The underlying premise of this question is whether large language models (LLMs) like ChatGPT will transform the reputation of chatbots from clunky, impersonal and faulty into algorithms so meticulous that (a) human interaction is no longer needed, and (b) traditional ways of building chatbots are now completely obsolete. We’ll explore these premises and give our view on how ChatGPT will impact the CX space.
Broadly speaking, we differentiate between conventional chatbots and chatbots like ChatGPT built on generative LLMs.
This category includes most chatbots you’ll encounter in the wild, from chatbots for checking the status of your DPD delivery to customer service chatbots for multinational banks. Built on technologies like DialogFlow, IBM Watson or Rasa, they are limited to a specific set of topics and are not able to respond to inputs outside of those topics (i.e. they are closed-domain). They can only produce responses that have been pre-written or pre-approved by a human (i.e. they are non-generative).
These can respond to a wide range of topics (i.e. they are open-domain) and generate responses on the fly, rather than just selecting from a pre-written list of responses (i.e. they are generative). They include Google Meena, Replika.ai, BlenderBot, ChatGPT and others.
LLM-based chatbots and conventional chatbots fulfill somewhat different purposes. Indeed, for many CX applications, LLMs’ open nature is less help and more hindrance when building a chatbot that can specifically answer questions about your product or help a user with an issue they’re experiencing.
Realistically, LLMs won’t be let loose into the CX domain tomorrow. The process will be much more nuanced. The name of the game will be marrying the expressiveness and fluency of ChatGPT with the fine-grained control and boundaries of conventional chatbots. This is something that chatbot teams with a research focus will be best suited for.
There are many aspects of chatbot creation and maintenance that ChatGPT is not suited for in its current state, but here are some for which it is already well-suited:
Brainstorming potential questions and answers for a given closed domain, either on the basis of its training data, or fine-tuned on more specific information — either by OpenAI releasing the ability for fine-tuning when ChatGPT becomes accessible by API, or through including desired information via prompt engineering. (Caveat: It is still difficult to know with certainty where a piece of information comes from, so this development process will continue to require a human in the loop to validate output.)
Training your chatbot: ChatGPT can be used to paraphrase questions a user might ask, particularly in a variety of styles, and even generate example conversations, thereby automating large parts of the training.
Testing and QA. Using ChatGPT to test an existing chatbot by simulating user inputs holds much promise, particularly when combined with human testers. ChatGPT can be told the topics to cover in its testing, with different levels of granularity, and, as with generating training data, the style and tone it uses can be varied.
We see the next generation of CX chatbots continuing to be based on conventional, non-generative technology, but generative models being used heavily in the creation process.
LLMs’ key impacts on consumer expectations will include increased visibility of chatbots, greater urgency to incorporate them into CX, a heightened reputation for chatbots and a higher standard. In other words, chatbots are getting a glow-up!
We’ve all experienced them — clunky chatbots with extremely limited dialogue options that churn out painfully robotic lines (if they can understand anything at all). While poorly performing chatbots are already on the way out, standards will now be shooting through the roof to avoid this experience, and the shift from human to AI will rapidly continue.
A recent report predicts that the number of interactions between customers and call centers handled by AI will increase from 2% in 2022 to more than 15% by 2026, then double to 30% by 2031. However, given the rapid adoption of and exponential advancements in AI over the past three to five years, we anticipate the true growth to be far greater.
Brands like Lemonaid, Oura, AirBnb and ExpressVPN have paved the way for excellent 24/7 support — so much so that today’s customers now simply expect a seamless experience. The consequences of missing out on delivering great service are no joke. Poor service can have a significant impact on a brand’s retention rates, causing would-be buyers to look elsewhere: According to Forbes, bad customer service costs businesses a combined $62 billion each year.
ChatGPT is certainly in a hype phase, but there are significant risks in using it as-is right now. We believe that the majority of the current risks result from ChatGPT’s unpredictability, which creates reputational, brand and legal concerns. Whilst the buzz around ChatGPT is good, you must not forget its associated risks, and the importance of selecting the right partner to avoid any pitfalls.
In particular, we see the following risks for big businesses adopting LLMs directly into their customer journey:
Harm to brand image — sharing of offensive content
Misleading customers — sharing false content
Potential for adversarial attack — people trying to break the chatbot to damage reputations
False creativity — users mistaking the “stochastic parrot” for genuine human creativity/connection
False authority — ChatGPT produces authoritative-sounding text which humans are notoriously bad at refuting.
Data security and data ownership and confidentiality — OpenAI has insight and access to all data shared via ChatGPT, opening huge risk floodgates for confidentiality breaches.
In other words: “Just because you can doesn’t mean you should”
Startups and established organizations will inevitably try to introduce safeguards and other measures to mitigate some of these risks. However, a lot of companies, including many of those we work with, still want (or are legally obliged) to retain full control of the content. Our legal and FCA-regulated clients are a good example. With generative LLMs like ChatGPT retaining full content, control is impossible.
When it comes to chatbot development itself, players using open-source stacks like Rasa or Botpress will have the advantage of agility due to the flexibility and versatility these open systems enable. In the short to medium term, chatbot developers with experience in NLP and using LLMs will be the ones to bring this technology to the chatbot market, because they are able to effectively leverage and fine-tune the models to their (or their clients’) needs and use cases.
In the long term, small companies will continue to be better positioned to swiftly implement changes than large, established platforms like ChatGPT. Amidst the current financial market volatility, however, we anticipate a potential market consolidation of players in the next 12-24 months, with the larger players acquiring smaller players, and — a common occurrence in the chatbot space — clients buying their chatbot suppliers.
Despite ChatGPT only being in beta and no API yet available, there has been a myriad of exciting use cases published by individuals, including a number of browser extensions, mainly via Twitter.
As long as ChatGPT is available to the public (we expect a volume-based pricing model to come, as with previous models like GPT-3), small players will continue to be the ones pushing the boundaries with novel applications.
(Copyright: VentureBeat ChatGPT and its implications for customer experience | VentureBeat)