Nationwide Indonesian bank · 7.5M customers
Survey sentiment for 7.5 million customers, processed in real time
A nationwide Indonesian bank surveys customers with everything from 50-question deep-dives to single-question NPS. Coding the sentiment and themes by hand took two to three months a year. An agent now does it as the responses land, in real time.
Customers surveyed
7.5 million
Survey types
NPS to 50+ questions
Processing time
2-3 months/year → real time
Human role
taxonomy + sampled validation
The backlog
The bank serves 7.5 million customers and surveys them constantly: long-form questionnaires of 50 questions or more, mid-length surveys, and single-question NPS. Every response carries sentiment, what customers like, what they don't, and why, and the bank needs that read across the whole book, not just a score.
A team coded it by hand, and working through a year of responses took two to three months. Insight always arrived late: the bank learned how customers felt about one year somewhere in the middle of the next. By the time a theme surfaced, the moment to act on it had usually passed.
Why it resisted off-the-shelf tools
Surveys are not uniform. A 50-question deep-dive and a one-line NPS need different handling, and free-text answers arrive messy: mixed Indonesian and English, sarcasm, typos, and complaints wrapped in praise. A generic sentiment tool scores a line positive or negative and stops there.
The bank did not need a polarity score. It needed the themes and the reasons: which product, which branch, which part of the app, and whether the feeling was rising or falling. That is extraction and judgment across varied documents, which is agent work, not a plug-in.
How the agent runs it
The agent reads each response as it arrives. It classifies sentiment, extracts the themes behind it, and segments by survey type and customer group. Long, mid, and NPS responses all feed the same pipeline, each handled to its length, so a 50-question survey and an NPS score end up in the same live picture.
Results land in a running dashboard instead of a quarterly report. When promoters climb on a new app release or detractors cluster around call-center wait times, the bank sees it the day it happens, while there is still time to respond.
Where a person stays in charge
Sentiment coding is analysis, not an action that moves money or reaches a customer, so the agent runs it end to end. The judgment that stays human is the framework: the bank's team owns the theme taxonomy and the edge-case rules, and refines them as the picture sharpens.
Accuracy is measured, not assumed. A sample of the agent's coding is checked against human coding on a regular cadence, so the bank knows how far to trust each segment. The agent does the volume; people keep the definitions and the confidence honest.
From the run log
A batch of responses, coded as they land
A representative slice: mixed survey types arrive together, and each is classified, themed, and segmented in seconds, so the dashboard never carries a backlog.
Want a system like this in your business?
We build one workflow at a time, deliver it into your accounts, and monitor it after. Pricing is published up front.
Start with a free workflow audit
A 45-minute call. We map three processes agents can take over, estimate what each would save, and quote what it would cost. No obligation.