A Presentation at TADSummit Paris by Webio Chief Stratgey Officer, Paul Sweeney
How we interact with AI and each other is making a big leap at the moment from ‘communications’ to ‘conversations’. What does this mean and how will it play out?
When voice activated devices like Alexa came on the scene, we were enthralled at how smart they were. It was a piece of genius for its time. While revolutionary in that they could understand instructions, these devices struggled with the nuances of true conversation. They excelled at tasks but faltered in maintaining context over time, in what we call long-term short-term memory. If you told it something a while ago, it wouldn’t really ‘remember’ it.
Enter ChatGPT, a precursor that kicked our grand plans in the nose, forcing us to recalibrate and adapt. Like Mike Tyson said: “Everyone has a plan until they get a punch in the face.”
Fundamentally, the world of conversational AI has undergone a significant shift in what some call a "Netscape moment." It's not just a new web story; it's a new API, a fresh interface. The grand narrative here is moving from just connecting the channels to connecting what's happening in the channels. This marks a pivotal shift from merely connecting channels to comprehending and connecting the content within those channels. It’s calling up a new service into a new UI.
For example, before stepping into a Sales meeting you may ask the AI, "I'm going to a Sales meeting. What likely objections am I going to get?" The AI then gives you a couple of things to be concerned about. It could also include product information or details of a likely contract.
Another example: from a Zoom meeting you could gain insights like, "How come Dan doesn't speak as much when Paul is chairing a meeting? How come one person speaks 30% of the time?”
Despite initial misstarts, conversational AI is directionally correct. People were willing to embrace a new way to interact with technology, as evidenced by the adoption of devices like Alexa. However, were just a behind in our capability of delivering it.
LLMs, for all the talk of what they can do and what they can't do, they are good disambiguators. They're good at figuring out what the conversation is about. And that was the problem at the front end of your Amazon device. It couldn't figure out what the conversation was about. The LLMs excel in figuring out the context of conversations, a critical improvement in the user experience.
LLMs are going to be masters at discovering third-party services. How this will pan out is unclear, but it could go with approved and discoverable LLMs that you can pull and call up.
The potential for LLMs extends beyond mere discovery; they're poised to usher in dynamic integrations that occur on the fly. Envision a future where companies utilise custom LLMs, trained on their specific conversations, offering a nuanced understanding of organisational dialogues.
We're moving from a text-first experience to a multi model one where we can use LLMs and generative AI to solve problems from visual inputs. For example, now with ChatGPT, we’re able to upload images and query the LLM. You can, say, load a picture of a puzzle and ask, “How do I solve this?”
Instead of simply typing in keywords, with AI search we could ask for more detailed responses. For instance, before a trip you could say, “I'm going to Paris. I'm arriving Monday next week and I'd like to see some interesting things. Can you organise a schedule for me?" And the AI generates a prompt chain and pushes through various technologies and services.
In the near future, the Intelligent Assistant will become the core interface, and every person in an organisation will have some form of AI co-pilot and that’s how we’ll experience the next user interface.
The infrastructure supporting this advancement involves Natural Language Understanding (NLU), Large Language Models, dialogue managers, and rules engines working in tandem. Every company will have a Customised Language Models trained on conversations that happen in their company. This approach ensures control and a deep understanding of organisational nuances.
However, the complexity is undeniable, with channels, knowledge bases, enterprise data, external data, and actions converging to control the CLM. Despite the intricacies, the end result is a user-friendly experience, offering concise summaries and a visionary shift in how backend processes shape the interface's functionality. The future promises a smoothly integrated and visually enhanced conversational experience driven by sophisticated backend systems.
Ultimately, the CLMs are going to be based open source LLMs (not just one, but many) that can be customised and trained.
To conduct a conversation between an AI chatbot and a customer, you need information matching, interaction enablement, and transaction facilitation. Even checking a simple query like “When is my next bill due?” requires many different systems to work together.
The trick with generative AI (GAI) is to remember the 'generative' part. Generative AI is great at creating options but falls short in transactional and interactive capacities. While co-pilots offer great experiences, their generic nature often leads users to contemplate disabling them to protect their data.
Generic generative AI is also not that useful for individual company tasks. However, the closer you are aligned to a person's specific role and help them do a specific job, the more chance you have of GAI being sticky.
Generative AI's strengths in copy generation, summarisation and labelling are well-established. However, at the enterprise level, the demand for oversight prevails. In a contact centre situation, companies prefer human review before content generated by AI is sent out to customers. There remains, and rightly so, an emphasis on maintaining control and privacy in data handling. This hinders enterprises from freely using AI capabilities, and the lack of transparency, and the uncertainty about the handling of data, pose challenges which leaves the trajectory of these advancements in question.
Voice data's richness, offering more than just words, presents both opportunities and challenges. Voice interaction has still been, to date, instruction-based, like "Turn off the lights", but it hasn't been great at interacting. AI behind IVRs should work well, but at the moment, it’s not getting adopted, and voice conversational IVR does not seem to be driving down overall call volumes.
However, voice data carries 10x richness. It has so much more in the voice pattern itself than just the words they're saying. And with that comes the ability to know that when someone says, "I'll pay that by Friday," and you know by their tone that they will surely not pay by Friday!
Language is hard, but if you can unlock it, it's got 10x the data potential.
On the flip side, voice technology can be used negatively. Voice cloning is fast and accurate, and it works. With this tech, you're vulnerable to social cloning and a myriad of fraud issues.
Voice conversations are a natural match with the metaverse; it's a natural match for AI and VR. They're made for one another. So, expect to see a lot of developments in the entertainment business. But from experience to date, nobody's doing customer service in the metaverse.
Some thoughts and questions to consider:
The personal assistant co-pilot as the future of the customer interface: If you believe the personal assistant or the co-pilot is the interface of the future, what does that change for your company?
Communications data, and knowledge base data, are both required to facilitate deeply personalised customer conversations: Companies must grapple with questions of data access, partnerships, and a technology stack capable of real-time processing at scale.
Deep data driven insights into 'everything' that could matter: What is happening in the conversation and what are the outcomes?
The need for customer intimacy to deliver on conversational AI: You need to understand what a company is about and what they are trying to do.
Technology stack performance at scale for real time processing: The platform oversees the end-to-end journey and architects the necessary components.
Verticals have own workflows, data, integrations, and models: Companies aiming for a superior customer experience must grasp not only the intricacies of those companies but also their workflows. This reveals where integrations are needed to deliver the full service.
Where do you partner, with whom, and for what reason?
Contrary to headlines predicting widespread job losses due to automation, the reality is different: nobody is getting automated out of a job. Automation has streamlined basic tasks, allowing a shift towards more complex and meaningful responsibilities. Focusing on the "jobs to be done" philosophy remains key.
Let’s illustrate with a quick example. In a large company handling credit collections, they spend time organising payments, etc. which all takes up agent time. Now, imagine reaching a point where you can identify vulnerability in a person using AI through subtle cues in conversation. For instance, a customer may hint at personal challenges during a routine inquiry, like “I'm vision impaired,” or a family issue. Now, you'll need to get this person to an agent.
This vulnerable customer requires human interaction, where a human can understand messy lives, and responding appropriately. No one is losing their job; instead, the focus is on the old-fashioned concept of "jobs to be done." Figure out what job your service is meant to do for customers and the figure out how AI can assist that task.
People are starting to interact with their systems and are figuring out how to think differently.
If you need to improve your customer engagement, talk to us and we'll show you how AI automation via digital messaging apps works.
You will love the Webio experience.