Transcending the term buzzword, AI has proven it is here to stay. It is manifesting itself in many different ways; from futuristic ‘robots are taking over our jobs’ visions up until basic voice search solutions. It’s easy to feel lost in the search of how AI can be of any help for your business right now. The technology has been around for decades, but practical solutions for marketing are fairly new. So, is it necessary to invest in heavy machinery to be able to use AI? Luckily, the answer is ‘no’. Because by bringing it back to machine learning there are plenty of possibilities available right now.
Machine learning is a technique to make predictions, based on historical data. Which is a broad definition, so before you start using this technology ask yourself what sort of predictions are valuable for your business. If you can answer that, dig deeper to specify what sort of data is necessary, what you already have and what third-party or open sources you could use. Having that set-out makes the right usage easier. Or approach it the other way around: what data do you have available, and what kind of predictions can you make out of it?
The future looks bright, start now
As Kai-Fu Lee describes in his book ‘AI Superpowers’, the technology comes in four so-called ‘waves’:
- Internet AI: mainly used by big internet players, the FAANG (Facebook, Amazon, Apple, Netflix, Google), to make their services smarter.
- Business AI: this is very applicable for many businesses like banks, telco’s and utilities. It’s about using business data in predictions with algorithms.
- Perception AI: usable for consumers. Think Google Glass, which was a first (although not very successful) step into consumer AI. We will have more and more applications reaching consumers.
- Autonomous AI: this is very much in the future, like self-driving cars and autonomous warehouses. We’re working on it, but we’re not there yet.
Although it is tempting to start dreaming about how AR glasses will unlock the full potential of perception AI, or if drivers licenses are part of the past any time soon, for now, the second wave is the most relevant for businesses. That’s where the biggest opportunities are at the moment. Why? Because most companies are not Amazon or Google, their data, knowledge and tools can be used in a smart way to enhance businesses now.
Four layers for business AI
Within business AI there are many ways to embed the technology. Breaking it up into four layers gives a clear overview of how to make business easier. An important thing to note is that the focus is not on building self-driving cars, but creating smart enhancements on your existing service.
1. Cognitive services
By using existing cognitive services like a Vision API, customers get to insert data on request much easier. Take your meter readings for example. It’s always a hassle to find the tiny numbers in the dark of the fuse box, remembering them and then entering them on the website. Wouldn’t it be so much easier to just take a picture of meter readings, uploading them and having AI to recognise and extract the numbers? It’s possible. The same goes for speech to text and text to speech API. You can use this to monitor your call centre, and to record and transcribe all conversations taking place. This data can become very valuable in the coming years when you want to train your robocall centre agents with appropriate dialogues or to automatically subtitle instruction videos. When converting text to speech it can be valuable to switch from a human agent to a robot-agent at the end of dialogue to read out specific terms and conditions and end the conversation politely without wasting human agents time. The next step in this would be natural language and dialogue API, with which full robocall centre agents are a futuristic possibility.
2. Use your own datasets
Leverage the power of transfer learning by feeding algorithms with existing knowledge. There’s always a starting point, but the convenient part of algorithms is that the first level of knowledge (see the image below) is only to expand and improve. For example, when you have a large catalogue with images of products, you could use transfer learning to get a model which specialises in accurately identifying your products. This technique uses the baseline of the general image recognition models, but then you train it a bit further with your own limited labelled image sets. So, if a consumer wants to place a repeat order of a pair of jeans, he or she just has to take a few photos of his old ones to get a personalised suggestion of a new pair in the current collection of your brand. This is applicable to all types of products. Take a b2b customer for example. You could easily detect the exact type of equipment that you’re using without looking for the type code, or a utility consumer can identify the type of meter in his cupboard.
3. Yield management
When you have inventory that you need to sell, you want to make sure that the remaining inventory is connected to pricing and advertising models. Because
knowing beforehand how much you have left and how much extra budget you need to spend on selling the remaining part really boosts efficiency. Connect the data you collect per season to optimise pricing and advertising strategy, therefore saving budget, or spending it more efficiently.
4. Segmenting & targeting
Basically, this is the classic advertising game, but on steroids.
Customer segmentation for personalisation is one of the best practices today. However, you can take it to the next level by integrating back-office data in order to get a better view of real customer lifetime value. It will give you detailed information on real-time conversion behaviour in certain consumer groups. Use these insights to choose customised targeting. On top of that, historical behaviour also shows you the worth of each type of consumer. Take the die-hard shoppers for example. They may sound like the perfect customer to target to buy more, but when it’s the group that also returns 80% of every order, they need a different approach as their intention differs from other groups. Targeting them to order more, just ends up in losses for your business. Connect your internal first-party data sources (CRM and ERP) to your advertising and online data to pick the best customers. With segment-based buying behaviour, you’ll be saving on your marketing spend and end with increasing revenue.
Predictive (re)targeting
When it comes to targeting, many companies do so on gut feeling, such as ‘target all people who visited the website’, or ‘all people who have viewed the product’. However, there are better ways to go about this. Preferably, you want to target visitors who are likely to convert when you reach out to them again. And if you are able to gain insights into the actual buyer intent per individual visitor, this is possible; allowing you to only approach relevant audiences. Not only will that improve the conversion rate percentage, but it will also solve existing frustration among users about irrelevant targeting and improve the overall brand experience for rightly targeted persons.
How-to: a predictive model
Create a model that can predict buyers intent in real-time by smartly using clickstream-data and several machine learning techniques. Seems difficult? It’s actually easier than it sounds.
1. Collect clickstream data of enough sessions (for example 500.000, but that can differ per website), with all interactions of the visitors during their time on your website.
2. After that, use the algorithm suite on the collected data to figure out which algorithm best recognises browser behaviour that signifies buying customers. Use many indicators to recognise this, among which data from the current and past website visits. Finally, after countless tests and repetitions, the algorithm that predicts most accurately will be used on the website.
3. The next step is to implement the predictor model on the website so it can make real-time predictions and make sure the needed information is collected and stored. This can be done with several JavaScripts via Google Tag Manager.
4. When the predictor model is in place, predictions are ready to be sent to all desired advertisement platforms so they can use it for focused retargeting.
With everything set and done, it’s possible to use the information about buyers intent to improve the retargeting audiences in campaigns, as to only reach out to ones who are likely to convert. This actually works, as proven in an experiment whereby the target group for retargeting, which consisted of all people who viewed a product or added a product to their shopping cart, was split into two groups. The first group was the group with a weak predicted purchase intention and the second group had a strong predicted purchase intention. The same bidding strategy and the same frequency range applied to both groups, and both groups were approached via Facebook with the same advertisement.
It may come as no surprise that the group with a strong buying intention is one-third of the total group that was targeted. On the other hand, it is interesting that the difference in conversion rate between the two groups is considerable. In other words, if only the visitors that are classified as visitors with a strong buying intent would’ve been retargeted, it would generate a 236 percent increase in conversion rate compared to the group with a weak buying intent (see the visual below). But if the objective is to optimise sales, it’s not advisable to exclude visitors who are classified as visitors with a weak purchase intention, because potential buyers are included in this group. However, it is recommendable to treat this group differently (e.g. with a different bidding strategy, frequency range, etc.).
Alternative purposes
This model can serve various purposes. For example, it can be used to deliver dynamic content to personalise the website. In addition, the predicted purchase intention can be used to expand the reports. Since we measure not only some interactions with this metric (purchase intention), but the actual intention of website visitors, it is a much richer metric that provides insight into the general intention of a specific group of visitors. Furthermore, only a small adjustment is needed to make this model suitable for predicting the expected revenue that a visitor is likely to generate when he or she is approached again, or which frequency range should be used for retargeting each individual visitor.
Finally, through the smart use of big data, machine learning and open-source software, it is possible to develop a prediction model that helps online marketers to approach the most relevant target group possible and leave uninterested visitors alone, resulting in more efficient use of budgets. This model is suitable for many different applications and thus creates a new playing field in online marketing, in which business issues can be translated into data solutions using machine learning and data.
Power POC
Applying machine learning to your business may come across as a big step, but this is for everybody, not just the big corporations. If you want to know whether to use machine learning within your organisation, it’s possible to use the so-called ‘power POC (proof of concept)’. Within one to two weeks you have a business case with possible machine learning solutions (delivered by Dept) that will convince managers and/or stakeholders of the added value of machine learning solutions.
Chatbots on the rise
With AI solutions on the rise, chatbots are a hot topic. With technology constantly improving, everybody basically has the same expectations about these digital assistants: a chatbot that understands what people asks them, answers them immediately and understands everything that’s being said. Preferably, the bot is also one that is autonomous in learning and becomes smarter. Call it the superhero among all chatbots. And like superheroes, it’s fiction. Because although we’re making great progress in the development stage, we’re not quite there yet.
However, since people are using messaging apps more often than ever before, there are many “real” options to use chatbots. Examples are customer services who have implemented a messenger on their website or the (re-)targeting possibilities via Facebook chat, and even sending out weekly newsletters via WhatsApp. The possibilities are numerous. And this may lead many to think that customer service is superfluous in the future or that in due time a website isn’t needed anymore, but this isn’t entirely correct. However, a chatbot is of great help when it comes to supporting and optimising your existing marketing.
How to get started
In order to apply chatbots to your business, you have to start off with the very basics and grow into the usage of chatbots. This is best done in four steps:
1. Start with outlining which questions your customers ask most often. Take Vacanceselect for example; they’ve set up a chatbot to answer one question in particular: “What should my next vacation destination be?”. With real-time user testing, the chatbot is made smarter by checking how the bots respond to the question and where issues occur. Even though this is not a self-learning chatbot, it works pretty well. And most of the time you don’t even need a self-learning bot, because come to think of it: about 80 percent of the questions are proposed more often, so using a hand-made chatbot like Vacanselect does, means higher efficiency, without complicated machine learning techniques.
2. Of course, 20 percent of the questions are new, so how do you go about that? One solution can be to install an automated response that the chatbot doesn’t understand the question and to reroute it to a human client service employee for example. Think about solutions for situations a chatbot can’t handle beforehand.
3. The logical next step would be to focus on personalised questions, such as ‘what will be my health insurance price?’. That differs from person to person. Use API’s to retrieve personal data out of your databases to enrich your chatbot. The chatbot becomes smarter and can answer on a more personal level. However, this is still done with some old school programming.
4. Lastly, the self-learning chatbot is one that needs to be trained, continuously. Start with collecting data, for example all of your call center conversations and use ASR (voice to text) to provide the algorithms with input. This is where machine learning steps in, because if you want to make your chatbot self-learning ASR is one way to go, whereas NLP (Natural Language Processor) is the next – it will help the chatbot to understand actual intentions from the sentences. Using machine learning for your chatbot lets it understand certain indications of time and place. Meaning you won’t need a decision tree that defines all questions; the systems will find out what people want on its own.
Just do it
For all these examples there is a pretty common denominator: the better the quality of your internal data – and the easier the availability of this data to your marketers and CX builders – the easier it becomes to start working on these enhancements in AI. It’s actually pretty simple, you just have to put your back into it and use the enormous amount of opportunities and services available to incorporate AI into your marketing. So, what is the road to AI? To just do it.