In the previous blog in this series we looked at the main categories of Machine Learning namely; Supervised Learning, Unsupervised Learning and Reinforcement Learning.
Machine Learning - Technical Background
Dive deeper into Machine Learning models and discover the differences between Supervised, Unsupervised and Reinforcement Learning.
Whilst we just skimmed the surface of these topics, it may have left you wondering how to get started? In this blog we will look at some issues and considerations that may leave you wondering whether it’s even worth starting!
As we covered previously, not embracing what Professor Stuart Russell called “the biggest event in human history” may be short-sighted.
Don’t get despondent, Salesforce can help.
Salesforce has over 20 years of history of taking complex technologies and making them ‘business friendly’. The same is true for Machine Learning.
Anyone who has been around the Salesforce ecosystem will be familiar with the breadth of the ‘Salesforce Customer 360’. Salesforce has solutions for Marketing, Commerce, Sales, Service, Analytics (and many more). The good news is that Machine Learning capabilities are available throughout the the Salesforce Customer 360 - with a ‘business friendly’ interface. Let’s take a look at just a few of these solutions.
As we covered previously, not embracing what Professor Stuart Russell called “the biggest event in human history” may be short-sighted.
Don’t get despondent, Salesforce can help.
Salesforce has over 20 years of history of taking complex technologies and making them ‘business friendly’. The same is true for Machine Learning.
Anyone who has been around the Salesforce ecosystem will be familiar with the breadth of the ‘Salesforce Customer 360’. Salesforce has solutions for Marketing, Commerce, Sales, Service, Analytics (and many more). The good news is that Machine Learning capabilities are available throughout the the Salesforce Customer 360 - with a ‘business friendly’ interface. Let’s take a look at just a few of these solutions.
Marketing
Imagine you are in a marketing function.
You want to predict the likelihood that a customer will engage with your campaigns in order to maximise campaign effectiveness. A good way of achieving this would be to predict the likelihood that a subscriber will engage (open, click-through and convert) through each channel (email, social, etc).
This is a job for Supervised Learning. With Supervised Learning we will be able to create a score for Engagement based on historic engagement data. You need historic data to train the model - 90 days of engagement data including Opens, Clicks, Unsubscribes, etc.
With the predicted Engagement score, you could personalise the email messages and campaigns each subscriber is sent to optimise the chance of positive engagement with that content.
Using predictive Engagement Scoring, a Salesforce customer in the travel industry was able to achieve a 66% drop in unsubscribe rate and a 13% revenue increase.
Make every moment count to deliver success now
Make your spend go further and drive efficiency with Marketing Cloud.
Commerce
Your marketing team have been very successful in driving subscribers to your e-commerce site. You’ve got a load of great products and you want to ensure prospective customers can find the correct products quickly. A good way of achieving this would be to use personalisation to surface relevant product assortments throughout the shopper journey.
This looks very much like an Unsupervised Learning task. First we have insights into customers’ buying patterns, site browsing tendencies, relationships between search terms and products purchased. We couple this with information about the current shopper and their specific browsing histories.
The result of the model is personalised (predictive) product sorting. Products relevant to the customer are predictively served up resulting in a much smoother shopping experience.
Using AI powered Predictive Sort has been shown to result in a 9.1% increase in revenue per visitor and a 3.8% increase in conversion rates.
Maximise revenue now
Get value fast — from discovery to delivery — on the platform trusted by industry-leading brands.
Sales
Your e-commerce sales are going well - but a large proposition of your business is via ‘assisted sales’. Luckily you have a lot of Opportunities through the good work the marketing team are doing - but with so many opportunities, where do you best spend your time? A good way of achieving this would be to predict a score for each Opportunity indicating the ‘quality’ - allowing you to concentrate on the best Opportunities in the given financial period.
This type of prediction modelling can be achieved with Supervised Learning. With Supervised learning we need training data - historical Opportunity data in this instance. Specifically, you need more than 200 Closed/Won Opportunities and at least 200 Closed/Lost Opportunities over the past 24 months.
The result of this model is a prioritised list of Opportunities - allowing the individual seller to maximise their revenue creation potential. In addition, because scoring can be used in forecasting, more objective sales forecasting can be achieved.
Using AI to indicate the optimum use of time has a positive impact on Opportunity close rates. One large Salesforce customer in the consumer goods space experienced a 48% increase in win rates by concentrating on the correct Opportunities.
Reduce costs and increase productivity
Max out efficiency with better automation, data, and intelligence. How? With sales force automation powered by AI.
Service
Sales are going well - but your responsibility to your customers extends post sale. You have set up a self-service support channel and that’s going well but now service agents need to address the more challenging cases that remain. It would be great to learn what resolved previous similar cases so we can maximise agent productivity (and customer satisfaction).
Again, this is a great fit for Supervised learning. As you know by now, we need historic data. In this instance we need at least 1000 cases - at least 500 cases with knowledge base articles attached to them.
The result of this model is that agents get Article Recommendations to resolve a current case based on its similarity to previously resolved cases. Agents can save time searching for answers and customers can get their issues resolved promptly.
Using Salesforce AI powered solutions to assist agents, a large Salesforce company in the electronics space was able to save 5 hours per agent per week in productivity gains.
Scale service that customers love
Maximise ROI and drive efficiency from the contact center to the field - on a single platform.
Above are just a few examples - there are many more. What these have in common is that complex technology is being used to address business use cases. Another thing they have in common is that, although the technology being utilised is complex, no specialised Machine Learning or Data Science knowledge was needed to realise the benefits.
As we discussed at the start of this post, Salesforce has a rich history of allowing businesses to realise the benefits of complex technology - without having to be experts in that technology.
A ‘classical’ Machine Learning approach to the business needs above would involve steps such as; data collection, data transformation, data sampling, feature selection, model selection, score calibration and integrating the model results into the user application.
Only when all of these steps are complete (and most likely many iterations) can business value be realised.
Salesforce has a data model which is often customised for each business - this allows us to automate feature engineering. Salesforce also provides automatic model building and customer specific model selection. As the Machine Learning capabilities are part of the Salesforce Customer 360, the model outcomes are automatically available to the business users - and can be continuously refined based on business interactions. All of these complex stages are abstracted away behind simple declarative admin friendly interfaces.
If we look at just one step - model selection, we can see the complexity Salesforce is relieving us from. For each use case that we build (e.g. Opportunity Scoring), Salesforce tries many different models for a given customer. We make the hyper-parameters vary automatically: for example, Model 1 will be a Random Forest with 50 trees, Model 2 will be a Random Forest with 100 trees, Model 3 will be a Logit with L1 regularisation and Model 4 will be a Logit with L2 regularisation. We run a model tournament to select the model that has the best accuracy, measured over a hold-out data sample. If all of the above sounded like nonsense - that’s fine. The result is that you can be sure you have the appropriate model with the appropriate parameters without having to worry (or even know) about the details.
For more details on the complexity Salesforce saves you from worrying about check out this video from our Salesforce Developer community.
As we discussed at the start of this post, Salesforce has a rich history of allowing businesses to realise the benefits of complex technology - without having to be experts in that technology.
A ‘classical’ Machine Learning approach to the business needs above would involve steps such as; data collection, data transformation, data sampling, feature selection, model selection, score calibration and integrating the model results into the user application.
Only when all of these steps are complete (and most likely many iterations) can business value be realised.
Salesforce has a data model which is often customised for each business - this allows us to automate feature engineering. Salesforce also provides automatic model building and customer specific model selection. As the Machine Learning capabilities are part of the Salesforce Customer 360, the model outcomes are automatically available to the business users - and can be continuously refined based on business interactions. All of these complex stages are abstracted away behind simple declarative admin friendly interfaces.
If we look at just one step - model selection, we can see the complexity Salesforce is relieving us from. For each use case that we build (e.g. Opportunity Scoring), Salesforce tries many different models for a given customer. We make the hyper-parameters vary automatically: for example, Model 1 will be a Random Forest with 50 trees, Model 2 will be a Random Forest with 100 trees, Model 3 will be a Logit with L1 regularisation and Model 4 will be a Logit with L2 regularisation. We run a model tournament to select the model that has the best accuracy, measured over a hold-out data sample. If all of the above sounded like nonsense - that’s fine. The result is that you can be sure you have the appropriate model with the appropriate parameters without having to worry (or even know) about the details.
For more details on the complexity Salesforce saves you from worrying about check out this video from our Salesforce Developer community.
Machine Learning - Issues and Considerations
Does Machine Learning really pose a risk to how we go about our daily life and business or can it be an ally in our path towards greater personalisation?
