axsa: How to transform an automated business process into an intelligent business process using Dynamics and Azure ML

[Blog bot]

Источник: http://blogs.msdn.com/b/axsa/archive...-azure-ml.aspx
==============

Dynamics automates business processes. And does so cheaply and simply and you know that. But what if business processes could be made more intelligent instead of just automated. Automation among other things, brings consistency and improves productivity and you know that too. That’s why you want it. But intelligence could do all of that at a different scale and do it in real-time and could be flexible as business changes perhaps at a fraction of cost. By intelligence I simply mean that instead of a system making rule-based decisions such decisions are made using big data technologies in the cloud in real-time. This opens up the possibility of using various data sources including social data and the possibility of using computing power of the cloud.

Variety, velocity and volume of data is increasing at tremendous pace. Technologies to work with this data are being offered in troves and data is the new oil…. Now we have all been hearing about such possibilities for some time now. But are companies really using and
leveraging such technologies and are these really useful? Jury is still out.

However I will like to share with you our first successful story in Dynamics space.

This post will provide you with a real example of a LIVE customer that benefits from this approach. And perhaps will inspire you to think of scenarios where you can use it in your own business.

CASE STUDY

Let's call our customer Contoso. Contoso is the UK’s leading foodservice delivery and collection provider, supplying full- range of food stuffs across the UK. It is a leading supplier to Restaurants, cars, pubs, café's and schools. It delivers five thousand high-quality, affordable food products to over 45,000 different establishments.

CUSTOMER BACKGROUND

Food service in UK is a highly competitive industry with relatively low barriers to entry. Most players source & distribute low-priced, low-margin simple food products and kitchen supplies from similar set of manufacturers and sell them to similar type of customers in their own regions in UK. The low-to-medium price-range food service business cannot be profitable by differentiating on product quality or by cost cutting alone.

Contoso therefore deploys advanced technology to provide highly differentiated customer service and has managed to grow at higher than market rate consistently for past ten years. Because customers are fickle and there are no real costs to switching suppliers for them, to keep existing customers and to grow the customer base Contoso strives to delight customers by offering new services and reinventing existing ones with laser-sharp attention to detail.

Contoso uses Dynamics AX 2012 as the central source of truth for all its other applications. Contoso has two sales channels - both online portal and call center contribute to sales equally. Call-center is the traditional channel through which all of business was routed once upon a time but Contoso introduced online portal to stay ahead of the curve, reduce call center costs and make it possible for customers to place an order anytime anywhere. Call center agents use Dynamics AX to record call history and create sales orders. Call center agents make guided selling possible which improves customer service experience and can also help with up-sells.

Online portal allows the flexibility for customers to place their orders whenever they like. Product catalogue is pulled from Dynamics AX and shown on the portal, customers complete the order online which creates sales orders in Dynamics AX for warehousing and delivery processes.

Contoso's quest to improve customer experience led them to big data technologies. In the beginning they only had one key requirement - how to predict what a customer is likely to buy today given all their past purchases in the current and past months. This prediction
will allow them to present only those products on the main page. And you may ask what is the benefit of this? Does the benefit justify the cost of building a predictive model? The key benefit is that ordering time is reduced, user does not have to scroll on the page or go to other pages or search through the portal. Same applies in the call center scenario. In a B2B scenario where chefs or restaurant managers have to order food ingredients and kitchen supplies they need online almost daily from many different sources, shaving off a few minutes and making their experience with Contoso stand-out among the competition is a very significant differentiator. What impressed us was the attention to customer experience and how central customer experience is to Contoso's investment decisions.

And if the predictive model could not only be built using simple and inexpensive tools but also could be changed, tested, trained, deployed and consumed without any up-front investment - what more could Contoso ask for?

Awareness that such technologies are already being offered by Microsoft is low currently in the Dynamics world. And more importantly that the real business value of these technologies is only in lighting-up business processes which live inside Dynamics.

So from the first key requirement this project with Contoso is now at a stage where we provide a number of different services.

1. We present next purchase
2. We offer recommendations
3. We analyse customer churn
4. And then we classify customers value to
   the business

We build predictive models in Azure ML, we leverage Azure ML apps available in Azure marketplace and we call Azure ML services real-time from within Dynamics.

Purchasing history and clickstream data are collected and ingested into these models automatically.

There are various predictive models, some are trained weekly others are trained on the fly.

There is still some work remaining to make the data flow seamless and automatic and we are working on improving this.

Let's see each of the Azure ML services created in a bit more detail.

CUSTOMER CLASSIFICATION

Customers are classified into different levels based on a mining model built in Azure ML studio. This classification helps call-center agents in decision making. Going forward, we will base promotion offers on this dynamic customer classification.

CUSTOMER'S NEXT PURCHASE

Based on purchase history, customer’s next purchase is predicted and presented products on the portal and in call-center are ordered by likeliness to buy today. This prediction is based on a model built in Azure ML studio. Customer’s next purchase is expected to delight the customer by presenting what she needs on a given day. It also increases order taking efficiency.

RECOMMENDATIONS AS A SERVICE

Customer has been live for a month and we have seen consistently that 20% of all the provided recommendations are clicked on by the users which is a very high number given this is a B2B space and customers/users are busy people whose main focus is to get done with the
order quickly. Almost 5% of the items in the final shopping cart are coming from the recommendations provided. If this number continues to hold this would lead to about 5% sales uplift which is also very high. Customer's expectation is 1-2% lift in the long-term. For Contoso this translates to £1-2m lift. Contoso consumes about half-a-million prediction events per month.

Three types of recommendations are offered in real-time

Market basket analysis produces frequently bought together recommendation when customer choses to put an item into the cart.When user clicks on a product, a real time call is made to Azure ML. The Azure ML service returns the items that are frequently bought together with this item. Azure ML models do all the math on more than billions of possible combinations to show the most relevant items. The most significant benefit here is to help customer place the order quickly to make sure s/he is not forgetting something most people buy together with this item.

Item-to-item recommendations are provided to the customer on the item landing page in the portal. When user clicks on an item to go to the item landing page, a real-time call is again made to Azure ML. The service returns other recommended items. These items are predictions based on what other users buy when they buy this item. Azure ML service not only recommends relevant items but makes sure that there is some novelty and diversity in its recommendations - nudging & subtly urging the customer to click. This helps in exposing the darker parts of the catalogue which the user may not be aware of. This not only can lead to upsells but also helps in bringing back the concept of guided-selling which Contoso lost to some extent when they moved half of their sales from call-center to the portal.

User-to-Item recommendations are provided to the customer just before she checks out both in the call-center and on the portal. Just before check-out another Azure ML call is made in real-time. This time user is recommended items based on the total basket of items user already has. Azure ML service here does what is referred to as "training on the fly", real-time personalization. This is unique to Azure ML.

CUSTOMER CHURN ANALYSIS

Contoso analyses monthly data to predict which customers are likely to churn in the current month. Sales department uses this information to call customers-at-risk and take necessary action.

DISCUSSION

As you can see the sales order business process has been surrounded by a number of machine learning services at critical decision points - what product is this customer likely missing in the cart?, what product is this customer likely to buy today? What product is this customer likely to want if revealed to the customer? Should I offer this promotion to all my customers or to the ones who are likely to churn or to the ones who are rock solid customers? Customer wants a delivery slot that is reserved, should I offer that slot to him given that he is a diamond customer for three straight months? This type of decision management capability is what transforms sales order process from merely being an automated process to an intelligent process. With the simplicity and affordability of Azure ML Studio and services in Azure marketplace Contoso can chose to stay engaged and continuously develop newer models to answer deeper questions about their business. For instance, say a particular product is recommended several times but gets rarely clicked by users, is it time to discontinue this product and save all costs related to the product. Or if certain recommended products consistently end up in the cart, is there some type of latent demand and perhaps more such products should be included in the catalogue? Or perhaps work in a different area for instance do text analytics on case history logs stored in Dynamics AX to identify customers who have complained often in the past or delivery drivers against whom most complaints have been registered or products against which most pre-orders have been placed.

What has been achieved at Contoso with Azure ML and Dynamics working together cannot have been achieved if Contoso relied on Dynamics alone. For instance clickstream data being analysed to monitor consumer behaviour would not fit in a SQL database. The recommendations offered are predictions that need heavy computing capability and would be too expensive if were deployed on premise. The business changes quickly, new products, new markets, new customer requirements, new competitor offerings all require
Contoso to respond quickly - Flexibility and ease of modifying a model and redeploying a service using Azure ML studio makes this possible to do in few hours.

This is a case study of a successful Dynamics customer. We are closely monitoring Contoso and at the same time working simultaneously on a number of other projects. If you have a scenario you would like to discuss please reach out aksheyg@microsoft.com.

Like I said initially, the jury is still out. Success depends on identifying right scenarios, having the requisite skillsets etc. but clearly the opportunity here is endless - surround the business processes with various Azure ML services, perhaps at each decision point and transform your automated business processes into intelligent business processes. Manage your decisions consciously and not by accident.

You are welcome to join us for EMEA Convergence in Barcelona on Tue, Nov 4, 2014 to watch this in action at the keynote or later in afternoon session on Wed, 5th Nov 2014.

Contributors: Royi Ronen, Akshey Gupta






Источник: http://blogs.msdn.com/b/axsa/archive...-azure-ml.aspx

[Diman]

Интересно бы на реальных данных посмотреть, как себя модель покажет с учетом какой обычно бардак бывает в данных у пользователей системы.