5 Golden Rules for a Successful Conversational AI Application
Look at the key aspects for ensuring a positive result on your conversational AI journey.
Join the DZone community and get the full member experience.Join For Free
5 Golden Rules for a Successful Conversational AI Application
In the fourth post in our series of how to get started with conversational AI, we take a look at the key aspects for ensuring a positive result on your conversational AI journey.
We covered the topic recently in a webinar given by our VP of Global Customer Services, Darren Ford and below is a partial transcript highlighting his 5 Golden Rules for Success.
If you'd like to listen to the webinar in full including examples of how these golden rules have been applied in real-world implementations then you can watch the replay here.
1. Prioritize the Business Case
If you can't articulate the business case and value you want to achieve from the conversational AI application, don't start the project. Or at the very least, be transparent with the partner or third-party supplier and work collaboratively with them to help define the value you expect to achieve.
The idea is to use this as the lighthouse to guide you towards business success and to help you make design decisions in order to deploy the right technology. Prioritizing the business case helps to focus on the right goals when venturing into the unknown.
By knowing the features you need to develop in order to achieve the desired end result, you can then shape the implementation, bearing in mind any business restrictions such as time or budget.
Whether it's a proof of concept, a pilot, or a full production project it's important to stay true to these goals before moving on to other phases within the project. Otherwise, you may be distracted by cool features that aren't really necessary to achieve your goals and potentially make the end application below par.
2. Consider User Experience
The user experience will materially impact the adoption rate of the solution, the revisit rate and metrics such as net promoter score or customer satisfaction.
You can have an awesome dialogue and integrate data to deliver personalized and relevant answers, but if the user experience isn't what it needs to be then you won't get great results.
To achieve this, a number of factors come into play.
It needs to speak like a human. But don't pretend to be one. People expect digital employees to have intelligence, memory, context, and maybe even participate in small talk. However, passing a chatbot off as a real human is likely to end in mistrust of your solution by the end user.
Be very clear about the state of your application's knowledge. In order to meet expectations, you need to set them correctly. If a user can't achieve the results they need and have to find another channel in order to resolve their issue, they will not come back to the conversational application. They will go back to the alternative channel that served them better.
For this reason, make sure the user understands the capability of your intelligent virtual assistant and provide early on in the customer journey the right links for information beyond the scope of the chatbot.
Don't waste the user's time. Take advantage of data and integration to be more efficient than traditional channels and give users an experience they can't get any other way.
Be on brand. Work hard to have the tone of voice correct for your corporate identity. Focus on the personality you want for your chatbot. Is it humorous, sassy, formal, or authoritative? By being on brand and adhering to the company's identity you will find in subsequent phases with additional regions or business divisions that the internal adoption will be taken up more readily. This makes it easier to achieve the ubiquitous deployment of your application, which in turn allows for a greater return on investment.
Use the intelligence of the AI platform, the context and the understanding that it has to drive an effective user interface. Keep it simple, using text for example, where simplicity works best. Make it rich using other media such as video, only where it makes sense. Consider that the user experience will be different over the various channels and devices and derive the right experience for the consumer.
Of course, it needs to look good too. It must appeal to the user and meet modern design techniques and expectations. Sometimes that might mean including channels that aren't yet in your vision, such as Facebook Messenger or WeChat.
Finally, ensure you draw to a conclusion at the end of the journey so users don't go off and try to ask the same thing of another channel. If you're building a pilot, cover a particular part of the user journey from beginning to end. Focus on a small part the business or a particular region with an expansion strategy to broaden the knowledge. Starting with a narrow domain and going deep to the conclusion is better than delivering broad and shallow, which often only forces only the user to go through a different channel.
3. Choose a Scalable Platform
If you're venturing into a new domain or channel, you may not be able to articulate fully all the features you need right at the beginning. Choosing a scalable platform will give you options moving forward.
Think big. Start small.
Think about what you want to achieve with this new technology, but start with a smaller project in order to see the results and measure the success before deciding on next phase. But in order to achieve the next phase and capitalize on your initial investment, you need a scalable platform.
For example, you may have built the application in one language but being a global enterprise you want to deploy in other languages. How easy is it to take the investment you've already made in dialogue flows and integration, including the tone of voice and branding, and reuse it in a different language? Does the conversational development platform support the languages you might need? Many platforms say they cover multiple languages, but often this really means a new build for each one.
How many intents can the platform handle? Think of an intent as something the user wants to achieve. Various conversational AI platforms handle different numbers of intents with different algorithms to work with those intents. A pilot or proof on concept may have only 10 or 20, making it easy for the application to learn those intents with machine learning. What happens when it's hundreds of intents over multiple business divisions, languages, and regions?
Some of the API is driven machine learning solutions today are only capable of managing tens of intents at a time. Some of the more enterprise-focused platforms can deal with hundreds of intents, but often there is a hard stop on the number of intents, leaving enterprises having to build multiple solutions to cope. However, it becomes very hard to manage those intents on a machine learning basis only, because each time you apply new training data so the understanding changes making it hard for the application to be precise with its response. This is one of the reasons Teneo takes a hybrid approach using both linguistic and machine learning techniques.
Can you apply your own data and vocabulary into the conversational flows, and how quick and easy is it to do that? It might be that you want the consumer to personalize a product such as naming rooms within the smart home.
As well as using dialogue components within multiple conversations, can you reuse them and the conversational logic across different channels? In other words, how easy is it to take say an application built for a website and use it across Google Home, Alexa or Facebook Messenger?
As the application is expanding across different platforms and regions and developed further by different teams, does the conversational platform support collaborative activity including the reuse of dialogue components built by other teams? Other enterprise features to look out for include rollback, versioning, and the capability to lock records on particular aspects as they are being developed.
We often recommend that proof of concepts are run as SAS based solutions hosted by a third-party such as a partner or software provider so you can quickly get up to speed and see those proof points. However, it may be in the future you want to take the solution in-house to your own data center for security, privacy, or data integration reasons. Few conversational AI development platforms are able to offer self-hosted options, so it's important to check first if this is likely to be an important requirement for your business.
Finally, similarly to the issues with multiple intents, you may have a narrow field of knowledge where you need to be very precise. That may be for legally compliant reasons such as providing financial advice. The challenge with a narrow band of knowledge and yet needing to be very precise is you can't just use machine learning. To ensure a precise answer in a narrow field you will need language rules or conditioning.
To teach a machine learning platform the difference between "What's the weather is like in Barcelona?" or "Tell me a joke", is very easy. It doesn't require much data and is simple to annotate.
However, if you take, "You canceled my flight, can I get a refund?" or "I canceled my flight, can a get a refund?" these are very different conversations and potentially one where machine learning only systems will not understand the subtle difference.
If you're looking to ensure a precise response is given in certain situations, then you need to ensure the conversational AI platform is capable of delivering it.
4. Never Compromise Customer Data
First-person conversational data is very valuable to the business. While you might have hundreds of thousands of calls to your call center every day, it's not as easy to access that data in a way that it can be tagged, analyzed and understood as it is with conversational AI data. The information gleaned from your customers' conversations is like gold dust.
However, it's also very dangerous. You need to look after the data and privacy of your users. You absolutely cannot afford to compromise that data because you will lose your customers' trust in an instant.
Enterprises need to consider their options. It may be that your application is providing very general information and there is no opportunity for data misuse. Or it may be a banking app and you need to need to worry about the data collected.
You need to be able to anonymize or pseudonymize conversational data. To be able to replace identifiable data with placeholders such as customer_email_address, so you can still understand the intent for analytics purposes, but not know the customer identity.
On the other hand, you may need all of the identifiable data because you're carrying out a transaction. In which case you might want to encrypt that data, before sending it over the internet and into your systems.
Before the conversational AI application is built, an enterprise must consider the various situations in which information may be collected in order to ensure it can make use of the data, without compromising customers' privacy.
5. Recognize that Going Live Isn't Last Base
Going live is just the beginning.
Following the processes above you now have a clear idea of the business value you want to achieve, and you know the things you want to measure. You've got an awesome and complete customer journey. You've already reduced the number of calls coming into the contact center and have a goldmine of customer data by knowing exactly what your customers are saying.
You've got a 24/7 non-sleeping, intelligent virtual assistant running in multiple languages, over multiple channels. Congratulations! You've launched.
But the solution can learn. This is only the beginning of the value you can obtain. It can always do better and increase customer satisfaction even further.
With all this information at your fingertips you know if you're having a positive conversation or a negative one. You know what people are asking, what they want. But you also know what they can't get so you can improve, enhance and expand the knowledge to make the experience even better.
It's important to make provisions to provide continual and continuous improvement to the system. It doesn't have to be time intensive. Much of the process can be automated through machine learning or by reporting directly into corporate dashboards.
At the same time, it is also essential to have KPI reporting in place and to use the traditional measuring methods already used by the organization, such as first call resolutions rates.
By enabling the conversational AI application to continue to learn and improve, so you will increase the value of the overall solution.
To discover more about the 5 Golden Rules for a Successful Conversational AI Application and for additional information in the Q&A session at the end of the webinar, follow this link.
Published at DZone with permission of Andy Peart, DZone MVB. See the original article here.
Opinions expressed by DZone contributors are their own.