI performing well.

The reasons are instructive. An AI-based conversational agent cannot interpret body language, cannot develop a genuine therapeutic emotional relationship, and cannot share lived experience. For certain situations — patients who find it difficult to open up to humans, educational simulations for training psychologists, 24/7 support between sessions — AI offers real value. For others, the absence of genuine human presence remains a significant limitation that better technology alone will not resolve.

AI in Customer Service: What Customers Actually Want

The picture in customer service is cleaner. Studies using the UTAUT model (Unified Theory of Acceptance and Use of Technology) show meaningful differences in how customers relate to AI versus human representatives in e-commerce contexts — differences that vary by age, prior experience, and the nature of the interaction.

Shorter, more transactional interactions — order status, return policies, account questions, product information — are strong candidates for AI handling. The interactions are well-defined, the information is structured, and speed and accuracy matter more than emotional resonance. For these, customers who use AI and have a good experience often become advocates rather than detractors.

Complex or emotionally charged interactions are different. Customer preferences for human contact are stronger when stakes feel higher, when the issue is ambiguous, or when the customer is already frustrated. Organisations that remove the option of human contact from these situations — to optimise deflection rates — consistently damage the customer relationships they were trying to protect.

Key Findings and Implications

The research yields three conclusions that should inform any AI deployment in communication-intensive roles:

• Transparency is non-negotiable. Customers and patients must know when they are interacting with AI. Concealment, however well-intentioned, destroys more trust than it saves.
• Choice must be real. Offering a human option that is buried, discouraged, or absent does not meet the standard. The option must be genuinely accessible.
• Task type is the primary criterion. Routine, well-defined, lower-stakes interactions are excellent candidates for AI. Interactions requiring emotional depth, lived experience, or genuine relationship are not — regardless of how convincing the AI's outputs may be.

The article also identifies a gap in the research: there is a need for further studies based on the TAM3 model (Technology Acceptance Model 3) to investigate both customer service specialists' and customers' willingness to adopt AI tools — particularly in contexts where the boundaries between routine and communication-intensive tasks are blurred.