AI-driven Personalised Marketing: How We Got Here and Where We’re Going Next
Across the entire customer journey, AI is helping marketers provide more assistive, enjoyable experiences for shoppers and improving performance with it: from one-to-one messaging that doesn’t sound like spam, to dynamic website design that adapts to individual users, and even retail stores that recognise and welcome their visitors.
In part two of this series, I’ll show you what this looks like ‘in the wild’, with fun, innovative examples from real companies, both large and small. But first, let’s take a look at how we went from simple, rule-based marketing campaigns to using AI, Big Data and Machine Learning to open up new possibilities for marketers, and consumers.
The thing is, unlike gimmicky tactics of the past, AI-driven personalisation in marketing is actually being demanded by today’s customers. Why? Well, we are all distracted and time-poor, but tech-savvy. We’ve been trained by the likes of Google and Amazon to get what we want, right now. The result?:
Customers expect brands to convey the right message, with every communication. In fact, 81% of consumers want brands to know when and when not to approach them, while 63% of consumers expect personalisation as a standard of service and believe they are recognised as an individual when sent special offers (Accenture). From a brand or retailer’s perspective, this means that in this attention economy, content which fails to resonate is a missed opportunity, or worse.
How did we get here?
In order to really understand just how revolutionary AI-driven personalisation has been for marketers, let’s look back in time at how thing’s used to be.
It was always true that understanding consumers is vital, but in the past, the required information was hard to attain. Marketers had to use assumptions about what the different types of customers want, and then trial and error to verify this. It was of course expensive, time-consuming and an inefficient use of marketing budget, but they had no other choice.
In the 90’s, basic rule-based personalisation became available via Customer Relations Management software (that is, databases of information on customers, built via a company’s interactions with them). For example, greeting customers by name in emails, or sending them coupons on their birthday.
The problem here, was that segmentation into customer types was limited by the manual effort required. Furthermore, marketers had to decide, rather than learn, what content to show to which visitors. And most challenging of all: data was restricted to information from customers who were logged into the retailer’s website.
In the 2000’s, social networks began collecting the data which was required for tracking tools and insights. In addition, years of improvements to A/B testing techniques and segmentation strategies meant marketers were getting a little more granular with their targeting efforts. However, the data was becoming bigger than any human could ever make sense of!
Turning point: The Benefits of Big Data, Compute Power, AI and Machine Learning
The situation now looks very different, thanks to a wave of technological breakthroughs we are all familiar with.
Firstly, big data techniques can now be used to collect, process and analyse huge amounts of behavioural data , like search logs, email clicks and shipping cart history. Such tracking, combined with users’ wilful use of the same platforms across different devices (mobile, desktop, etc), give marketers an ‘omni-channel’ view of the entire customer journey — and no, they no longer need to be logged into the retailer’s website.
Artificial Intelligence practitioners can leverage many variables which reveal how, when, and what customers shop for. Variables could include factors like contextual data, behaviour, demographics, expressed interests, customer rewards programs, seasonal data (such as Christmas and summer trends), and even local weather information!
Machine Learning can identify patterns — like customer needs — from this, and use them to make predictions. And finally, GPU -based compute power enables us to build ever more complex (and usually more accurate ) models for a huge variety of tasks.
So what’s the result of all this?
Granular segmentation means we can now identify more specific customer types, like ‘female, aged 20–25, interested in high fashion, is highly price sensitive’, using machine learning-based classification . Knowledge about these fine-grained segments can then inform product development, pricing, targeting, messaging and performance measurement.
Accurate Attribution allows understanding the customer journey, which channels and touch-points are effective, and which content drives results. Of course, knowing this can help marketers allocate their budget efficiently to improve their return on investment (ROI).
Predictive analytics provides insights into customer needs, industry trends, and intent signals, now and in the future. Again this is useful for marketing strategy, and it can also inform inventory management, which is about determining ‘how many of which products should I stock at any given time?’
Predictive marketing lets marketers adjust to customers’ next actions and experiment with personalised experiences, with the confidence that their content will be more contextually relevant: that is, delivered via the channel, time, and medium which suits the customer most. Naturally, this can improve the customer relationship.
Clearly Big Data and AI power are bringing about many benefits. A bonus is, that all of this can be done more easily, at scale, and over the whole customer journey — from search to loyalty and advocacy. Furthermore, marketing data, like clicks, time-on-page, and purchases, is (often) high volume, enabling fast model training and making it easier to test personalisation techniques.
For consumers, the experience becomes easier, more relevant, personal and human, and the gap between physical and digital worlds can be bridged. This is more important than ever, now that we are restricted to shopping online and no longer have a salesperson waiting to greet and assist us.
If you think that sounds exciting, you’re in for a treat! In part 2, I’ll illustrate just what an impact all of this is having on marketing, by providing some real-world examples (and for those not so comfortable with machine learning, there’s a handy guide to help you interpret AI solutions based on their marketing material). Read it here.
 Behavioural Data: this data is gathered by cookies, e.g. Google Analytics for PPC accounts in the Google universe, or tracking links which web designers can embed in their various product and landing pages.
 Graphics Processing Units: First used for graphics in gaming, GPU’s were designed to accelerate tasks like mapping texture and lighting, which requires a lot of linear algebra, specifically, matrix and vector operations. That is, they had to perform many, many calculations on grids and series of numbers. They handled this by ‘parallelisation’, which means doing many of these calculations in parallel. Neural networks require the same kinds of parallelisation, which is why neural machine learning and deep learning began experiencing a boom once GPUs became available. Fun fact: the first work on neural networks began in the 40’s and 50s, they just didn’t have the compute power then to do much with it.
 I mentioned that GPU’s enable us to build more complex models, which are usually more accurate. Why do I stress the ‘usually’? Neural networks have accomplished remarkable feats, but at the cost of huge amounts of data and training time. Thus for a lot of real-world projects — outside of academia — they are not feasible. Many organisations simply do not have the resources these models require, and if they were to proceed with them anyway, they would enter a danger zone. That’s because if you train a neural network on too little data, it can become ‘overfit’, meaning it has essentially memorised the data perfectly, but would struggle to produce useful outputs for future, unseen test examples.
 Using machine learning classifiers to identify customer segments even works for emerging ones. For example, when people first started searching for ‘bio’ or ‘vegan’ products, classifiers would have struggled to categorise them, which would have alerted marketers to this new trend.