I am delighted to present my new blog – AI Business Transformation Playbook for Executives. originally posted here. I get into the nuts-and-bolts of AI Systems Solutioning in this rather lengthy blog but the “First Ten Plays” at the end summarizes the key steps. I look forward to your thoughts and comments.
“AI, IoT & 5G” – the confluence is a “perfect storm” of business opportunities that will appear in the next few years. This is an article on how executives at enterprises can get ready to thrive in this milieu. My focus will be on value-creating business opportunities and how to grasp it, written in an easily understandable and logical manner incorporating best practices that I have learned myself or seen in the past 15+ years.
We have a few years of AI applications in business under our belts by now (in 2019). The executive lament has been that while there are many PoCs, most do not mature into significant revenue generators. This should not be surprising – I believe that the best way to look at PoCs are as “early startups”. Much like startups, very few PoCs go on to become unicorns!
There are two approaches to AI business projects –
2. Major transformation projects
They have two implementation routes –
a) Refactoring existing products
b) New product business initiatives
I will focus on 2-b) but many of the lessons are also applicable to other combinations. In this article I will use Artificial Intelligence (AI) as a “suitcase” term encompassing Machine Learning (ML), Analytics, Data Science, etc. I also state my personal beliefs at the outset – yours will be reasonably close to these; if not, you can adopt these beliefs for the purposes of this article.
1. AI is “applied” ML. There is more to AI but applied ML is 90% of AI in 2019.
2. IoT is an integral part of this transformation – “IoT turns Products into Services”.
3. 5G will be a major enabler technology – path to distributed intelligence.
4. ML & AI will usher in a revolution like the Printing Press in the 1500’s – both reduce entropy!
5. Learning paradigm shift coming . . . from “model-free” to model-BASED!
The last two beliefs are not directly germane to this article; however, they are worth considering on their own. For more discussion, see “Machine Learning – Industrial or Printing Press revolution?” where I point out that while Printing Press augmented Memory, AI will augment Intelligence. The last belief is an important one that should be on the radar of AI technologists (see “Future of Machine Learning and AI . . . the BIG opportunities” ); in virtually every technology field, the transition to Model-based methods have been game-changers in terms of accuracy and utility.
A General Scenario
Let us consider a specific AI business example before delving into the details of each step of AI solutioning. Here we take a overall “systems” approach and not just point solutions applied to a particular task within business operations.
If you have been to Japan, you know that convenience stores (such as 7-Eleven) are everywhere. These stores provide literally anything that a neighborhood shopper will need in small quantities. Some stores are larger with more products varieties staffed by a store manager and a couple of store clerks. One of the serious problems affecting convenience stores in Japan is the demographic-shift driven labor non-availability. Japan faces a lack of people due to aging population to perform stores operations. As such, a solution where AI will *augment labor* is highly desirable.
The key elements of the AI solution are (1) Robotics, (2) Natural Language Processing and ML. These elements are stitched together with IoT and 5G.
Autonomous Robots (AR): There are 2 major activities that AR will accomplish – (1) Complete restocking of all shelves between 2 AM and 4 AM, (2) As-needed restocking of specific SKUs that are running low during busy periods (such as towards the end of lunch hour).
Natural Language Processing (NLP): For engaging the shoppers to find out their wants and needs, we can implement NLP systems in the ARs and at fixed Q&A stations. Customers can converse in Japanese, English, Korean & Chinese naturally and let their desires be known. AR may also ask shoppers waiting in the automatic checkout line if they found everything they wanted on the shelf.
Machine Learning: One of the keys for success is the tight coupling among shelf availability, backroom storage levels and external supply chain. The new system will integrate the information gathered by ARs via video and voice. Supply chain inventory data will be tied to backroom levels to update existing reordering system for timely delivery from external suppliers.
In a systems approach, integration into store chain’s ERP system cannot be overlooked. Descriptive, predictive and even prescriptive analytics results will be displayed on dashboards for consumption by Operations and executives.
While addressing the labor non-availability issue head-on, this AI solution will minimize the out-of-stock problem with robotic shelf replacement which will increase same-store revenue. Improved customer satisfaction will lead to “stickiness” – up-selling opportunities will exist when there are repeat customers. This AI-assisted Retail Store solution can be extended to include product recommendations, coupons and discounts to selected shoppers to increase same-store sales and shopper satisfaction.
Role of IoT & 5G as Enablers:
Sensors placed on store shelves and backroom, video tracking of product SKU facings and depths as well as shopper traffic sensors form the IoT infrastructure to enable this solution. With the arrival of 5G with 1-msec latency, the heavy computation processing (for visual, route planning and NLP) can be offloaded to the Cloud, thus reducing the hardware to be carried on the robot and battery size. This will make the robots significantly lighter in weight and cost. There is a small possibility of robots bumping into a shopper and causing physical harm; however, since the robots are light-weight, physical harm will be minimal.
General AI Business Solutions
In the previous convenience store example, we have seen the overall systems architecture for an AI business solution with ML, Robotics and NLP as its components and IoT and 5G as key enablers.
Now, let us step back and look at how we build such a solution starting from first principles.
Why do we do AI in Business?
AI when applied to the enterprise has the following unique value proposition –
“Do more at higher quality with better UX”
In our convenience store example, “Do More” product sales; “Higher Quality” refers to elimination of Out-Of-Stock problem and “Better UX” relates to natural language interactions with the shoppers. We should achieve some or all of the three value propositions in all of the AI business solutions that are developed.
What does the Customer want?
One can make the case that what the end customer wants is a Service and not a Product. I do not necessarily want to own a smartphone; if I can surf the web, talk to my family and play games in a personalized virtual “dome” surrounding me, I do not care to carry around a piece of hardware in my pocket. You can think of many other examples.
Providing a useful service that sustains requires “closing the loop” with the customer in it. This loop typically contains machines and people (customer service, coaches, etc.). Let us consider a concrete example.
Suppose you live in the city but your aged mother is in the small town where she was born in many 100’s of miles away. Clearly, this is an anxiety producing situation for you and your mother (and other relatives and friends). You will happy subscribe to a Service that helps keep your mother safe and you sane.
It is not enough that you install a Home IoT system that tracks her activities (using devices such as a Fitbit, video, IoT pill box or phone), you need an AI system that munges all this data, comes up with a dashboard with an overall Happiness Score that you can scan. Her healthcare providers will also be in the loop, drilling down into the Score when it is outside a satisfactory range.
The reason that an AI solution is highly desirable is the following:
A human processing the loop is not feasible –
o Too much data
o Tiresome (non-events most of the time)
One of the key tenets of when and where to apply AI is to identify tasks within a job that are repetitive and replace it with AI solutions. This tenet is also a great starting point for identifying AI opportunities in any enterprise situation, be it on the shop floor, agricultural field, utility grid and so on – “AI does not displace JOBS; AI can substitute for repetitive TASKS within that job”.
AI-driven Business Model Transformation:
Many experts see the transformation from Product to Service as the driver of the next Industrial Revolution!
In Industrial, Commercial, Infrastructure, Consumer, Medicine and other domains (such as Finance and Law to a lesser extent), AI and IoT (with 5G) together will drive the transformation of business from Product to Service. This will impact the business models greatly – instead of one-off sales, a service is best offered as subscription service (freemium to tiered levels). After further evolution, the model may evolve into a client-partnership where the payment is based on outcomes (there are already stories of GE Aviation and Rolls Royce selling Thrust per Hour instead of jet engines to the airlines!).
If you are a maker of gadgets, you may develop a device based on video that can detect out-of-stock for a retail store but what that client really cares about is a full shelf so that the shoppers do not leave empty handed. You will see this type of Service business models springing up in all enterprise and other domains in the next 5 years.
Putting AI into Practice:
In this and the following sections, we will describe the best practices to develop AI solutions for business model transformation from Product to Service.
One of the best ways to convey business value and convince decision makers is to use the language of Analytics – Descriptive, Predictive and Prescriptive. Then, we will link them to ML methods and make explicit how they are related to artificial intelligence. In human intelligence, there are 4 main processes in sequence: (1) Sensation, (2) Perception, (3) Cognition followed by (4) Action. Sensation is the realm of sensors and IoT, networks and tables of data in the business domain.
Once the data are collected, the first quest is to identify patterns in them; structure in these patterns are the “information” that will make the rest of the steps work well. Humans are good at detecting patterns but it turns out that there are undetectable subtle patterns which will require specialized ML tools such as Deep Neural Networks. In “intelligence” terms, this is Perception. Descriptive Analytics is the human-accessible display of these patterns and summary dashboards.
Once the human brain perceives something, it can “cogitate” on it. ML methods such as constrained and unconstrained optimization provide the tools for predictions. We have decent function approximation and combinatorial optimization methods for Predictive Analytics.
Descriptive and Predictive Analytics have been around for a while and are well-understood by the cognoscenti. *Prescriptive* Analytics is less clear because such analytics is not readily available – but this is where the rubber meets the business road! This is where AI Solution TELLS you what to do, not just provide analysis and insights. “Action” is a synthesis step that will lead from the “what-if” analysis done using methods such as simulation or causality analysis which will provide the rationale for prescribed actions building confidence in the AI Solution.
In a business context, all THREE Analytics solutions are necessary. It takes the combination of dashboards, predictions and action recommendations to truly benefit a business. When we combine all three Analytics in the context of a Service as discussed in the previous section, we enter a virtuous cycle that continually enhance the Service, thus achieving “Do MORE at HIGHER quality with BETTER UX”!
AI Solutioning is a Team Sport:
Turing our attention to developing and deploying AI solutions, we note that many of the tools and methods in the ML row of the last table are fairly esoteric still; fortunately, AutoML is maturing fast with solutions available from major software companies and purpose-built startups.
But I have no confidence that all the steps of development in the last section can be automated or even “democratized” – meaning anyone can do it. This is because there is no definitive way to test the outputs of ML solutions! Each trained ML model will produce slightly different results due to the stochastic nature of the learning procedure. This uncertainty can mask deeper non-robustness.
The major stumbling block (beyond qualified resource availability) is the very first step of Problem Formulation! If this is done right, the rest of the development will be accessible to qualified data science resources with their automated ML tools.
Problem formulation challenge goes beyond Feature Extraction which itself will improve considerably by applying human ingenuity and broad knowledge of Mathematical theories and tools. The simple fact is that data as given may not be the best form in which to detect patterns in them – an example is a complicated looking time series which may have a very simple spectral structure (only an expert knows the many different mathematical transforms that lead to many kinds of spectra and that not all transforms will reveal the simple “peak” patterns!).
Further, knowing what feature elements ought to be included in the feature vector so that the end result is more likely to be robust is a “black art”. One may claim that you can throw “garbage in” but Deep Neural Network will take care of it is not a recommended approach – wise composition of features and transformed features to fit the business context is a pre-requisite for robust AI solutions. With this in mind, how do we put together teams that will produce good AI solutions?
As the section heading says, AI solution development has to be a team effort. I do not believe that this is a passing phase since all the skills required will never be found in one individual developer – the training and thinking style required are too diverse.
I think of one AI solution developer full-time equivalent (FTE) as a composition of 3 individuals with differing skills: Data Science (1) Guru, (2) Engineer/ Developer and (3) Biz Vertical specialist. This is the Data Science Trio (DS Trio) three-in-one!
For each of the three phase of solution development – Descriptive, Predictive & Prescriptive – the level of person-hours contributed by the 3 resource types will vary. For example, DS Guru will be more involved in the early phase of each while DS Engineer will be more active during the middle and later stages, etc. The dynamic nature of resource requirements dictates novel approaches to resource hiring and deployment plans; of course, you will need more of DS Engineers than the other two types.
AI Systems Solutions for Services
While “digital twins” are for Products, “Digitalife” is for Services! Digitalife emphasizes the fact that the solution fully encompasses a specific need of the customer. AI Solution is a
combination of humans and computers in a network where the individual steps are abstracted away and highlight information is presented at the right time to customer. AI components of ML, Robotics and NLP play their parts with the support of IoT and 5G technologies. Such business models are ready to be put into practice at the present time.
“FIRST 10 PLAYS”: How to develop and deploy AI Systems Solutions
1. Analyze the various jobs in your enterprise; assess which tasks within a job are repetitive and require less interpersonal interactions. Automate those tasks using AI.
2. Form a virtuous cycle incorporating those tasks, experts and customers into a “Digitalife” service. Assess the revenue potential from various subscription models.
3. Assess your AI Systems Solution’s social impact; Architect in security and privacy at all levels.
4. Many of the AI methods such as ML, Robotics and NLP are reasonably mature; do not wait for them to be bullet-proof before embarking on a Digitalife service offering – set expectations that you will “perfect as you go”.
5. Utilize IoT and 5G proactively to close the service loop.
6. Interface with the enterprise ERP using all-three Analytics: Descriptive, Predictive and Prescriptive. Assure that Operations and executive interfaces have great UI/UX.
7. Form Data Science Trios to develop AI Systems Solutions; deploy DS Trios planning for their engagement in each project to be fluid depending on the stage of development.
8. Use automated AI tools carefully; democratization goes only so far – in AI and ML, things will seem to work okay but there is no deterministic way of testing performance as discussed; this can lead to instability in the service offering over time after deployment.
9. Brittleness of AI /ML solutions are best countered by insightful problem formulation and an ingenious expert paying attention to feature selection. Remember that “garbage in, garbage out” situation may be masked at launch due to the forgiving and non-global-optimal nature of AL/ML trained models.
10. PoCs are like “dipping your toes in the water”; instead commit to significant Digitalife Services. Develop using agile methodology where sprints are for “PoCs” in addition to customer validation.
With AI Systems Solutions, transform Products to Services and “Do more at higher quality with better UX”!
I am convinced that the time is NOW to transform businesses along the roadmap we discussed in the Model Transformation section in all major business segments. It is heartening to see that most major enterprises have started on the journey; it is time for small and medium businesses also to join the AI revolution.
Dr. PG Madhavan’s career in corporate technology includes developing multiple AI and ML startups for NEC X, Inc., a subsidiary of NEC, as its CXO and product leadership roles at Microsoft, Bell Labs, Rockwell Automation and GE Aviation. PG founded and was CEO at 2 startups (and CTO at 2 others) leading all aspects of startup life.
He has led the development of large-scale ML products at major corporations (GE Aviation, Rockwell Automation and NEC as well as other software solutions at Microsoft and Lucent) and startups (Syzen Analytics, NEC startups and Global Logic) involving ML algorithms to cloud software development to business operations.
After obtaining his Ph.D. in Electrical and Computer Engineering from McMaster University, Canada, and Masters in Biomedical Engineering from IIT, Madras, PG pursued original research in Random Field Theory and Computational Neuroscience as a faculty member at University of Michigan, Ann Arbor, and Waterloo University, Canada, among others.
PG’s recent major contribution in Data Science is the creation of “Systems Analytics”, a blend of Systems Theory and Machine Learning (published in 2016) providing a pathway to formally incorporate “dynamics’ into Machine Learning.