Done well, chatbots enable better customer self-service and improve customer satisfaction. However, a good implementation requires some knowledge of how chatbots work, and you need to get a few things right if you want to get the best out of your new virtual employee.
Natural language understanding
Even though there are infinite ways your customers could ask the same question, most humans understand language well enough that they can figure out the intent of what’s being asked. Traditional computer programmes can’t do this because each variation would need to be written in the code in order to be recognised.
For example, different ways of asking the same thing:
Hello, what’s my rent balance please?
How much rent do I owe?
What’s left to be paid of my rent this month?
Hi, my name is Julie, can you tell me the outstanding balance on my rent account?
Your chatbot uses artificial intelligence that recognises language patterns and can be trained to figure out the query’s intent regardless of how it’s been asked. If trained well, your chatbot can also deal with text-speak, lack of punctuation, misspellings, some people remembering their manners while others might not feel the need to do so.
It’s not a search engine
It’s a common misconception that chatbots just need a single keyword such as ‘rent’ to understand the intent of the user’s question, much like a search engine would. While it’s acceptable for a search engine to provide dozens of possible search results, for a conversation with a chatbot to feel natural, it typically needs to provide a single response or at most a few options that the user can select from. This is relatively easy to achieve if you only need your chatbot to answer one or two rent-related questions. But you’ll probably want your chatbot to be able to provide a unique response to a good number of rent queries, such as:
Can I get a new rent-card?
Is there a rent holiday this year?
Is there a rent freeze because of coronavirus?
Can I pay my rent via direct debit?
What should my mum do if she can’t afford to pay her rent?
Like a human, your chatbot can’t guess what the customer wants if they only type in a single word such as ‘rent’. Your chatbot uses a mathematical process, sometimes called an algorithm, to predict the intent of the customer’s query. It will assign a percentage likelihood, or confidence rating, for each of the potential intents and respond to the most likely one.
Some chatbot jargon
A good chatbot should handle the variety of ways tenants might ask the same question, and so you’ll need to train it to do so. We call the query the customer is trying to convey the ‘intent’. And we call the variations of the original question the ‘utterances’.
You’ll need to tell the chatbot the response you want it to provide for a specific intent, and these might change over time. For example, as we move in and out of lockdown due to coronavirus, the response to “I’d like to book a repair for my dripping tap” will change.
When training your chatbot, you’ll need to provide around 20-30 utterances for each intent. A chatbot for social housing can easily have over 400 intents and 10,000 utterances.
And this is where AI comes into its own. From this relatively small number of variations, chatbots use AI to identify patterns and create a neural network so that it can recognise a much broader set of utterances. Even if your chatbot hasn’t been trained to recognise the exact set of words or phrases that a customer uses, a well-trained chatbot will still be able to identify the intent. Once it knows the intent, it can provide the corresponding response that it’s been trained with.
How to train your chatbot
If you buy a chatbot product or platform, you have a few options to train your chatbot with the intents and utterances you need it to know:
Create them from scratch;
Extract them from customer emails and/or support-call transcripts;
Buy a pre-trained, sector-specific chatbot.
If you’re building your own chatbot, rather than buying a platform or product, and have access to an extensive data set of utterances and appropriate responses, you can use machine learning so that the AI can train itself.
Training your chatbot to recognise the correct intent accurately takes a lot of effort, and a lot of testing to get it right and so buying a pre-trained chatbot, where all the hard work has already been done, can make a lot of commercial sense.
Beyond the utterances and intents, there are many other things you’ll need to do to tune the AI. Things such as pluralisation, spelling errors, text speak and recognising entities – such as different rooms in a house, different types of units and various family members also need to be taken into consideration.
To get the most out of your chatbot, it’s important to invest the time to train it properly. Anything less will result in customers not getting the responses they need and a frustrated call into the contact centre.
Scott Summers is the co-founder of Fuzzlab.