otso Annotator is live
Trial our cloud-based annotator built for Machine Learning Engineers and Data Scientists now. No credit card required.
Learn More
Customer Insights
Project Feelings
Book a Demo
Log in

Supercharging Customer Insights with AI

If you missed our recent webinar on Supercharging Customer Insights with AI, you can access it on-demand here.

For those that prefer written content, this blog explores the same content.

Machine learning is a fantastic tool in your customer insights arsenal when it is used correctly and relied upon and trustworthy; otherwise, it's just another buzzword.

This post is split into three sections. Firstly, we go over what AI is. Not in an exhaustive manner - AI is a deep topic. But we aim to leave you thinking; "okay, when I see AI in the wild - when I see other companies talking about AI - I know a little bit about what they're talking about."  We then give some practical use cases in the customer insights space before explaining domain-specific modelling.

What is AI

Machine learning is fundamentally a shift away from deterministic or rules-based programming, to probabilistic programming. What this means is that rather than using rules to find patterns within data, we are using mathematical algorithms and statistical algorithms to find patterns.

You can think about a machine learning model as a mathematical algorithm trained on a particular data set, to perform a specific task. The algorithm's complexity level is what is referred to when you hear terms like deep learning or shallow learning.

Once you've got a model that is trained on a particular task, that is what we would call an AI capability. And it's where you start getting something that resembles artificial intelligence - the training of a machine learning model to replicate a human's cognitive function. That could be sight with image recognition, or hearing with voice to text. It could be some understanding or comprehension with natural language processing.

We take these AI capabilities and combine them to form an application. It's the application that creates value within an organisation. The nature of creating an AI application means that it is ill-advised to do machine learning for the sake of doing a machine learning project. A favourable business outcome can only be achieved when an AI application has been created with an understanding of the human cognitive function it is replicating.

The next key element for understanding AI is data. All organisations are full of data, and we are sure you have heard that “data is more valuable than oil.” We like to view data in an organisation a bit differently. It is nearly always an iceberg.

Above the waterline, you've got your quantitative metrics. Data that is well understood within the business. There are already good analysis methods to unlock value in this data, such as NPS ratings and product reviews. But below the surface lies all of the qualitative data - and many organisations often overlook this.

This is unstructured data and can include customer feedback, voice recordings and images. The things that are not readily used within an organisation. It is crucial to organisations remaining competitive in the market that they extract value from these data sets and bring hidden insights to the surface. Otherwise, only a fraction of the story the data holds is told.

The last concept is domain-specific modelling. This is something that we don't think a lot of people touch upon when they are explaining AI foundations. Domain-specific modelling describes training a model to do a particular task. The more complicated or specific that task is, the more critical it is that the data that the model is trained on relates to that task.

Think about this by way of an example. Imagine that you're putting together a trivia team. You have one place left in the team, and three of your friends all want to join; Gary, Mary and Dennis.

Gary is a bit of a generalist. He knows a little bit about a lot of things. He's what we would call our generic model or off the shelf model. He has broad coverage, but low depth. If you need to get deep on a particular subject, he's not probably going to be the man to help you there.

Mary, though, is a film and TV buff. She goes to film festivals, and she subscribes to all the streaming services. Mary is what we would call an industry model. She is someone that has a lot of expertise in a particular vertical.

Finally, we've got Dennis. He is your eccentric friend that spends all his spare time watching reruns of the 1999 cult classic being John Malcovich. No one knows more about that movie than he does — not even John Malkovich. Dennis is similar to a highly bespoke, ‘specialist’ model — extremely capable, for a very specific niche.

So who would you take? The answer is, it depends on the trivia.

If this is just your local pub trivia, general knowledge, then you probably best off taking Gary. He's got the best range of knowledge across topics. If it is a TV themed night though, Mary would be the best pick. Suppose you've got the annual “Being John Malkovich" trivia tournament though, with a hundred thousand dollars in prize money riding on the outcome. Can you afford not to take Dennis?

Now that we know what AI is, we turn to how AI can supercharge customer insights.

While the trivia metaphor is a simple example, a more practical example is an engagement otso.ai did with an Australian Utility Provider. They had recently started using a cloud provider to enrich their customer feedback with sentiment. They found, however, that the sentiment wasn't quite accurate enough.

So then as a stop-gap, they put two of their full-time resources into reviewing and verifying the sentiment for their upstream reporting - not a very efficient use of valuable customer insight time.

We took the sample that they'd been annotating and trained the domain-specific model for them as illustrated below.

This is no slight against the cloud provider because they are trying to provide a model that's everything to everyone, which is a really valuable tool. They are providing a Gary, in our earlier metaphor.

The problem is that when you've got a highly specific domain, which this utility provider did. Their customers spoke to them in an incredibly specific way. They needed a Dennis. They needed to have a model that is trained on their data.

The model trained on their data resulted not only in a highly accurate sentiment, but they were also able to free up those resources for value-creating tasks.

Three pillars of AI for Insights Quality

The three pillars of a value-creating AI project for customer insights are accuracy, scalability and speed to insight. When making a decision based on text, you have to be able to trust the underlying modelling used.

You might have decisions being made by people who have been with the organisation for 10 or 20 or 30 years. You might have other decisions been made by new graduates who have only been with the organisation for a couple of weeks. Domain-specific AI modelling can democratise knowledge across the organisation. It can take all of that organisational historical learning and use that to level the playing field in the team's experience levels.

You've then got scalability. If we gave you a list of a hundred customer verbatims, you could tell me the pain points you could tell me what's wrong. If there are any churn risks, what the key themes are and more. But what would happen if we gave you a thousand examples or 10,000 or a hundred thousand, it wouldn't be that easy.

And on that same point, how long would it take you to get back to me? Those hundred responses analysed would likely be a couple of hours. Whereas a domain-specific model could get through this work in minutes.

Extracting key information from text.

Customer feedback text comes from a wide variety of sources, internal and external solicited and unsolicited.

The challenge is, how do you analyse all of those together without a manually intensive time consuming, costly process? Machine learning is excellent at doing just that. It is fantastic at extracting information at scale, from a large range of data sources. This includes things like sentiment analysis or key entity analysis. But also more specialised metrics like churn risks, or intent to renew or cancel. And what this gives you as an organisation or as an, an insight team is the opportunity to be proactive. It allows you to come to the conversation with evidence in hand. It allows you to spot things in near real-time, so you can act faster.

A great example of proactivity AI for CX can afford is a project otso.ai did with a New South Wales government agency for Vivid Sydney. We were monitoring feedback on the event in real-time. We found a small trend that some art installations weren't wheelchair accessible. It was small in number, but the momentum was slowly growing. The agency was able to take action early, and as a result, an incredibly negative story was avoided.

Sometimes being proactive is impossible. Sometimes you have to be reactive. In these cases, domain-specific modelling also gives you the ability to provide evidence-backed insight to help you support or disprove an organisational anecdotal hypothesis.

Let's say you just have sat down on a Monday morning, ready to get stuck into work. You then receive correspondence that a senior executive has seen a negative review on social media about pricing.

With your domain-specific model in your toolbelt, in minutes you are sending back that actually in the last seven days, there have been no complaints about pricing across all channels.

Or you might come back and say "yes, we have had complaints as we just upped our pricing. And every time we've upped our pricing in the past, we've had a similar number. So it's no cause for concern." Your model allows you to provide that evidence.

Break down data silos

You know that in an organisation, how many times do you only analyse the data you've got available without touching the other data that you could potentially analyse because you don't own it or can't access it, or you don't even know it exists. That is a very common problem.

We worked with two different teams within one bank. One team was responsible for home loans, the other internet banking. The amount of cross-pollination between their survey responses was massive. Their customers were leaving app reviews such as "Hey, I love this app, but can you get back to me on my Homeland application" and "Hey, the home loan went really smoothly, but then I went to try and access my redraw and I couldn’t."

Customers view any single channel as a way to interact with the business. We need to make sure that we're considering all channels and analysing all data to genuinely understand how the customer feels.

Otso is a machine learning company specialising in the enrichment and analysis of unstructured text at scale, using a combination of in-house products, and state-of-the-art natural language processing and machine learning.

We support our clients across a range of business use cases, including voice of the customer and text analytics through to operational examples, like claim automation.
If you like to explore what an AI project would look like for your team, get in touch.

About UsCareersContact UsBlog
Copyright © 2021 - otso.ai. All rights reserved.
Follow us:
LinkedIn LinkTwitter Link