Text analytics for customer feedback: how machines read what customers write

Open comments usually contain the most useful part of customer feedback, but they are also the hardest part to process consistently once volumes grow. Text analytics structures those comments so teams can compare topics, sentiment and trends without reading every answer by hand. It does not replace reading; it gives the analysis process a consistent structure at larger volumes.

Key takeaways:

• Text analytics makes four moves per answer: detect the language, split into subjects, classify against a taxonomy, attach sentiment per subject. The output is a measurable dataset, not a summary.
• The taxonomy is the craft: it must mirror your journeys deeply enough to point at an owner, and stay stable enough to keep trend lines readable.
• Language models have transformed reading quality, and they wobble: even at "deterministic" settings, research measures accuracy swings of up to 15% across identical runs. Pin down the taxonomy, prompts and model version.
• Nine checks separate a measurement system from a demo, from native multilingual reading to a route back to the raw comments.
• A weekly human sample remains the audit on all of it.

What text analytics does with one answer

Consider this example: "Delivery was fast, but the driver couldn't find our address and the app kept crashing."

Text analytics reads that answer in four moves:

1. Detect the language, and read the answer natively in it.
2. Split the answer into subjects. This one contains three: delivery speed, driver navigation, app stability.
3. Assign each subject to a category from a fixed taxonomy, several levels deep (Delivery > Speed; Delivery > Driver; Digital > App > Stability).
4. Attach sentiment per subject: positive, negative, negative.

Multiply by every answer, every day, and the open text becomes a dataset: topics with volumes, sentiment and trends, linked to the scores and metadata of the customers behind them. That link is what makes a key driver analysis possible, and it is the difference between reading feedback and measuring it. What you then do with the emotional layer, trends, early warnings, segment comparisons, is covered in our guide to customer sentiment analysis; this article is about the machinery that produces it.

The taxonomy is the craft

The category structure is as important as the model. A bank needs topics for branches, cards and claims; a retailer needs returns, sizes and store staff. A generic list often becomes too broad to assign a finding to an owner.

Two requirements make a taxonomy work:

• It fits the business. Categories mirror the journeys and vocabulary of your customers, deep enough that a finding points to an owner.
• It stays stable. Trend lines are only readable when the categories underneath them stand still. Extend the taxonomy when new themes appear; resist redesigning it every season.

This is where the craft sits in a platform such as Hello Customer: a taxonomy built per client, applied consistently by AI that reads open text in dozens of languages, several levels deep.

Depth is what separates a reporting taxonomy from a working one. "Delivery" as a single bucket produces a chart; "Delivery > Timing > Missed window" produces a work order for the carrier team. As a rule of thumb, keep going until the deepest level names something one team can fix, then stop: beyond that point extra levels add maintenance without adding ownership. For a multilingual base, hold one taxonomy across all markets and let the model map each language onto it; separate taxonomies per country feel respectful of local nuance and quietly destroy every cross-market comparison the leadership team will ever ask for.

From keywords to language models

The technology has moved through three generations. Keyword rules came first: cheap, transparent, and blind to context ("not fast" counted as fast). Trained classifiers came second and learned context within the categories they were trained on. Large language models are the third: they read negation, tone and mixed feelings close to the way a person does, in most languages, without translation.

The economics moved even faster than the quality. Stanford's AI Index measured the inference cost of GPT-3.5-level performance falling more than 280-fold in two years, which is why reading every customer comment, in every language, every day, went from a research budget to a line item. The 2026 edition of the same index has 88% of organisations using AI somewhere in 2025; in customer service specifically, a 2024 Gartner survey found 85% of service leaders planning to explore or pilot customer-facing conversational GenAI in 2025. The reading capability is now table stakes.

What has changed is reading quality and cost. What has stayed is everything around the reading: the need for a domain taxonomy, the link to scores and metadata, and the verification habit below.

The reproducibility problem nobody demos

Language models vary their answers across repeated runs, even at the settings meant to hold them steady. This is measured, not anecdotal: a systematic study of five models across eight tasks found accuracy varying by up to 15% between identical runs at "deterministic" configuration, with not one model producing repeatable results across all tasks.

For a chat session, that wobble is harmless. For a measurement system, it is fatal if unmanaged: a trend line that moves because the model moved is worse than no trend line, because someone will act on it. A measurement system therefore has to pin down everything around the model: a fixed taxonomy, fixed prompts and a known model version, with changes logged like any other methodology change. When the March delivery-complaints line jumps, the first question should never have to be "did we change the model?"

What to check before you choose tooling

• Native multilingual reading. Translation before analysis flattens tone; sentiment should be read in the language the customer wrote.
• Sentiment per topic, not per message. One answer, several feelings; the technology must keep them apart.
• A customisable taxonomy. Your categories, your depth, stable over time.
• Stable classification. Ask a supplier to run a sample of your own feedback twice and to show you how often the two runs agree. No model reaches a hundred percent, and a supplier who has measured it can tell you the figure.
• One taxonomy across every source. Surveys, reviews, calls and tickets should run through the same categories and the same sentiment layer; otherwise a trend line across channels cannot be drawn.
• Privacy by design. Feedback contains personal data. Check where it is stored, whether PII is recognised and anonymised before analysis, and whether the supplier will state in the contract that your customers' words are not used to train anyone else's models.
• The link to scores and metadata. Text without the surrounding data cannot feed a driver analysis.
• Continuous processing. Feedback arrives daily; batch analysis answers last month's questions.
• A route back to the comments. From any chart, the team should be able to open the underlying answers and check what the classification means.

A comparison of the platforms in this space against these criteria is maintained separately in the text analytics tools overview.

Running a proof of concept that proves something

Most text analytics evaluations test the demo, not the system. A proof of concept worth the name has four parts, all run on your own feedback:

1. The same sample, twice. Give the supplier five hundred of your real comments, have them classified, wait a week, submit the identical file again. The agreement rate between the two runs is the stability figure no brochure prints.
2. A blind human panel. Have two of your own people tag a hundred of those comments against the proposed taxonomy before seeing the machine's labels. Where the humans disagree with each other, the comment was genuinely ambiguous; where the humans agree and the machine differs, you have found the model's real error profile.
3. The per-language slice. If a fifth of your feedback is in French, check the French subset separately. Averages across languages hide exactly the market where the tool might be weakest.
4. The route back. From the resulting topic chart, ask to click through to the underlying verbatims. If that path does not exist in the demo, it will not exist in the quarterly review either.

A supplier comfortable with this exercise is telling you something; so is one who is not.

Keep a human quality check

Read a fresh sample every week and compare the comments with their labels. This catches drift and keeps the team familiar with customers' language. When evaluating a tool, a labelled sample of your own feedback is more useful than a general accuracy claim.

Frequently asked questions

What is text analytics for customer feedback?

The technology that automatically structures open customer text: detecting language, splitting answers into subjects, assigning topics from a taxonomy and attaching sentiment per topic, linked to scores and metadata. It turns open answers into a dataset you can measure and trend.

How is text analytics different from sentiment analysis?

Sentiment analysis is one of its outputs: the emotion per topic. Text analytics is the whole machinery, including topic classification and the link to scores.

Can't we just paste our feedback into ChatGPT?

For exploration, yes: a general model summarises a batch of feedback impressively. A measurement system asks for what a chat window does not offer: a stable taxonomy applied identically every day, the link to every score and segment, volume processing, and a verifiable trail from chart to verbatim. The two work best together: connect the classified feedback to the AI tools your team already uses, for instance through an MCP connection, and ask your questions there, grounded in consistently classified data.

How much feedback do you need for text analytics to be worthwhile?

The technology works from the first answer; the business case arrives when reading everything by hand stops being realistic. In practice that is somewhere in the hundreds of answers per month.

How accurate is automated topic and sentiment classification?

Do not treat a single accuracy percentage as decisive. Human reviewers also disagree on ambiguous comments, and research shows even deterministic model settings produce run-to-run variation. Test the system on your own sample, look at the types of mistakes it makes and judge whether classification stays consistent over time.

What this changes in practice

Text analytics turns a large collection of comments into something a team can measure and investigate. The models will keep improving underneath it; the taxonomy, the metadata links and the weekly human sample are what make the output trustworthy enough to act on. What it should deliver is a reliable route from a recurring customer issue to the people who can address it.