Generative AI in customer service and analytics

Generative artificial intelligence in customer service

Generative artificial intelligence is defined as deep-learning models that can create high-quality texts, images and other content based on the data they were trained on. The category includes, e.g. the popular ChatGPT, Google Bard or Midjourney.  

From the perspective of customer service, this article will focus on text-to-text processing models. We will zoom in on language model training, including large language models (LLMs). 

How do you train a text model?

The first stage involves training your model on a large corpus of texts. You can choose from the following training strategies:

• MLM (Masked Language Model) – a strategy based on context rather than specific word representations;
• Autoregressive learning – a strategy that automatically predicts the next word in a sequence based on previous inputs.

Fine-tuning

Your AI (your language model) is then presented with a series of tasks; for instance, it has to summarise texts, solve logical puzzles or answer questions. This helps adapt it to the context of the situation. The process is known as fine tuning. Fine tuning can be identified with limited learning, e.g. only a small proportion of parameters or model layers can be modified.

Prompts

All this serves as a springboard for your model, which then generates answers based on its analysis of previous text inputs. In order to obtain accurate and valuable results, the specialist responsible for the interface of the model needs to formulate effective prompts (hints). The process consists of many stages; the user prompts the model iteratively, gradually narrowing down their query, until they are satisfied with the answer.

To formulate effective prompts, you need to:

• understand the model,
• act iteratively,
• validate (test),
• experiment,
• avoid generalities.

Target parameters

Importantly, in your interaction with generative AI models, you need to set up the prompt parameters that will allow users to fine-tune the model to match their specific demands. These parameters are the key to manipulating the quality and character of generated results.

Please note: Parametrisation is only possible via an API or in dedicated sandboxes.

A few important settings include: 

• Repetition penalties: this parameter tells artificial intelligence to avoid repeating the same words or sentences.
• Temperature: temperature affects randomness. A higher temperature will lead to responses that are more diverse, but less consistent. A lower temperature, in contrast, will generates responses that are more predictable.
• Top-k sampling: a technique that limits the pool of potential words or tokens in the response to the most likely candidates. The k parameter defines how many of these probable tokens should be considered. The higher the k, the more diverse the responses.
• Maximum length: tells generative AI how long the response should be. This is important if you don’t want it to operate on responses that are too long and non-specific.
• Maximum token number: defines the maximum length of the response in tokens. This is useful when you want to avoid getting excessively long responses.
• Stop sequence: tells generative AI when to stop generating the response. You can set it to stop when it reads the word “stop”.

Business applications of generative AI

Since it can handle almost any task, generative AI now shows a growing impact on employee productivity, especially in knowledge-intensive processes that require high cognitive ability.

Knowledge-intensive processes depend on a thorough, often specialist knowledge of a given subject or domain. Examples include e.g. legal analysis, medical diagnostics, scientific research or financial analysis.

In processes of this kind, generative AI may, for instance, generate synthetic data that can be used to train other models, create reports or analyses based on available data, or even suggest possible solutions based on previously learned contexts and templates.

Collaborative intelligence

Cognitive abilities include the ability of our mind to process information, solve problems, take decisions, learn new things and engage in creative thinking. Despite its huge potential, generative AI does not replace, but rather collaborates and cooperates with people, creating what we call collaborative or cooperating intelligence.

Collaborative intelligence describes the synergy that arises when people and machines work together, pooling their unique abilities in pursuit of a shared objective. This approach is particularly important in fields that require thorough understanding, empathy, ethics and other features that are difficult or impossible to teach to a machine.

In these fields, mixed teams consisting of people and generative AI models can work together, shaping an environment in which machines help automate and streamline tasks that can be automated, while people focus on aspects that require a human touch and judgment.

Artificial intelligence and security

Artificial intelligence may bring many benefits, but also creates certain security challenges. Below are some examples of issues that you should keep in mind when using AI-based solutions:

Ordinary and sensitive personal data

Ordinary personal data include, for instance, first names and last names, while sensitive data include race and religious beliefs. Since data is processed and stored on servers, even in commercial solutions from providers such as OpenAI or Google, you might want to avoid transferring particularly important data to such platforms.

You can prevent this in several ways: you can use an API, opt for paid plans, e.g. Enterprise plans, or mask your data.

An illusion of hallucination

AI can lull our vigilance. After many correct outputs, you can get one output that looks credible, but is in fact wrong. The fact that it is correct linguistically does not mean it is also factually accurate. Let’s remember that AI is a probability-based model; it is not 100% reliable. You should always verify the data it produces.

Bias

AI Bias is particularly pronounced if the training sequence was not sufficiently diverse. For instance, let’s suppose that an image generator was trained on data that consisted only of photographs of Arabian horses.

In the future, when asked to generate an image of a horse, the model will generate a target based on the template it has known. Most likely, it won’t be able to tell the difference between an Arabian horse and, e.g. an Andalusian horse.

Generative AI in customer service

In many sectors, potential applications of AI, even only those limited to customer service and data analytics, are very broad, and can be divided into the following categories:

• Points of contact with the customer (front-office) – AI can assist you in, e.g. taking orders, but also creating quotations or ad-hoc translations. Apart from content, AI can also take care of visual messaging and correct language;
• Back-office operations, e.g. processing documents;
• Marketing-sales processes. Sales involves many interactions and transactions and generates a lot of information, such as e-mails, phone call audios and videos from meetings. Such disorderly data can represent more than 80% of all data in your business, and as much as 43% are never used at all (source: Report: “Generative AI in Business” – Portal sztucznej inteligencji – Portal Gov.pl (www.gov.pl));
• Advanced analytics, such as partner verification, preparing agreements, legal analysis;
• Risk and security management, e.g. cybercrime, fraud;
• Operational and management reporting.

Practical AI applications in customer service

We have developed an AI-based transport application that takes over the repetitive processes of e-mailing, clarifying shipping details and creating orders. The app is based on a trained LLM GPT model that uses autoregressive learning. A key feature of such models is that they can understand the context and produce consistent, meaningful outputs.

If you want to learn more about the app, read this post: Altkom AI Assist. A generative AI-powered transport application for customer service and analytics or directly explore the Altkom AI Assist tool and its mail support features for a range of sectors.

Pros and cons of AI in customer service

Let us now go back to our subject, i.e. customer service and analytics. In this context, AI successfully supports data collection, tagging, classification and grouping. It also helps detect anomalies, enabling a quick response to irregularities and effective fraud prevention. Below are listed some of the pros and cons of AI-powered solutions (such as Altkom AI Assist), covering aspects related to cost, quality and performance:

Pros:

• Automation: A software that classifies messages can automatically sort and assign them to categories, which saves your time and minimises the risk of error.
• Better performance: Performance is improved since AI can quickly identify and reply to important messages; it also allows you to define your priorities.
• Better customer service: Since AI-based software can quickly access all the necessary information and forward messages to appropriate departments, it improves the quality of customer service and increases customer satisfaction.
• Easy integration with other systems: Many cutting-edge software solutions of this kind are easy to integrate with other tools, such as CRMs or ticketing systems, allowing you to tap their potential within your core IT infrastructure.

Cons:

• Necessary monitoring and training: The software requires regular monitoring and conservation to maintain adequate performance and functionality, which may require extra effort and increase system maintenance costs.
• Misclassification risk: There is a risk that the software might misclassify some messages, which can cause response delays or loss of important information.

Training the model to deal with specific business scenarios

Importantly, every LLM model needs to be retrained to ensure that it will communicate in a natural way, ask the right questions and produce meaningful answers. Whenever there’s a change in correspondence patterns, the model needs to be provided with a new training set.

This could be considered as a drawback, but is it not the case that employees also need to step up when many other systems are developed? Users need to provide a training set, i.e. correspondence, which will reflect real situations and help address their problems (as a result, the AI will be able to take over repetitive tasks and take the load off your employees).

Fears around AI deployment

Especially at initial deployment stages, you may worry that a misclassified e-mail can generate losses (such as a loss of sales opportunities). Indeed, this may happen. Such situations should be monitored; in Altkom AI Assist, this is done by a built-in reporting module.

But we should also ask ourselves: are we monitoring the losses that are caused, for instance, by our delayed response to customer messages? Do we keep any updated statistics to that effect? Even if the answer is yes, we are not going to lose any of these data once we have deployed the system.

AI-based analytics in customer service

Thanks to AI-based customer service analytics, companies can better understand the needs of their customers, manage them more effectively, and offer a more personalised and satisfying experience.

Pros:

• Quick analysis: analysis can even be done online.
• Granularity: multi-level analysis, including individual customers.
• Automated scoring: customers may be scored based on your correspondence and various parameters taken from other sources.

Cons:

• Low-quality data: There is a risk that an assessment will be made based on insufficient data, e.g. contact history that is too short.
• Overconfidence: this may encourage employees to neglect verification and app monitoring.

AI in customer service. Important takeaways

While generative artificial intelligence can help automate many routine tasks, allowing employees to focus on the creative aspects of their work, it is still up to people to decide what should be deployed and how.

The article covered a wide range of generative AI applications in customer service, data analytics and risk and security management. From points of contact with the customer (front-office) to back-office operations, generative AI can significantly improve effectiveness, customer service quality and repetitive task automation. However, in order to understand and adapt AI models to specific business needs, you need to be diligent and open to continuous learning. We encourage you to explore what generative AI can do for your organisation and recognise its potential as a tool that can support and facilitate, rather than replace, human labour.