Covid-19 accelerated the adoption of virtual assistants (VAs). Today, we use the terms “chatbot,” “conversational artificial intelligence” (AI), and “virtual assistant” interchangeably. But what we have in mind are no longer those “rule-based” systems that ask users a million questions and wait on their selections to move forward. We are imagining AI-powered applications and well-connected information systems that allow humans to interact naturally and dynamically with devices, machines, and computers using speech or text. These applications and systems help pass information and context back and forth between humans and sources such as websites, APIs, and databases.
In response to the spike in volume at their contact centers, many companies across industries have started to explore and offer VAs as an additional communication channel. This happens especially in areas like hotel bookings, insurance claims, search and FAQ interfaces, and general customer support. It is exciting to put a chatbot into production. But, it is challenging to develop a successful chatbot, and it’s not uncommon for the product team to hit a plateau during VA product development. After all, building conversational AI is different from building traditional software products. That makes it especially important to understand conversational AI best practices.
Check out these 10 best practices to make sure your organization is taking the right approach to conversational AI. These practices come straight from Slalom’s own experience with clients, and they can be broken down into three categories:
- Start simple and iterate fast
- Drive cross-functional, cross-team collaboration
- Continuously upgrade VA product & validate ROI model
Start simple and iterate fast
1. Start with a FAQ bot and mature your VA over time.
The best way to approach and build a functional bot is to start with a prototype — a basic bot that answers the frequently asked questions (FAQ) from your end-users. As a minimum viable product (MVP), the FAQ bot should provide static, navigational information to users and an “escalation” path or fallback flow to handle the unexpected questions and/or use cases. Starting with your MVP, consider how well the technical architecture, AI model, and conversational experience meet your end-user goals. A more sophisticated bot can include integration to different databases and production systems to retrieve personalized information for users, replacing the one-size-fits-all bot responses and increasing credibility. To make VA a preferred channel for users on par with a human agent, your VA needs to be able to drive transactional interactions and complete workflows, enabling persistent changes in your production systems. The ultimate version of a disruptor bot should leverage AI and automation to prescribe solutions and intelligently suggest the next best actions for users, delivering a superior user experience and becoming the brand differentiator.
2. Leverage analytics to drive continuous product improvement.
To better understand user persona, user goals and decide when and where to invest development efforts, the product team should consider designing and implementing a bot monitoring and VA performance evaluation system. A comprehensive list of data should at least include bot technical metrics, model performance, product analytical metrics, and user feedback, including surveys. It’s equally important to track the aggregated KPIs and periodically review the user/bot/agent conversation transcripts to understand end-user needs and overall conversational experience.
3. Select high-value use cases to build out experiences.
The primary goal of a VA should be serving its end-users. The business stakeholders and product owners can prioritize a VA backlog based on the user need(s)/impact and the level of development effort. In practice, the product team should focus on and prioritize the high-volume, high-escalation cases with a low or moderate level of development effort, then begin tackling complex use cases in an iterative approach. For a chatbot, it is nice to have some features that help human agents to be more productive, or that help the business promote brand and/or marketing campaigns. But because these features don’t directly serve end-users, they should not be the highest priority for developers.
4. Set the right target for different intents.
There are different levels of VA maturity, and then there are different types of chatbot intents. Intents reflect the variety of tasks that a user might want to complete. An example of a knowledge task might be a user seeking answers through the FAQ. An information task is one in which the user seeks a more personalized and on-demand response. An action task drives transactional results and persistent changes via a guided, long-form conversational experience, which appeals to users by being more efficient than other channels or products. In addition, there is the dialog task, which addresses the need to handle nested/multiple intents, support fallback and escalation, and process basic conversations like greeting and closing. When designing an intent, the product team should clearly define the types of tasks involved and the relevant user base/persona. This will help set the right target for the intent and conversational experience to build.
5. Be agile. Test and iterate early and often.
Trying to conduct exhaustive user research/use case discovery, going straight ahead for an Action type of task, or setting a goal to deliver a disruptor bot first are proven less effective in developing a successful VA. The nature of your business could change (e.g., think about emerging user questions that appeared after COVID). Major advances in technology happen faster than ever (e.g., the basis of the current popular language model/algorithm was created and published less than 3 years ago). And, if your VA is already in testing or production, your end users are probably providing an enormous amount of feedback, which should be taken into consideration. The best way to embrace the uncertainty in the VA development process is to stay agile, pivoting fast while maintaining a roadmap and backlog for future product enhancement.
Drive cross-functional, cross-team collaboration
6. Maintain a growth mindset.
Even though conversational AI may be one of the most mature use cases in AI applications, the technology and development methodology are still in their infancy. As compared to other traditional software products, you will likely need a full product team and cross-functional members to be able to deliver a successful VA. Software/machine learning engineers, data scientists/analysts, conversational designers, product managers, and domain experts all play important roles in shaping and developing the VA product. This team needs to work together to align on user/development goals, as well as understand how the AI model is plugged in, how chat messages are bounced between bot and user, and what the technical constraints are. That requires the entire team to develop and keep a learning mindset to stay on top of the tech trends and progress to be able to make the best decisions for product changes from a technical perspective.
If the product team is new to building AI products, it should also be prepared to adapt its approach in design, development, and testing for VA, an AI-enabled product. For example, the conversation designers need to translate complex business processes into concise conversational experiences and understand when to plug in conversation, or when to leverage AI, backend scripts, or other UI elements to deliver the best experience. Then, the data scientists need to help the entire team understand the right way to architect and implement AI in the product, acting as a bridge between user experience design and tech implementation. And then, the engineers need to consider how to test the implemented conversational flows at different development stages, since using AI as a core service component can sometimes yield unexpected results.
7. Let customer experience, AI, and automation go hand in hand.
Using a state-of-the-art AI algorithm or fine-tuned the language model alone does not guarantee a performing VA. A successful VA product requires a mix of solid technical implementation to keep the bot running and responding, an intelligent AI model to identify the user intent and recognize key information at scale, and a well-designed user experience in a conservational format to engage and help the end-users. Instead of paying attention to and concentrating the investment in one area, the product team needs to consider how to balance the different development goals through close collaboration.
Continuously upgrade VA product & validate ROI model
8. Review tech architecture periodically and decide when to refactor code.
As your VA product matures and your team gets more in sync through the development cycles, it is a good time for the product team to tweak the existing technical architecture to enable more AI capabilities. This allows for better availability, efficiency, scalability, and reliability for the VA product. The team can also explore and consider different language algorithms, chatbot development frameworks, or architecture to support more sophisticated user or product needs. Sentiment analysis, semantic search, image recognition, and prediction/recommendation are all popular and mature AI capabilities that could drive better user experience if incorporated and implemented appropriately in the chat experience.
9. Consider when to expand the team and adopt a new operating model.
When you have too many intents for your VA, it is natural to think about converting it to a mega bot or using a dispatching model to expand one chatbot to a family of VAs. Most organizations start with a centralized product team to ensure strong governance and to support different lines of business. So, when you think about shifting and scaling your technical architecture, it is also a good time to evaluate a different operating model. A hybrid or federated model could allow the different business units to design, build, and own the capability and better accommodate the various business needs.
10. Manage impediments and eliminate blockers proactively.
Depending on where your organization is on various fronts, developing a successful VA product in-house could be a business that requires high investment before yielding high returns. In addition to keeping users’ pain points in mind and managing changes while developing VA, you may also want to work with other business stakeholders to proactively:
- Understand the organizational/business process
- Improve customer-facing content quality and the user experience of your core business services
- Explore points for database/system integration
- Support the API development and infrastructure upgrade
Ultimately, the goal of these activities is to drive and accelerate the digital transformation of your organization. Doing so will pave the road for more accessible and intelligent VA.
About the author
Alice Chen is a Principal Data Scientist, AI/ML Solution Architect at Slalom. With an educational background in Applied Mathematics and Advanced Analytics, Alice focuses on data & analytics in real-life applications and has served clients across industries including Hospitality, Consumer Products, Financial Services, Healthcare, and Telecommunications. Alice is not only a certified conversational AI developer and cloud solution architect, she is also a thought leader and subject matter expert in designing and building AI products.