Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
Conversational AI is hot and getting hotter. Like other forms of artificial intelligence, conversational AI has been around for a while (in this case, since the 1960s). But a few recent developments have rapidly made it a point of focus for technology leaders.
ChatGPT, for instance, has shown how natural and flexible conversational AI can perform in real-time dialogue with users, both through voice and text. At the same time, announcements like Apple Intelligence have app developers imagining a future where Siri acts as a quasi conversational AI assistant.
But the excitement around conversational AI is matched by the misconceptions around it. Is it the same thing as generative AI? (No.) Does developing conversational AI mean you have to build a chatbot? (Again, no.) How do you know what use cases to start with? (We’ll get to that.)
These conversational AI best practices are based on our Data & AI Research Team’s (DART) client engagements here at WillowTree, which range from building secure conversational AI assistants for financial services to prototyping AI tools in our eight-week GenAI Jumpstart accelerator program.
Use them to shorten your learning curve and develop conversational AI solutions with greater confidence.
It’s easy to use conversational AI and generative AI interchangeably, but they're different types of artificial intelligence that offer different possibilities. That means they also pose different tradeoffs (e.g., security risks, ability to deliver your brand voice).
RELATED: How to Evaluate Conversational AI for Politeness
The better you understand the realities of conversational AI and gen AI, the less likely you’ll bias your thinking toward either. Instead, you’ll more objectively see which one best fits which use case:
Imagine you want to design a conversational AI assistant for your contact center agents. Conversational AI might offer advantages such as limiting your risk of AI hallucinations (because you can predetermine your responses). This level of control also 1) makes it easier to consistently deliver responses in your brand voice, and 2) highlights why developing AI tools for employees first is a smart move.
“Developing AI internally first gives you a much safer sandbox to play in than the consumer side.”
“Developing AI internally first gives you a much safer sandbox to play in than the consumer side," said WillowTree Senior Director of Product Delivery Stephanie Lewandowski. "It also provides hands-on experience you won’t get from just reading about AI. Then you can apply what you've learned to building customer-facing products.”
So what advantages would generative AI offer, knowing it also introduces risks such as hallucinations and vulnerability to malicious users? LLM-based systems are better at 1) matching responses to a broad, diverse range of user requests, and 2) tailoring these responses based on the entire context of the conversation. That’d be an advantage for human agents tasked with handling complex requests (e.g., customer service calls that take 15 minutes or longer to resolve).
In reality, developing an AI assistant could combine methods from both conversational and gen AI. For instance, when building a secure conversational AI assistant for financial services, we placed an automated check in the conversation flow by first passing requests through an intent classification layer. The assistant then decided to send the request to an LLM for response generation, or to a retrieval augmented generation (RAG) system with access to a custom knowledge base.
By understanding underlying technologies like these, you’ll lead projects more confidently and collaborate effectively with your dev team and AI consulting partner.
AI makes product development principles matter more, not less. To start, risks like AI hallucinations will probably never go away. A disciplined development process reduces risk, and reduced risk increases ROI.
But more concerning is how the hype around AI seduces many technology leaders to skip key product development steps, like starting with the problem. Instead, they go solution shopping. This approach results in lost time and frustrated teams because without a clear problem to solve, there’s no north star to unify everyone involved. It also means that whatever solution is developed, it will likely have low product adoption due to ineffectively meeting user needs or delivering a good user experience.
“Good AI products are just good products.”
Based on the brands we’ve partnered with, the technology leaders who consistently develop successful AI products do things like:
Strong product development practices like these bring a level of specificity where the most impactful and cost-effective AI solutions emerge. You can begin mapping specific conversational AI technologies (e.g., natural language search) to specific pain points (e.g., builders needing time-sensitive technical information in the field).
“Good AI products are just good products," said WillowTree Principal AI Research Engineer Nish Tahir. "The hype cycle tends to create a focus on the technology first rather than the problem it’s helping to solve. Be mindful of that. You won’t see robust product adoption unless you’re meeting a true user need.”
Here are a few products that started as a request to build a chatbot, then evolved into solutions that applied other AI techniques:
“It makes sense why a lot of directors and VPs drift toward thinking ‘I need a chatbot,’ because it’s the most familiar interaction point most of us have with AI," said WillowTree Delivery Lead, Data & AI Conner Brew. "Their C-suite is probably pushing a chatbot, too. But that thinking creates a lot of friction. It’s more effective when a VP keeps their dev team and AI partner focused on the most important problems to solve.”
“It makes sense why a lot of directors and VPs drift toward thinking ‘I need a chatbot.’ But that thinking creates a lot of friction.”
Notice how refined the above solutions are compared to a chatbot (and how awkward it would’ve been to force a chatbot as the solution). This is where AI shines: secure, cost-efficient, and problem-specific applications that drive productivity.
The better you understand different types of AI — and the closer you stick to product development fundamentals like starting with the problem — the easier time you’ll have not fixating on a solution like a chatbot. Instead, you’ll see it’s just one of many options AI offers to automate complex tasks.
At WillowTree, our data and AI consulting services span the entire development lifecycle. In addition to end-to-end strategic advice, we offer hands-on development to help you bring new AI solutions to market faster. That includes identifying the right types of AI to develop across your enterprise.
Our approach has led to results like TELUS launching the world’s first gen AI-powered chatbot internationally certified in Privacy by Design (ISO 31700-1), a milestone indicating the highest standards in privacy and data protection.
Bring your own responsible AI solutions to market safer and faster with help from WillowTree. Learn more and connect with us through our conversational AI consulting services.
