Direct customer feedback

In a highly competitive environment, knowing customer complaints and expectations has become a major challenge for all companies. Direct customer feedback has become an increasingly important contributor since the emergence of the Internet. Whether it be results from a speech recording converted into text, a comment from a marketing survey, an incoming e-mail or a tweet, direct feedback neatly reflects the voice of the customer because it is unfiltered and without any interpretation bias.

What are the stakes of analysing customer feedback?

The main challenges are to :  

  • Reduce churn by understanding customer irritants and taking appropriate actions… before they leave,
  • Increase sales – typically cross selling – by anticipating customer expectations,
  • Improve the customer experience with the brand and increase its natural promotion,
  • Deliver more value with the provided service.

The difficulties to analyse customer feedback messages

The difficulty of its analysis comes, of course, from the nature of the data (free text), the number of messages that companies now receive, the variety of comments, their evolution over time, and their dependence on the language and culture to which they relate. Human reading may be envisaged but quickly encounters its limits as soon as the volumes to be analyzed are large (from a few hundred to tens of thousands per month for large corporations).

Automatic processing is an option and has been implemented in the last few years using NLP (Natural Language Processing) technologies, which by the way have made remarkable progress during the last 18 months.

Traditional solutions to automate feedback analysis

For large corporations, the solution has often been to use agencies specializing in opinion analysis (BVA, Sofres, Médiamétrie, etc.) or a few customer journey specialists. These agencies have enough “Customer Experience” analysts to deal with these volumes and have developed automatic processes whose costs have been partly amortized thanks to the volumes to be processed.

For all the others, the only possible solution was to use software based on purely statistical methods associated with keyword search or, more recently, semantic analysis using linguistic models, generally based on rules.

The shortcomings of traditional software approaches: quality, cost and language support

The results are:

  • The detection of fairly generic trends for the statistical or keyword-based solutions which make a detailed analysis difficult.
  • Quite relevant results for rule-based linguistic models, that are developed by computer linguists, but often expensive to use because their development is delicate and dependent on each language. Moreover, any change in the criteria to be analyzed requires specific developments, most of the time carried out by an external party, often the supplier of the solution used.

Why is Artificial Intelligence a real innovation in this field?

First, by offering high-performance algorithms, in particular those based on neural networks. Moreover these algorithms are all accessible in an “open source” format. Anyone can use them.

Then because it is possible to use pre-trained language models to better understand the “language” of the feedback message. These models can be trained on large quantities of monolingual or multilingual messages that already exist on the internet or are held by the clients themselves or even opinion agencies (which have large volumes in their archives). These pre-trained language models ultimately allow a significant increase in quality. 

Finally because with Artificial Intelligence the most difficult challenge is to create a dataset to learn from and train the models. The prerequisite is to have a sufficient set of customer messages which generally all companies have available. Annotating these messages is now within reach of any analyst specialized in “Customer Experience”. All it takes is a good knowledge of the customer’s journey and a little common sense to put one or more annotations on each message in order to create a reference on which the machine will be able to learn.

Nowadays there are easy-to-use tools for analysts requiring no programming knowledge at all. It is therefore the business analysts themselves who can build and evaluate their customer feedback message classification using the most advanced and powerful algorithms such as neural networks.

How does the analyst proceed in practice?

The “Customer Experience” analyst must firstly define categories or “labels”. He or she can do this either thanks to existing experience or after exploring the messages. At any time in the process the category definition can be reworked and improved:

The analyst will also define the labels to identify the sentiment.

Then the analyst starts to classify a few messages in each of the defined categories. The messages can also be conveniently filtered in order to work at a finer grained level. The analyst will ideally assign a sentiment to each category. 

This work can be long and tedious if done manually. But Artificial Intelligence can help the analyst to create samples faster in each category.

As soon as a few samples are created, Artificial Intelligence can start to learn and suggest new samples that the analyst will have to validate, correct or reject. This saves precious time for the analyst and opens the possibility to subcontract this task if needed.

Once the analyst has built a representative dataset, its quality can be reviewed, for instance by visualizing the distribution of the different categories.

When the messages are annotated sufficiently, the analyst will want to find the best algorithm providing the best possible quality to automatically classify new messages. This can be achieved by testing different algorithms among the most powerful on the market, and adapting them to his particular challenge (text classification in this case), such as Spacy, Sklearn or Flair. 

It is even possible to create two models in the same project: one to identify the category and the other to identify the related sentiment.  

Subsequently these algorithms will be trained then tested on the dataset and finally be compared to their respective performance… category by category:

It is often necessary to refine the results by changing the parameters of the algorithms, which is the domain reserved for Data Scientists and AI specialists.

The Customer Experience analyst will be able to filter the dataset on a particular category and proceed with a new experimentation phase.

The last step is to put the model thus created in production. The model will be able to automatically classify a flow of messages by theme and sentiment without any human intervention and with an accuracy of about 95%, i.e. a performance superior to what a human being can produce.

Presentation of the results in a “Customer Experience” portal

The example below shows the output of an automatic analysis in the BVA CX Insights portal. The objective is to provide the end customer access to the results including: 

  • An interactive online dashboard,
  • Statistics on categories and sentiments extracted automatically from the messages to provide decision support to improve the quality of leads, or train customer-facing personnel
  • Access to message details and filters…


Artificial Intelligence offers exciting new possibilities to analyse customer journeys and makes these possibilities available to a much wider audience through easy-to-use, effective and inexpensive applications.

The capacity to listen to the voice of your customers is becoming an increasingly critical competitive advantage to deliver the products and services that customers want and will promote.