The Empathy Gap: Why You Shouldn't Use AI to Summarise Sensitive Customer Support Tickets

The Temptation of the Flawless Dashboard

Every customer experience lead shares the same recurring nightmare: a mountain of unread support tickets, escalating response times, and a team running on caffeine and sheer willpower.

When you are staring at thousands of angry, confused, or distressed messages, the promise of modern language models sounds like salvation. It is incredibly easy to pipe your support queue through an API—using a tool like those we write about on our /platforms/claude—and have it output a clean, prioritised, categorised dashboard.

Instead of reading through pages of human misery, you get a neat bulleted list: User frustrated by billing delay (Priority: High)* User unable to log in after update (Priority: Medium)*

It is clean. It is efficient. It also happens to be an empathy-killing disaster.

Using AI to summarise standard technical queries (like "how do I change my billing address?") is perfectly fine. But when we apply these same automated summaries to sensitive, emotionally charged, or highly personal support tickets, we create an "empathy gap" that can actively harm both our customers and our product's integrity.

The Sanitisation Problem: How LLMs Wash Away Human Pain

To understand why this is a massive ethical issue, we have to look at how LLMs process language. A large language model does not "feel" the desperation of a user who has just lost access to their primary business account. It predicts the next token in a sequence based on statistical probability. You can learn more about how these text structures are defined in our /glossary.

When you ask an LLM to summarise a long, emotional message, its natural training bias is to make the text concise, professional, and objective. It strips away what it perceives as "noise."

But in customer support, the noise is where the truth lives.

Consider this hypothetical customer support message: > "I've been trying to log in for three days. My business partner is hospitalised, I am trying to run this launch entirely on my own, and your system just keeps throwing an error code. I am literally crying at my desk right now. Please, just tell me how to fix this or I am going to lose my biggest client."

An LLM running a default summarisation prompt will likely turn that into: > "User experiencing login error code for three days during a high-stakes launch. Requesting troubleshooting steps."

Technically, the summary is accurate. It captured the core functional problem. But it completely erased the human emergency. The desperation, the panic, the absolute breaking point of a real human being has been sanitised into a sterile task.

If your support team only interacts with the summary, they will reply with a sterile, templated troubleshooting guide. The user, already at their wit's end, receives a robotic response that feels profoundly dismissive, destroying whatever brand loyalty was left.

The Ethical Danger of Delegating Judgement

When you outsource your reading comprehension to an AI, you are also outsourcing your moral judgement.

In sensitive contexts—such as reports of online harassment, accounts of financial hardship, or access issues tied to physical safety—the tone, vocabulary, and pacing of the message are vital diagnostic clues. A human support agent can read between the lines. They can spot the silent red flags that tell them a customer is not just annoyed, but in genuine distress or danger.

An LLM, operating entirely in vector space, cannot do this. It can easily miscategorise an abusive message as "assertive feedback," or downplay a critical security vulnerability because the user wrote it in a polite, non-threatening tone.

Furthermore, relying on summaries creates a dangerous psychological distance between your product team and your users. It is easy to ignore product bugs when they are presented as a percentage point on a chart. It is much harder to ignore them when you are reading the raw, unfiltered frustration of the people paying your bills.

The "Empathy First" Support Pipeline

How do we balance the operational need for scale with the ethical necessity of human empathy? The answer is not to ban LLMs from your support stack entirely, but to draw a hard line around where they are allowed to operate.

If you are setting up or troubleshooting your support flows using Anthropic's tools, you might want to consult Claude Support to understand their specific rate limits and context window handling. Once your tech is configured, apply these rules:

1. Separate the Functional from the Emotional Build a routing layer that detects emotional intensity. If a message contains high-stress language, bypass the summariser entirely. Route the raw, unedited text directly to a human agent who has been trained in empathetic communication.

2. Never Use Summaries as the Source of Truth If your agents use summaries to get a quick overview of a ticket, the raw message must always be displayed immediately below it, fully expanded. The summary should be treated as an optional index, never a replacement for the customer’s actual voice.

3. Track "Tone Shifts," Not Just Keywords Instead of asking your AI to summarise what the customer said, use sentiment analysis tools ethically to flag when a long-term customer's tone shifts from friendly to hostile over a series of tickets. This highlights systematic issues in your product without silencing the individual's voice.

Your support queue is not just a list of problems to be solved; it is the heartbeat of your user base. Treat it with the respect it deserves, and leave the summaries for the API documentation.