machine learning

Conversational Commerce, without the BS

I’m investing substantial time and effort in helping people to better understand what all the hype around AI and conversational commerce technologies like chatbots and intelligent digital assistants is about. It can be rather overwhelming in my experience. And I don’t blame you. I’m overwhelmed at times by all this news on A.I. induced jobless futures and fully automated and self-learning machines as well. It looks as if A.I. will change the world before we can even say “what?”. But we know better.

Why use statistics if we can apply math?

I am convinced Natural Language Processing (NLP) and Machine Learning techniques can do great things, and will be able to do even greater things in the future. But it will take some time and many learnings before we get there. Progress has to be made, and will be made for sure. Just not through claiming ever greater disruptive breakthroughs that are based on assumptions, not evidence. Progress will also not be made by seeing A.I. and machine learning in specific, as the panacea to all problems.

As I’ve been saying in this post: Stop asking the machine to find the most probable answer if you know it with certainty yourself. Just provide the machine with the answer and instruct it to provide it when the question pops up. In most cases this will be a faster fix than asking the machine to learn the answer by comparing 100k+ inputs (Q’s) and outputs (A’s) to come up with the best estimate for the answer. This is specifically true since machine learning algorithms tend to make mistakes and therefor require human supervision.

Invest in relevance, not long tail understanding.

Also, NLP has intent recognition rates of 85% to 95%, depending a bit on the width of the application and the time spent on improving it (should be on the high end after a year). You can of course spend hundred thousands of dollars into getting it to 99,8% or better. But should you? In my experience the return does not often justify the investment. What does justify the investment, is making the answers more relevant to the customer/user by using context and personalisation.

Think of it, what is the reason your customer still makes a call after he’s read the airlines luggage policy? Right, he wants to know what that means to him as an occasional flyer with a last minute discounted ticket on a transatlantic flight. Better NLP will not provide him with the answer he’s looking for. A more personally relevant answer will. And of course the threshold for your company could be at 96% or 84% intent recognition. The exact rate is not the point, the point is you should know when to apply other strategies to improve the customer’s experience and get their jobs done.

AI has reached the peak of inflated expectations

Gartner believes that AI has now passed the peak of inflated expectations. That’s good news for all of us, but mostly for you. Because now you can start using AI, without the BS. And you can start using it to solve your problems and those of your customers, not the problem of a company that lacks experience in the field and has a ‘feeling close to certainty’ that conversational artificial intelligence will change the way the world spins.

 

Thoughts on machine learning for customer service chatbots

I’m thinking out loud a bit about machine learning strategies for customer service chatbots. Bear with me. More question than answers, because some of the strategies I see, I just fail to understand. I get the impression that some try to find a machine learning solution to a problem that is hardly there.

Machine learning takes too long to find an answer

Let me start with putting out here that machine learning is not very effective when it comes to finding the right answer to a question.

In a customer service context, machine learning can be useful when it comes to parts of natural language understanding, just not so much in providing the right answer. Because, once you understand what the customer is asking, your company should be able to provide an answer, start a process of getting one, or get the job done with straight through processing.

And if you know the answer, there’s no need to wait for the machine to estimate it. You can directly instruct the chatbot to provide the answer relevant to the question, no?

If so, why do so many startups use machine learning strategies that require tens of thousands of input and output examples to do this? That way it takes months for your chatbot to answer FAQ’s automatically. Let alone how long it takes for the less frequently asked questions.

Of course, you can also use the numerous examples to have the chatbot understand all the variations customers use to ask the same question. But that’s what natural language processing can already do. Maybe you need to help it identify synonyms, industry specify language, stop words etc., but in general this should be covered.

You can have the machine identify gaps, suggest fixes and have humans make it perfect

Much of this ‘helping the machine’ is done by humans. Machines can identify where it goes wrong and even suggest how to improve or fix it. This isn’t supervised machine learning nor reinforcement learning. It’s effective, but human work.

And, here’s the thing: You can wait for 18 months to let the machine have its own way, or you can have the machine identify gaps, suggest fixes and have humans make it perfect. And you can do all that the same business day!

So, on the next day, not the next quarter or year, your chatbot will have the right answer to the question and your customer is happy. How would that score on agility? So please tell me, why the wait? To prove that tech can do this? To prove that you can do this nifty technological trick?

I frankly would not care about that. I would care about serving my customers, fast and right, the first time. How about you?