Empathy and Tone in AI Customer Service: Where Bots Fall Short

By Linh Ngo · Updated Oct 11, 2026

Empathy in AI customer service is how caring, understood and respected a customer feels in a conversation with a bot, and the research says it has two halves: the words a model writes, and what the customer believes about who wrote them. The words are often good. Several studies find AI-written replies rated as warm, even warmer than human ones. The belief is the problem. Once customers know a machine is doing the caring, many discount it, and in some service situations the emotional language backfires.

This article is for support leaders and CX managers deciding how a bot should sound and when it should stop talking. It's part of the AI in customer service collection. We'll go through what the studies actually measured (and on whom), where customers want a person, and a set of tone guidelines you can put into a bot's instructions this week.

Key Facts

• In three vignette studies with 714 participants, customers reported lower satisfaction and less willingness to return after a chatbot interaction than after a human one, and that gap was fully explained by perceived empathy (Markovitch, Stough and Huang, Journal of Retailing and Consumer Services, 2024).
• When people were told a reply came from AI rather than a human, they felt less heard: 5.13 versus 5.81 on a 7-point scale (Yin, Jia and Wakslak, PNAS, 2024).
• Across nine studies with 6,282 participants, identical LLM-written empathic text was rated as more empathic and supportive when labelled human than when labelled AI (Rubin et al., Nature Human Behaviour, 2025).
• In a medical-forum study, evaluators rated 45.1% of chatbot answers empathetic or very empathetic versus 4.6% of physician answers (Ayers et al., JAMA Internal Medicine, 2023). That's health Q&A, not customer service.
• Half of consumers worry AI will stop them from reaching a person (Qualtrics XM Institute, 2026).
• Nearly 75% of consumers want to know whether they're talking to an AI agent, and 45% are more likely to use one when there's a clear path to a person (Salesforce, 2024).

What "Empathy" Means for a Bot

Researchers split empathy into parts, and the split is useful for support teams. Cognitive empathy is understanding what the person is feeling and why. Affective empathy is sharing the feeling. Motivational empathy is caring enough to act on it. A model can do the first well, because it's pattern reading, and it can write sentences that sound like the other two. What it can't do is feel anything, and customers know it.

For a support operation, it helps to translate that into three practical things a reply either does or doesn't do:

1. Names the actual problem. "You were charged twice and the refund hasn't shown up" beats "I understand your frustration."
2. Acknowledges the cost to the customer. Lost time, a missed deadline, an awkward conversation with their own boss.
3. Does something about it. A refund, an escalation, a concrete next step with an owner and a time.

The third item is where bots are limited by permissions, not by language. A bot that sounds caring but can't issue a refund is a bot that sounds caring and then fails.

What the Research Shows

Most of this literature uses short written exchanges, online participants and single interactions, not live support queues. Here's what each study did, so you can judge how far it travels.

AI text can read as more empathic than human text

The best-known result is from a JAMA Internal Medicine study. The authors took 195 public questions from a Reddit health forum, with answers from credentialed physicians, and generated a ChatGPT (GPT-3.5) reply to each. Three licensed health professionals, who were also coauthors, rated the pairs without knowing which was which. They preferred the chatbot in 78.6% of evaluations, and rated its answers higher on empathy (a mean of 3.65 versus 2.15 on a 5-point scale). One detail matters a lot: the chatbot answers were far longer, with a median of 211 words versus 52 for the physicians (Ayers et al., 2023).

So what can a support leader take from it? Not that bots are more empathic than your best agents. The physicians were answering strangers in a free public forum, probably in a few lines, and the raters were health professionals reading text. The safe reading is narrower: when a model is asked to write a careful, acknowledging reply, and a person is writing a quick one, the model's version can read as warmer. That's a point about effort and length, and it's a reason to use AI to help agents write better replies (see AI agent assist).

A USC study points the same way. Participants described a stressful situation and then read a response. In the main experiment, 540 US participants were recruited on Prolific, and 455 completed the final stage. Responses written by Bing Chat (GPT-4) made people feel more heard than responses written by other participants (5.74 versus 5.17 on a 7-point scale). The authors tied that to emotional support: the AI gave more of it (5.76 versus 4.64) and fewer practical suggestions (3.28 versus 4.31), and emotional support, not advice, was linked to feeling heard (Yin, Jia and Wakslak, 2024). The human responders were anonymous strangers and the replies were about 50 to 60 words, so this doesn't tell you how an experienced agent who knows the customer would compare.

The AI label takes some of it back

The same USC experiment crossed the true source with the label. The label effect was about as large as the writing effect, and in the opposite direction. A reply believed to be AI felt less heard than one believed to be human, whoever actually wrote it. The two effects added up almost evenly: AI text under an AI label (5.41) landed nearly where human text under a human label did (5.51), while AI text that people believed was human scored highest (6.12). The authors note there was no unlabelled condition, so they can't say whether an AI label lowers feelings or a human label raises them (Yin, Jia and Wakslak, 2024).

The Nature Human Behaviour paper pushes the same finding much harder. In every one of the studies, the researchers had LLMs write empathic responses to people's emotional situations and told participants either that a human or an AI wrote them. The text was the same. The "human" version was rated as more empathic, more supportive, and it produced more positive and fewer negative emotions. The gap held across different response lengths, delays and models. It was strongest for responses that stressed sharing the feeling and caring, and weaker for the purely understanding kind. People who suspected, unprompted, that AI had helped write a human-labelled reply rated it lower too. When they wanted emotional engagement, they chose a person (Rubin et al., 2025).

That has a clear implication for support. Part of what customers value in "I'm so sorry this happened" is that someone chose to say it. Take the someone away and the same sentence is worth less. This also means you shouldn't hide the bot to get the empathy score back: the same research suggests discovery is costly, and most consumers say they want to know anyway (the survey data is in the table below).

Customer-service studies: empathy explains the satisfaction gap

Two papers look directly at service settings. In the first, Markovitch and colleagues ran three vignette studies. Customers rated chatbot service lower than human service on satisfaction, repatronage, recommendation acceptance and willingness to recommend, whether the outcome was good or bad. Perceived empathy fully explained the effect. And there's a design lever: when the chatbot was written to communicate more empathically, evaluations rose to the human agent's level, while making it look more human didn't help. A fourth study, with 100 people using ChatGPT for medical advice, supported the same pattern (Markovitch, Stough and Huang, 2024). Note these are vignettes and short interactions, not repeat customers living with a bot for a year.

The second paper adds the warning. Han, Yin and Zhang ran three experiments in MIS Quarterly, one with a live LLM chatbot. Chatbot replies that acknowledged and mirrored a customer's negative feelings often triggered psychological reactance, the sense that a boundary or sense of control is being pushed on. That made the bot seem less competent and lowered perceived service quality and satisfaction. Empathy from a human agent stayed effective (University of South Florida, summarizing the study). The coverage doesn't give the sample sizes, so treat the size of the effect as unknown and the direction as a caution.

Put the two together and you get a more useful rule than "add empathy." Empathic wording lifts a bot when it's about the customer's situation. Emotional mirroring ("I can feel how upsetting this is") can backfire, especially after something has gone wrong.

Where Customers Want a Human

The empathy research explains why the survey data on human access looks the way it does. Customers aren't against AI doing the work. They're wary of AI standing between them and a person when the problem matters.

Qualtrics' own recommendation is that AI handle simple, transactional requests and then supply background and suggested solutions to human agents on complex ones. The surveys don't directly measure "empathy needs," so don't claim they do. But combine them with the lab results and a sensible routing rule falls out:

• Let the bot handle requests where the customer wants the answer, not the relationship: order status, password resets, policy questions, address changes.
• Hand to a person when the customer is upset about a failure, when money or a cancellation is at stake, when they ask for one, or when the bot has missed twice.
• Assist the person in those cases with a summary and suggested wording. The automate or assist page covers how to decide, and what customers expect from AI service goes deeper on the survey side.

Tone Guidelines for Your Bot

Here's a set of rules to write into the bot's instructions and review against. They're drawn from the findings above, plus common sense about short written replies.

What those examples share:

1. Say what happened, in the customer's terms. Specifics read as understanding. Generic sympathy reads as a script.
2. Skip the mirrored emotion. Don't tell a customer how they feel. Given the reactance finding, "I can see you're upset" is riskier from a bot than "That shouldn't have happened."
3. Pair every acknowledgment with an action or a next step. Empathy with nothing behind it is the thing customers complain about.
4. Keep it short. The medical-forum result was largely about long answers beating hasty ones. Long isn't the goal in chat. Complete is. Cover the problem, the fix and the next step, and stop.
5. Disclose, plainly and early. "I'm an AI assistant. I can help with orders and billing, or connect you to a person." Don't give the bot a human name and typing delays. The label research says discovery costs you.
6. Apologize for the event, not for the feelings. "Sorry the refund took this long" is accurate. "Sorry you feel that way" is a classic way to make things worse.
7. Match register to channel and customer. A terse reply fits a developer pasting an error. A warmer one fits a billing complaint, and account context and history tell the bot which customer it's talking to. Read the intent first with customer intent detection.
8. Don't promise what only a person can decide. Refunds outside policy, exceptions and goodwill credits belong to the human the bot hands over to.
9. Offer the exit once, clearly. One line, same place every time.

Step by Step: Setting Tone for Your Support Bot

1. Sort your top contact reasons by emotional load. Mark each as informational, frustrating or high-stakes. The first can be fully automated. The last should route to people.
2. Write the tone brief as examples, not adjectives. "Warm but professional" gives a model nothing to copy. Five good replies and five bad ones do.
3. Ban the scripted phrases. List the lines you never want ("I understand your frustration," "sorry you feel that way," "I apologize for any inconvenience") and test that they're gone.
4. Set escalation triggers. Anger markers, cancellation words, a second failed answer, a direct request for a person, and any money dispute above a threshold. Sentiment signals can help; see the sentiment analysis glossary entry for what they can and can't tell you.
5. Make the handoff carry context. The person should see the full transcript and account details. Nothing sinks a bot's empathy faster than the customer having to start over.
6. Test with real transcripts. Take 50 recent conversations, run them through the bot, and have two agents rate tone blind. Disagreements show where the brief is vague.
7. Score tone with a rubric, not a vibe. Observable checks work best ("named the specific problem," "gave a next step"). The AI quality assurance guide shows how to calibrate that scoring against human reviewers.
8. Measure customer-side outcomes. Re-contact within a week, handoff rate, time to reach a person, and a one-question satisfaction check after bot-only chats.

Common Pitfalls

• Treating one study as universal. A forum study of health questions says little about a refund dispute. Whenever you read an empathy statistic, ask who the participants were, what they were rating and whether it was a real interaction.
• Hiding that it's a bot to win empathy points. The label research says the gap is real, and customers say they want to know. Disclosure is the honest path and the cheaper one when discovered.
• Using emotional mirroring after a failure. This is where the reactance finding bites hardest.
• Confusing length with care. Long replies scored well in one study. In a chat window they just slow people down.
• Giving the bot empathy but not authority. A kind reply that ends in "I can't do that" is worse than a plain one that routes quickly.
• Training agents to sound like the bot. If the AI writes better acknowledgments than your team, use it as a drafting aid, not a template to copy word for word.
• Never retesting. Models, policies and customer mood change. Re-run the transcript test every quarter.