AI came about as big data, powerful algorithms, and cloud hosting reached a level that could support the need for lots of data, processing data quickly, and running on demand.
Microsoft’s AI solution includes image recognition, language and speech, information search, voice recognition that achieves 5.1% accuarcy (the best human is 5.9%), deep learning.
The MS Bot Framework for enterprises includes payments, integrated API’s Q & A built from website FAQ’s, analytics, and can be surfaced on many channels.
Bing can serve chatbots related to search terms to help provide users answers questions on the SERP’s rather than going on to webpages to find information.
Age UK – Contact Support Chatbot
Age UK get 26,500 calls a year to their national advice line, each call lasts an average of 4 minutes, and 30,000 of those go unanswered. They developed their chatbot to try to answer some of the simpler questions and reduce the strain on the call centre with the thinking that chatbots can offer a 24-7 customer experience.
Some of the things they learned included:
- Most people don’t know what a chatbot is.
- People expect a chatbot to be able to answer any question, even if it’s not including in the purpose of the bot. The bot needs a means of filtering these out.
- People talk to chatbots in all kinds of strange (human) ways, often telling their whole story before asking a question, make spelling mistakes, and use metaphors and similes. This makes it difficult for the bot to understand the intent and so answer appropriately.
- Developing and maintaining bots takes lots of time and resources. They need to be constantly updated as you learn from user interactions.
- Bots should sound human, but not too human. They should have some personality and use emoji’s and GIF’s, just as people do when chatting.
- Know what success look like. 18% of conversations ended with the user going to the safety net to speak to a human. The target is 5%.
When surveying users they found that 33% were happy with their interaction, 48% were indifferent, and 19% were unsatisfied.
Cancer Research UK – My Alcohol Tracker Alexa Skill
Cancer Research UK Digital Innovations Team are looking at how they might be able to use AI, and trained an Alexa Skill as an Alcohol Tracker to help build awareness of the effects drinking too much alcohol has on cancer rates as part of Alcohol Awareness Week.
They mapped, sketched, designed, prototyped, and tested in a five day design sprint. This took some shift in thinking for them as they moved from thinking in screens to thinking in user intents. They learned that voice interaction needs to be very concise, and that the technology can struggle with different accents and how people say the same thing in lots of different ways.
This is Alcohol Tracker being tested in CRUK shops:
What I learned
People will expect to be able to ask anything, in any way, and get an answer
No chatbot can answer every question, but people will expect it to. Not everyone understands that a chatbot doesn’t understand messages in the same way as a human. Even if a chatbot is very clear with the user about what it can and can’t do they will still ask all kinds of questions. The only way to handle this is to provide a escape route so that users can be passed to a human if the bot fails to understand a certain number of messages in row.
Natural language processing is hard
Allowing people to type anything to the bot makes handling any message extremely hard. If a message mentions multiple things that the bot regards as intents then how does it judge which intent to answer? Developers can’t possibly predict and programme for every eventuality, but they should prepare responses for users swearing at the bot and telling it that they love it.
Chatbots work best when they are simple up front and complex behind the scenes
Simple bots that use buttons give a clearer indication to the user about what the bot can and can’t do. Asking questions to the user can help to define the context and understand the user’s intent.