Quick-start guide for Lyra AI

The world is on fire right now with anticipation about how artificial intelligence (AI) is going to change the business landscape. There’s good reason for it – early adopters are already seeing massive benefits in productivity, innovation, and automation that are returning solid commercial gain.

While there’s been a lot of hype about what artificial intelligence (AI) technology can do, there’s also recognition we’ve entered a new climate for business growth. According to a 2024 report by Accenture, more than $10.3 trillion in additional economic value can be unlocked by 2038 if organizations adopt GenAI responsibly

Chattermill has been working responsibly in this area for more than eight years with Lyra AI, our proprietary technology for analyzing customer feedback data. It uses a blend of AI technologies, including GenAI, to unlock deep insights from the data. Lyra AI has  been developed to answer the big question all businesses want to know – ‘What are my customers thinking and feeling about my brand?’

Lyra AI is a closed system to ensure customers glean accurate and consistent insight from their CX Intelligence. Here’s what you need to know to get started with Lyra and begin to drive Experience-Led Growth (XLG) for your business. 

Onboarding and training the Lyra model

It might seem like magic, but AI is only as good as the training it gets. Large language model (LLM) services such as OpenAI’s ChatGPT and Google’s Bard rely on algorithms that process natural language inputs and deliver human-like responses to queries. The algorithms are trained on billions or even trillions of parameters from myriad places including articles, blog posts, books, Wikipedia and other internet resources. 

You only have to look at what’s happened to ChatGPT to understand the impact of an “open” model. Joint research from Stanford University and University of California Berkeley in 2023 found the performance of LLMs can drop substantially in a relatively short amount of time, highlighting the need for continuous monitoring.  

That’s why Chattermill has been ultra-cautious about protecting Lyra from the bad influence of poor-quality data. We go to great lengths to ensure the integrity and consistency of the insights it delivers to our customers.

Using existing industry models for a quick start with Lyra 

Lyra AI has different options for onboarding and training the model. One option is to use an  industry model that’s already been trained by Chattermill. We have robust models for industries such as retail and ecommerce, finance, travel and hospitality, subscriptions, and all other areas from utilities to healthcare and telecoms. Our industry models are continually being trained on high-quality data so there’s a breadth and depth of relevant information for data analysis.

For example, we’ve been working with subscription companies such as HelloFresh for years. If another food delivery company wants to better understand refunds or pricing, we don’t need to retrain Lyra. It already knows how customers describe those themes when giving feedback.

If a customer is satisfied with a general approach to analyzing their data based on industry standards, we don’t need to do any additional training. It’s good to go from the very beginning, and onboarding is seamless.

Taking time to build a bespoke version of Lyra 

At the other extreme is building a completely unique model for new customers. The Chattermill machine learning team uses data from the new customer, creates a bespoke theme structure to fit the data, and sets up all the necessary tags. This takes time and there are several iterations with the customer to make sure they’re happy with the themes we’re suggesting. Once all the themes and tags are agreed, Chattermill trains their version of Lyra before we roll it out. It can take five days to train, but the advantage is they have as much granularity as they want – it’s up to the customer. 

In the happy middle with a self-supervised model of Lyra 

If a customer wants to take advantage of an industry model but also needs training for their specific business, Chattermill can provide self-supervised training. 

For example, if a customer has a niche retail business for something like custom tire stickers, they can use an existing retail model and Chattermill can create additional themes in Lyra AI directly related to tire stickers. That could include tags for adhesives or weather damage, things your average retail model doesn’t encounter. The customer gets the benefit of standard themes such as delivery, shipping, ordering, and returns. 

It usually takes two to three days to train a self-supervised model of Lyra. Every customer is different so Chattermill would need to confirm the length of training time required for your business.

How long does it take to get started?

A few different factors influence how long it will take to start benefiting from Lyra AI. If you have all your data ready and you want to use an industry model, you can be off and running within hours.

To fully train a bespoke Lyra model takes Chattermill up to five days. This also assumes all the data has been collected and is ready to go. A bespoke model can get bogged down when there’s not a clear strategy on what you’re trying to achieve or there are a lot of business areas that will be using Chattermill. It’s essential to consider all the ways a theme can surface deep insights across multiple groups before training begins. For example, retail product teams may have a different focus for returns than the fulfillment teams – product satisfaction compared to the delivery experience. Larger organizations with more stakeholders typically require more time to identify the themes and tags that will be used to train  Lyra. 

A hybrid Lyra model falls somewhere in the middle of a vanilla industry model implementation and a bespoke model. Self-supervised training of Lyra AI generally takes between one and three days depending on how many additional themes are required. As with the bespoke model, themes must be agreed upon before training Lyra AI can begin. 

Customizing Lyra for your business

There are four main components to getting your business set up with Lyra. 

  • Data Sources: Chattermill can analyze different types of data sources, such as reviews, surveys, and even conversational data from customer service calls. Customers choose the specific data sources relevant to their business and their growth goals and provide them to Chattermill for analysis.
  • Themes: Themes are used to categorize and analyze customer feedback. The theme structure is built to represent your customers’ journey. They can be customized based on business needs, including new themes for your industry or business segment. Existing themes can be modified to align with your unique requirements. It’s essential the themes are well defined and agreed on in your business because themes are the anchor to how Lyra AI develops deep insights. 
  • Model Training: If you have a theme or aspect of your business that is not part of a Chattermill pre-trained industry model, a custom model can be trained for you. This involves providing explicit instructions to the model and giving it examples of the desired theme or concept. The self-supervised method allows you to influence the training process for the model, which is conducted by Chattermill. Once your Lyra AI model is trained, periodic reviews and more training will guarantee you’ll always get the deepest insights and most critical information about what your customers think and feel about your brand.
  • Dashboards and Reporting: Chattermill helps you set up dashboards and reporting based on your specific analytics needs. If you require assistance in creating new or different visualizations as you begin to benefit from Lyra AI, Chattermill has the expertise to support you.

Updating the model, adding new themes, and discovering emerging themes

Chattermill takes responsibility for updating the model and adding new themes. It’s important to note that customizing your business model with Chattermill may involve collaboration with the Chattermill Customer Success and Machine Learning Operations teams. They work with customers to understand requirements, provide guidance, and check that the theme structure and analysis align with expectations. Most of all, Chattermill works to ensure Lyra is helping unlock growth opportunities based on your business goals. 

Why Lyra is a closed model

Currently, Chattermill customers cannot update the model themselves, but there are plans to allow for more client control in the future. Unlike “open” AI tools, we are very intentional about not allowing anyone to randomly change or add themes. It doesn’t take long for a cascade effect to occur that diminishes or completely corrupts the value of the insights Lyra AI delivers. 

This can happen, for example, when similar names are used for the same theme, like pricing, prices, price changes, or price shifts. Or, when you have completely different names, like promotions, specials, or bargains. It dilutes the value of the theme, becomes unwieldy to manage, and creates havoc with your CX Intelligence – and that can literally happen overnight. No matter how smart Lyra is, the whole system becomes useless when it’s open for anyone to train it.

The closed system approach maintains data integrity and prevents issues with metrics and theme definitions. This is our best guarantee of Lyra remaining the most powerful tool for customer feedback data. 

Mariana Lisboa

Mariana Lisboa is the Associate Director of Machine Learning Operations of Chattermill

Learn more about Lyra AI

For more information about how Lyra can transform the way you analyze your customer feedback data, take a look at our article, Why Lyra AI comes out on top when comparing AI platforms for text analysis.