Have we entered an era of Artificial Intelligence industrialization for enterprises? What is needed to embark on AI industrialization journey?

We sure have entered an era of AI Industrialization! I can say it emphatically because I am helping a Fortune 15 company execute its AI Industrialization Strategy. I  believe that most of the Fortune 500 companies will follow suit sooner rather than later.

So, what does AI Industrialization really mean? What are the business and technical drivers that lead any company to embark on this journey?  Also, why are they calling it Industrialization of AI? What are the prerequisites needed for companies to start on this journey, to set a solid foundation and reap all the benefits?

Let’s revisit the concept of industrialization by looking at history to understand the different phases of industrialization.  If we just look at first phase of US Industrialization from (1820 -1910), it went through phases of industrialization: cotton plantations, canals and boats, railroads, immigration, banking, steel mills, automobiles, and telephones.  If we analyze it further, we realize that the key characteristics of this industrialization process were:

1.          Inventions and Innovations

2.          Entrepreneurship

3.          Funding

4.          Improved Productivity

5.          Mass Production

6.          Cheaper Production

7.          Repeatability of Processes

8.          Critical Mass of people producing goods

9.          Critical Mass of people consuming goods

10.      Improved Standards of Living

11.      Better Experiences

12.      Cheaper to Buy

13.      Increase in Real Incomes and Return on Investment

 

Now let’s try to draw parallels with AI timeline:

1956:    John McCarthy coined the term AI

1997:    IBM's Deep Blue beat Gary Kasparov

1980s:   Neural network is popularized

2010s:   Amazing breakthroughs with Deep Learning technology

2011:    IBM’s Watson beats the two best human performers on Jeopardy

2015:    Google DeepMind’s AlphaGo beats human champion

2017:    Google announces AI First strategy

Last 3-4 years: Hundreds of new AI startups have been funded from Silicon Valley to Boston to Tel Aviv to Bangalore.

What have we really achieved in AI if we compare it against all 13 characteristics of industrialization listed above? The reality is that from an enterprise perspective, we have only achieved only the first 3 out of the 13. The majority of AI accomplishments in the enterprise context have been limited thus far to inventions, entrepreneurship, and funding. Of course, I am not including what pure play Web companies like Google, Amazon, Netflix, Uber etc. have accomplished. When it comes to mass production, faster production, cheaper production, as well as achieving the ROI, AI in enterprises falls far short of true industrialization. There are always few use cases to boast about, but it doesn’t mean that AI industrialization has truly happened. It means there is tremendous opportunity to build the right AI strategy for AI industrialization.

Traditionally, AI has been a hand-crafted, POC-driven and expensive initiative but the hype and the promise must now deliver at scale. It is high time; we have to start scaling AI initiatives across the enterprise in a cost effective and consumable manner so that it is truly democratized and is accessible to everyone. The operational efficiencies by AI across the board should never be questioned. The core value proposition of AI Industrialization is to take thousands of models in lab to enable and empower LOBs by embedding AI in business processes and in digital transformation journey.

What is missing today is an enterprise enablement view of AI. Let’s take a snapshot of sample of AI opportunities in a one of largest Fortune 15 company today.

SAMPLE OF AI USE CASES FOR A LARGE ENTEPRISE

 

These use cases are typically led by different lines of business who face many challenges in scaling AI. Typical challenges in implementing and deploying AI use cases include the use of different technologies or versions, difficulty governing the process, lack of repeatability and automation, and complications with collaboration and transfer of knowledge between AI engineers.

Another challenge is that you don’t want every LOB to worry about building their own AI COE, AI infrastructure, end-to-end AI platform, building data pipelines, machine learning pipelines, machine learning operations and catalogue, experimentation platform, design their own feature stores, data ops strategy, data quality, model decay, develop their AI cloud strategy, worry about bias, compliance, and AI ethics. You want them to focus on building AI use cases only and rest of the AI infrastructure is provided by centralized AI COE team chartered to develop the AI industrialization strategy as well as governance process to enable LOBs.

The reference architecture for AI Industrialization looks something like this:

Best practices around intelligent enrichment of data and methodology to deploy data pipelines need to be developed? What is the best way to build collaboration ecosystem between engineering, data analysts and data scientists? How do you figure out where manual intervention is needed versus complete automation? How do you create a map of where data starts, how it changes and where it is viewed from cloud as well on-premise perspective? How do you measure error rate decline in production? How to build features at scale from raw data for training? How do you combine features into training data? Calculating and serving features in production? How do you know which sources are biased? What are the best practices for model deployment, retraining and monitoring? How do you design a model workflow from approval perspective in a large enterprise? How do you decouple your data pipelines from your model development and governance? How do you democratize model building so that even data analysts and power business users can start contributing? How do you shorten the life cycle of model creation from deployment to deployment for hundreds of models? What AI components should you leverage from each cloud provider? How do you create a mix of on-premise and hybrid AI cloud strategy? How do you build standardization and repeatability across the board? Who is accountable when things go wrong?

There are hundreds of questions and answers that have to be figured out and best practices needs to be developed and communicated. AI Industrialization is a necessary journey for this era. But it is extremely important that it is setup for success for the long run by team of experts who understand what it takes to build this end-to-end.

Marketing in a Machine Learning World: Simplicity is the key!

I was really inspired to write this post after talking to a senior marketing executive of a Fortune 20 company. So if you work for a Marketing or E-commerce organization then you should read it.

Perhaps, it is one of the most exciting times to be marketing in this digital world because marketing is increasingly becoming more high profile job than ever in most of the organizations! But that also means that marketer’s life has become more complicated despite having more tools at her disposal. The reality is that the marketers have to worry more about problems like how to break through the noise in this hyper competitive era; how to drive loyalty; shift from product-centric to customer-centric model; deal with the saturation in social media – everyone is a content producer nowadays; changing demographics; understanding customer’s context; how to justify ROI internally; having consensus for creative assets; the confusion around attribution model and spend; the changing relationship between brands and consumers; keeping pace with technology; sentiment analysis; omni-channel customer journeys; rise of mobile; figuring out millennial; why so many shoppers are dropping out of their marketing funnel; lack of alignment with sales; lack of trust in the customer database and so on. The list is very long and it will continue to evolve. In the marketing world, what needs to be done is usually clear. But what isn't always clear is how to do it in an optimized way. 

Is there anything different Marketers can do to deal with these problems?

Yes, some of the answers are hidden in machine learning! It's time you start thinking about it seriously! 

As we move from hypothesis driven world to data driven world, we might realize that  we don’t need more theories – we need to rely on data to help us make practical decisions. Access to hundreds of data reports and interpreting it yourself is Not what I am talking about. I am talking about data recommending you a course of concrete actions that is possible only through Machine learning (ML). If companies are betting on building driver-less cars using power of machine learning then I am sure ML can help you in solving few of your problems also. Machine Learning in marketing is considered by many as one of the biggest game changing opportunity for marketers just because we have now more data than ever and it is no longer humanly possible to make sense of the data without the help of machine learning. Machine Learning can't solve all your problems but it does help you in giving a logical path forward to deal with many problems in marketing world.

So, Why is everyone talking about Machine Learning Now?

Machine Learning (ML), a branch of artificial intelligence, in simple terms focuses on the development of computer programs that can teach themselves to grow and change when exposed to new data. Machine learning is almost like an intelligent assistant that draws from fields like Artificial intelligence, Statistics, data mining and optimization.

The reality is that Machine Learning technology has been there for decades but two trends have contributed significantly to phenomenal rise of Machine Learning:

1. Big Data – You have more data than ever. Machine Learning becomes better and relevant with more data.  A lot has been written about big data so I won’t go into too much detail.
2. Affordability - Till few years back, the machine learning technology wasn’t accessible easily to marketers – the cost to setup infrastructure and build specialized team was just very high. In the past, successful use of machine learning algorithms required made-to-order algorithms and huge R&D budgets, but all that is changing. IBM Watson, Microsoft Azure, Google and Amazon have launched turnkey cloud-based machine learning solutions. At the same time startups like Idibon, MetaMind, Dato and MonkeyLearn have built machine learning products that companies can take advantage of.  I have used some of these libraries myself and found it very simple to use but very powerful. Again, these models aren't perfect, but they're very useful.

    So, What is happening in Industry with respect to Machine Learning?

     There have been notable acquisitions in the machine learning startups in advertising, sales and marketing: Oracle acquired Crosswire for $50 million; Twitter acquired TellApart for more than $530m, Google acquired marketing-management startup Granata Decsion Systems, and Israeli Unicorn Ironsource merged with in-app advertising startup Supersonic, to name a few. Machine learning startups in marketing space like Appier (cross device marketing), Databerries (in store traffic), Drawbridge (cross device advertising), Emarsys (content personalization), Lattice Engines (predictive scoring for leads), Oculus360 (marketing Intelligence Paltform) and Personali (Uses ML to form emotional connection) have been funded with tens of millions of dollars receently. 

      Deep learning - a cutting edge branch of machine learning inspired by the architecture of human brain -is the hottest thing happening in machine learning if you consider the recent acquisitions by Amazon (Orbeus), Facebook (Wit.AI), Google (Dark Blue Labs, Deep Mind, DNNresearch), IBM (alchemyAPI) and Microsoft (SwiftKey). Deep Learning was the underlying technology which was used by the Google’s DeepMind AI who beat Lee Sedol, a legendary Go player. There is a big race among major software players for technical superiority

      Even Salesforce.com has come out with their machine learning product called Einstein and Adobe has made announcement to embed machine learning in their  offerings.

 Is anybody in your organization looking at all the innovations happening in the  machine learning world; doing gap analysis and making recommendations about  your machine learning roadmap?

What can Machine Learning really do for a Marketer?

     Anticipating customer needs is not a new phenomenon but what is truly new is the ability to respond to customer needs automatically, in real time and at scale with the help of machine learning. The most common use cases of ML in marketing are primarily:

• Finding and predicting best and least valuable customers from lifetime value standpoint
• Building personas based on customer clusters and building appropriate creative, content and services for them
• Recommending new products and content based on who you are which prospects are most likely to buy;
• Tagging content with right keywords
• Testing countless paths consumers may take through content
• Programmatic ad buying
• Optimizing moments of interests by personalization of content
• Predictive lead scoring

There will be many more use cases in your organization if you explore and go deeper.

What is Missing? Why Everyone in my Organization is not using it? 

Despite having so much data and access to machine learning technologies, the rate of adaption should be better across the board in any marketing organization. There are two major reasons: Lack of training and Lack of Agility of Machine Learning Implementations. Let me explain in more detail:

• Democratization, Simplification and Training of Machine Learning concepts - I believe what is lacking is bare minimum machine learning training at conceptual level for marketers – from executives to marketers who  are in trenches - so that they can start having the right conversations.  The whole concept of using machine learning needs to be simplified and they should know how to make it actionable. Customized machine learning training should be developed for marketing and e-commerce organizations. No, marketers don’t need to become data scientists for that. They just should be able to exploit the machine learning technologies already existing either in their organizations or outside in the cloud. The idea is to develop a primer for them so that they can start having meaningful conversations with data scientists.There are some very good books out there for machine learning but they are either too technical or very high-level to be actionable. Marketers just need to articulate the problem in a better way with some understanding of ML! Defining the machine learning problem precisely is the hardest part and it should be initiated by marketers – not data scientists.

Lets take a look at one of the most popular machine learning  algorithms and see how a marketer needs to articulate a question to a data scientist:

Example of Marketer’s Question	Machine Learning Algorithm by Data Scientist
How much should I spend on advertising to achieve certain sales volume?	Linear Regression
What are the chances that the customer will a) stay a customer b) re-purchase a product or c) respond to a direct mail?	Logistic Regression
Can the machine help me discover segments or clusters in my customer base by using already known factors in my customer base?	K-means
How can I match my customers’ interests with the description and attributes of my products?	Support Vector Means
How can I predict customer retention and profitability?	Random Forests
How can I identify the right audience to target on social platforms and what should I say based on what data tells us?	Deep Learning/Neural Networks
How can I develop Amazon-like recommendations on my website?	Collaborative Filtering

Of course, you will have to dig deeper and iterate over each of these questions. But asking the right question is the first step to start meaningful conversation with data scientists and technologists in your organization. This will also relieve too much burden from data scientists as they are generally overwhelmed with work and the complexity of their work is never understood properly.

• Agility and workflow of Machine Learning Implementations – You can have a great team of machine learning engineers and data scientists but the process of deploying applications based on machine learning is usually so slow. How do you deploy models in 2 weeks versus 3-6 months? So there is a disconnect between what marketers are being asked to do versus what is produced by data scientists from timing standpoint. It becomes harder if you are dealing with millions of customers and large data. You will probably get two different viewpoints if you ask a marketer versus a data scientist in your organization! Agility in machine learning deployment models will not be exactly like agile software methodologies but there are many common elements. The big difference is that ML agile methodologies will be data driven. To solve that problem you need to customize good agile project management practices and apply them in ML context for your organization. But again you can’t do much about it unless you are trained, as mentioned in point above, in basic concepts. Machine learning in production is becoming less about algorithms and becoming more focused on the data workflows surrounding them. Data Flow is about all the steps required to train machine-learning models in the lab, deploy those models into production, monitor and evaluate their performance, and iteratively improve those models. If your data flows are long, expensive or manual than you have a huge problem. You need to rethink about strategy and consider alternate solutions to simplify it. The last mile problem of machine learning is well known. Most data scientists will not know deeply about marketing or/and marketing systems to embed predictions into the daily routine of marketers. 

In the end, it doesn't matter whether you own marketing strategy in your organization or have an operational role, you need to start thinking about taking concrete steps to integrate machine learning in the DNA of your marketing organization.  It really helps you in developing a solid long-term marketing strategy and an optimized operational model. The best time is now! Simplicity is the  key to success in this context.