Insights: Blogs

Dr. Mark K. Smith
Dr. Mark K. Smith CEO

‘Train a machine to recognize sheep and it will do a respectable job – until you show it a picture of sheep in a tree and then it will assume they are birds.’

My recent posts have focused a lot on what AI can’t do. It isn’t magic, it won’t take over the world, it won’t write a great novel. It would struggle to write a bad novel, come to think of it. Jeffrey Archer’s career is safe, sadly. Still – as someone who runs an AI company, it’s probably time I pointed out there are lots of useful things AI can do.

Mining data
Machines are only as clever as the data they receive. Train a machine to recognize sheep and it will do a respectable job – until you show it a picture of sheep in a tree and then it will assume they are birds. It doesn’t have enough ‘tree sheep’ data to make the right decision.[1]

At ContactEngine we hold tens of millions of conversations every year, so we have a lot of training data. We can see, per client, which are the best times to start a conversation to maximize the chance of a response. As we learn more, we can develop a nuanced call schedule and contact people only when we know they will answer. It’s the difference between a broad mailshot and a delivery timed for when you expect the recipient to be home.

We can see unusual details too, such as the fact response rates are best when the outdoor temperature is between 50-53.6°F. Or that people take two hours longer to respond to a text in summer than in winter. People just have better things to do in hot weather, I guess.

A/B testing
This is the next level after mining data. A/B testing – in which you test two variants of an idea – is as old as science. You have a control group and a test group. Applied to what we do, it means we can use AI to subtly vary messages to see what kind of language gets the best response. In fact, we employ linguistic experts to help determine this.

Research shows that people have a more positive response to words they learned before the age of 10. We use that and other knowledge to score messages based on their emotional content and then let the AI automatically boost the successful ones.

Overlaying data
When you know more about a customer, you can learn what they respond to. For example, if someone is older and has been with their energy supplier a long time, they are less likely to respond to a message about switching to a smart meter than a younger customer who has been with the company for less time.

There could be lots of reasons why this is, but what matters is that we can use this information to deliver a better experience. It may be best to contact the older customer by post first, so you can give them more information about how a smart meter will help them save energy. A younger customer could just get a text straight away.

Inbound messaging
When people do respond to a text, the majority just say “Yes”. The appointment is fine, they do want a smart meter, or whatever. But about 20 per cent respond with something else. If they reply “No, I can’t do next Wednesday morning, but how about Friday” for example, then that needs a response.

Without machine learning, you would pass those messages to a call center. That’s a worse experience for everyone: it’s more work for the call center and the customer doesn’t get an answer until the call center is open, which might not be until the next working day.

With enough data, machine learning can understand the intent of the message and respond accordingly. So, “No, I can’t do next Wednesday morning, but how about Friday” gets a response that says “Would 10am on Friday be ok?” for example.

As the machine learns more, it can respond more personally. If we send a message asking how an appointment went and get a reply saying, “It was great, please tell John he did a brilliant job!” then we can not only reply that we’re pleased it went well but, because we booked the appointment, we can forward the message to John – all without a person needing to make decisions.

This isn’t AI for rocket science but it’s still remarkable. With machines shouldering more of the load, call center staff can concentrate on the genuinely difficult problems and customers with simple requests can get what they want without fuss.

Now if we can just get computers to recognize sheep in trees…

 

[1] http://aiweirdness.com/post/171451900302/do-neural-nets-dream-of-electric-sheep