AI customer sentiment analysis works in four stages: it collects conversations, reads their emotion, separates it by aspect, and combines the results into a history per customer. That sequence decides whether your support team only knows that one message sounded angry, or knows the same customer has complained three times in two weeks.
Sentiment analysis is a technique for reading text to determine how the writer feels, usually as positive, neutral, or negative. AI performs it with natural language processing (NLP), the technology that lets computers understand everyday language.
This article covers why AI suits this job, its four stages, one customer message traced through each stage, and how to apply it in a support team.
Why Use AI for Customer Sentiment Analysis?
Businesses turn to AI because the two older methods, surveys and manual reading, both miss the substance of customer complaints. Surveys only capture customers willing to respond, and a team cannot read every message against the same scoring standard.
Research by the Cambridge Service Alliance at the University of Cambridge found this gap among customers who gave scores of 8.5 out of 10 or higher. Of those customers, 90% still used the comment section to voice significant complaints, and the high scores meant the company overlooked them.
The table below compares the three ways of reading how customers feel.
| Method | What it reads | When results are available | Scoring consistency | Main weakness |
|---|---|---|---|---|
| CSAT and NPS surveys | Scores from customers who complete the survey | After the survey closes | Consistent for scores, blank on reasons | Does not explain the score |
| Manual reading by the team | A sample of messages the team has time to read | Depends on how many people are available | Varies from reader to reader | Cannot cover every conversation as volume grows |
| AI | Every message from connected channels | As each message arrives | The same for every message | Needs testing on your language and business terms |
AI is most useful for filling the column surveys leave blank: the content of each customer conversation. CSAT scores still work as a quick summary number, while sentiment explains the reason behind that number.
AI’s weak point is accuracy on your own language and business terms. That is why you should test the results before using them for decisions.
The 4 Stages of AI Customer Sentiment Analysis

AI reads customer sentiment in four stages, and the output of each stage feeds the next. A mistake early on, such as text that was never cleaned up, carries through to the customer history.
Stage 1: Collecting and Cleaning Conversations on Every Channel
AI pulls text from every connected channel, such as WhatsApp, email, live chat, contact forms, and phone call transcripts. It also records the sender’s identity on each message so the results can be combined in stage four.
AI then cleans the text by removing anything the customer did not write, such as email signatures and agent reply templates. It also normalizes abbreviations and code-mixing (Indonesian and English in the same sentence), for example turning “udh” into “sudah” (already) and “cancel” into “batal”.
Stage 2: Reading Polarity, Emotion, and Intensity
Polarity is the direction of sentiment: positive, neutral, or negative. AI also recognizes more specific emotions such as anger, confusion, disappointment, or relief, then assigns an intensity score, for example from -1 for very negative to 1 for very positive.
How accurate this stage is depends on the technology behind the sentiment analysis tool.
| Approach | How it works | Results on Indonesian-language chats |
|---|---|---|
| Lexicon (word dictionary) | Matches words against lists of positive and negative words | Misses sarcasm and words whose meaning depends on context |
| Classic machine learning | Learns patterns from sample messages labeled by hand | More accurate than a lexicon, but weak on long sentences |
| Transformers (e.g., IndoBERT) and LLMs | Reads each word together with the words around it | Best at reading context, provided it has been tested on Indonesian text |
For Indonesian-language customer conversations, choose a tool built on transformers or an LLM (large language model). A lexicon treats “the wait was so long” and “I’m a long-time customer” the same way, even though only the first one is a complaint.
Stage 3: Separating Sentiment by Aspect and Intent
An aspect is the part of the experience the customer is talking about, such as the product, delivery, price, or customer service. One message often carries positive and negative aspects at once, so AI labels each aspect instead of giving the whole message a single label.
Intent is the customer’s purpose in writing, such as requesting a refund or threatening to cancel a subscription. Intent catches complaints that contain no emotional words, such as “balance deducted, transaction failed,” which reads as neutral yet is urgent.
Stage 4: Building Each Customer’s Sentiment History
At this stage, AI arranges the labels from every message into a history for each customer and each period. The result is a trend, such as one customer’s sentiment dropping from neutral to negative over their last three interactions.
This trend is what separates customer sentiment from the sentiment of a single message. A customer whose sentiment keeps falling is an early signal of a rising churn rate, the percentage of customers who leave in a given period.
Also read: AI Ticketing: How AI Is Transforming Customer Service Operations
Example of AI Customer Sentiment Analysis
In this example, AI reads one e-commerce customer’s message as a delivery complaint with refund intent, then raises the customer’s status to at risk of leaving. The message arrives on WhatsApp on the fifth day after the order was paid.
“Item’s good but been 5 days n still not here. CS just replies w a template. If no update tmrw, I want a refund.”
| Stage | Result |
|---|---|
| 1. Collecting and cleaning | Message logged under the customer’s name and normalized to “The item is good, but it has been 5 days and it still has not arrived. Customer service only replies with a template. If there is no update tomorrow, I want a refund.” |
| 2. Polarity, emotion, and intensity | Negative, disappointment, score -0.6 |
| 3. Aspect and intent | Product positive, delivery negative, customer service negative. Intent: requesting a refund |
| 4. Customer history | Third contact in 14 days. The two earlier contacts were neutral (asking about package status), and this one is negative. Trend falling, status raised to at risk of leaving |
A single “negative” label would mislead, because the customer actually rates the product well. The problem lies in delivery and the support response, so fixes should go to the courier partner and to how agents reply.
Stage four changes how the team handles this message. Without the history, it is one delivery complaint, but with it, the team knows this customer has already asked twice without getting a clear answer.
How to Apply Customer Sentiment Analysis
Apply sentiment analysis in stages, starting with one data source, then connect the results to real actions in your support team. A sentiment label that triggers no action only adds another column to a report.
1. Choose Your Data Sources and Labels
Start with the channel that carries the most messages, such as WhatsApp, so results show up quickly. Use three polarity labels and four to six aspects that match your current ticket categories.
Public reviews and social media can come later, once the labels for direct conversations are stable.
2. Add Sentiment to Ticket Priority and Escalation
Ticket priority criteria usually rest on business impact and SLA urgency (the agreed service time limits), so sentiment only needs to be added as a supporting factor. For example, a negative ticket with refund intent moves up one priority level, and tickets from customers with a falling trend go straight to a senior agent.
Messages that mention legal action or a complaint to a regulator, such as Indonesia’s Financial Services Authority (OJK) for fintech, need a separate rule. Route these messages to your legal or compliance team, and do not let AI reply to them.
3. Track Sentiment Trends by Customer and Segment
Review the list of customers with a falling trend every week, especially B2B customers on annual contracts. Also compare sentiment between new and existing customers to see whether the problem sits in onboarding or in routine service.
4. Test Accuracy on Past Conversations Before Going Live
Take 200 to 300 conversations from last month, label them by hand, then compare those labels with the AI’s. Pay close attention to two message types that machines find hard to read: sarcasm and complaints with no emotional words.
Turn on automatic escalation only once your team is satisfied with the match rate. After that, give agents access to correct wrong labels.
5. Measure the Impact on Satisfaction and Retention
Measure results with metrics that link sentiment to how cases are handled, not just the number of negative tickets.
| Metric | How to calculate | What it means |
|---|---|---|
| Label correction rate | Labels changed by agents divided by total labels in a month | An increase means the rules or the model need updating |
| First response time on negative tickets | Average first response time on negative tickets, compared with the average across all tickets | Should be faster than the average |
| Sentiment change at ticket close | Percentage of negative tickets closed with neutral or positive sentiment | Handling quality, not speed |
| Retention of customers with a falling trend | Percentage of customers with a falling trend who are still active 90 days later | Direct impact on revenue |
The last two metrics sit closest to business impact. If retention among customers with a falling trend has not improved after three months, check whether sentiment labels actually trigger action or only sit in a report.
Also read: Benefits of AI in Customer Service for Reducing Agent Workload
Conclusion
AI customer sentiment analysis runs in four stages, and the last one, the history per customer, does the most to determine its business value. Labels by aspect and intent show what went wrong, while the history shows who is most likely to leave.
One step you can take this week is to pull 200 WhatsApp conversations from last month and label them by hand. The result shows how many complaints never appear in your CSAT score, and it gives you material for testing a sentiment analysis tool.
Optimize Your Customer Service
Schedule a demo of Adaptist Prose and see how an integrated ticketing system helps bring tickets, conversations, and customer data together in a single dashboard. With a more structured workflow, teams can respond faster, reduce operational burden, and maintain consistent service quality as the business grows.
FAQ
A sentiment score is a number that represents the sentiment of one message or one customer, usually on a scale of -1 to 1 or 1 to 5. Averaging scores per customer or per month makes it easy to see whether sentiment is improving or getting worse.
Yes, especially transformer or LLM models tested on Indonesian text, which read “mantap” (great) in “Mantap, paket saya nyasar lagi” (Great, my package got lost again) alongside the complaint. You should still test the results on your own customer conversations, since sarcasm depends on context.
Social listening is the monitoring of public conversations about a brand on social media and forums. Sentiment analysis is one of the techniques inside it, and the same technique also applies to direct conversations such as chats and emails to a support team.
Yes. Call recordings are converted to text with speech-to-text technology and then analyzed like chats, and some systems also read the tone and pace of the caller’s voice.
Yes, because customer conversations usually contain personal data such as names and phone numbers, so processing them falls under Indonesia’s Personal Data Protection Law (UU PDP). State the purpose of the analysis in your privacy notice, and limit access to the results to the teams that need them.




