So, you’ve heard about the new large language model (LLM), ChatGPT? It’s been tipped to take over humans in disciplines like content creation, customer services and even in healthcare roles.
If you’ve tried it, it’s impressive. It responds to queries with a very human-like answer as fast as they are required, and it learns from every conversation it has, tailoring responses accordingly.
So, can we just plug it into our customer services function and see if it answers customer queries?
Don’t misunderstand us, we’re enthusiastic about LLMs and what they can achieve, but any experimentation with new technologies must be managed, and it should never be undertaken lightly or without consideration for the pitfalls. We like to look under the bonnet of new technology like ChatGPT to better understand how, combined with the knowledge of your employees, you can make use of it in a controlled way to best effect. Having the knowledge to know when the technology is wrong is important for staying one step ahead.
Correctly used, LLMs bring huge benefits. They free up time by doing some of the more monotonous and mundane tasks, answering repetitive customer questions for example. But your team must have control over the bigger picture and the impact on performance, and they should understand its limitations to be able to identify when things are going wrong.
What ChatGPT can’t do.
ChatGPT has five key pitfalls that impact the conversations it has:
- It has no emotional intelligence – it’s been exposed to vast amounts of text data from the internet and predicts the next word or sequence of words based on the context of the preceding words. Without the emotional intelligence of humans, it can’t understand nuances of human emotion and respond sensitively to complex situations. Deployed in a customer services solution, if it failed to respond to complexities of a question, the customer is more likely to default back to wanting to speak to a human.
- It struggles with context – it can deliver a response from the data, but that response may not be in the context of the conversation. For example, a human contact center agent will hear a number of indicators in a conversation such as tone of voice and style of question, rather than just the words used in order to resolve an issue, but large language models are unable to provide the same judgment.
- It can’t provide new answers to existing questions. It’s limited to the data it’s been trained with and therefore struggles to generate new responses which may lead to predictable or repetitive conversations. Without training, a business using an AI chatbot will find that it simply repeats itself and may result in customers defaulting back to more costly channels of communication once again reducing any financial savings that could have been gained.
- Be wary of bias. ChatGPT relies on the data it was trained with – if it comes from a single source or is not diverse enough it delivers a response that it knows, but isn’t satisfactory and in certain cases, may even offend. There must be checks in a customer service scenario to ensure that responses don’t offend customers and negatively impact a brand.
- It’s vulnerable to security concerns. Like all technology, hacking or security breaches are ever present, putting conversations at risk because the data they’re based upon has been breached. ‘Prompt injections’, where the LLM follows instructions provided by a malicious user to provide incorrect responses, threatens to undermine its use. The potential for reputational damage means that a successful deployment of ChatGPT and other LLM technologies must be planned and managed by professionals.
Make AI tools part of your wider strategy.
In customer services specifically, cutting corners simply isn’t an option, and can do more harm than good. Designing a watertight conversation AI strategy guarantees you can automate and reduce response times whilst improving customer satisfaction and protecting and building your brand. An LLM like ChatGPT is only one tiny piece of a much larger strategic jigsaw puzzle.
Designing a conversation AI strategy
- Define your business goals. Faster, more integrated communication and delighting customers are common aims for conversation AI.
- Who is your audience and what is your use case for conversation AI? Understand the parts of your business where an automatic process uses technology but perhaps still relies upon humans. Repetitive and menial tasks can be completed more easily, but you must understand the optimal channels and how best to communicate.
- Build or buy the platform? This is where language technologies may be introduced if you plan to build your own platform over time with experienced employees. LLMs can be incorporated into the strategy for many reasons - to talk to customers or to increase productivity for conversation designers by providing them with prompts and training phrases.
- Design for customer experience. If you have employees who are trained and understand conversation AI, customer experience will be at the heart of their daily goals.
At Conversation Design we’re regularly asked to provide our opinion on technology like ChatGPT, and whether they will make conversation designers redundant. Far from replacing people, AI systems will become a professional sidekick and exist alongside your trained teams of conversation AI specialists who can spot the pitfalls and errors. These new tools will help organizations be more efficient, more available around the clock and potentially scalable based upon demand and company growth if it’s part of a much wider strategy. In the hands of professionals, AI tools like ChatGPT can be powerful, but must be managed by humans.
To take a step further in understanding how ChatGPT and other LLMs can make you more productive in your conversation design, you can see our founder Hans Van Dam talk about it here. (https://www.youtube.com/watch?v=g-s3OXfoVjk)
If you’d like to find out more about how the team at CDI can help you and your team get the most from Conversational AI contacts us through the website or call the team on Call us +31 (0) 20 23 71 610.
