6 Key Tenets To Get Started with Enterprise AI
Recently I completed an executive program on the impact of AI on business strategy from MIT Sloan School of Management & MIT Computer Science and Artificial Intelligence Laboratory (CSAIL). While the program served many different needs of its varied and global student base, for me the key goal was to cut through the hype and really understand the basic tenets of this new set of technologies and does it really bring about the paradigm shift promised by the media and so called experts everywhere.
The following writeup is first of few that I intend to validate my understanding of the subject and interactions with the global human network build during the course. Many of the thoughts are also inspired by 20 years of experience of accelerating growth at Fortune 100 companies in different capacities involving emergent technologies.
Many business technology executives are excited about the application of Artificial intelligence (AI) in their respective businesses. There is no doubt in our minds that RPA (Robotic Process Automation), ML (Machine Learning), and AI (collectively sometimes referred to as Intelligent Process Automation (IPA) are presenting incredible opportunities to reimagine every business process in any company. However, while embarking on this journey of adding intelligent algorithms to your business context, please be aware that just because a powerful silicon chip based graphic processor can quickly identify a cat from a dog or can beat AlphaGo human experts doesn’t mean it will successfully solve your business problem(s) and help you get competitive advantage in the marketplace. There is more to these technologies and the teams that implement them than what meets the eye and so its harder to know how to move the proverbial needle.
So here are 6 best practices to get an AI initiative started in your company.
1. Spend time to really understand the actual business case to be solved using ML/AI
Though many people would like to believe the reverse, its now well understood among those in the know, that most machine learning applications will augment and not necessarily mass-replace, human efforts. When implemented, such AI enabled solutions will change what and how your teams are actually doing now, — in the new target state many non-specialized tasks will be eliminated and replaced with skilled tasks that require educated judgment and domain experience.
Given that, as an experienced and senior business executive, it’s imperative that you really hone in on the actual business cases that the organization wants to apply AI and the corresponding benefits and associated trade-offs in implementing such a solution. While someone can surely help you build and run a recommendation engine (Facebook’s ‘People you may know’, Netflix’s ‘Other movies you may enjoy’ are good examples of AI based recommendation systems) or plan your supply chain control towers better, here is a rule of five to structure your business cases for RPA,ML, AI
- applications that improve decision-making or prescribe a decision
- applications that improve client / customer experience
- applications that lead to improvements in business operations
- applications that generate revenue
- applications that can help predict or prevent fraud or risk
Additionally, here are a few use cases with verifiable business outcomes in various industries and subdomains that are already using machine learning and AI technologies — if there is a close match with your need then it’s easier to get started.
2. Know the ML/AI specialization you will need to address the business challenge
We are at the dawn of the age of AI. The AI we are looking at now is made possible by huge advances in computational power, wide availability of structured and unstructured data created at a high speed, and engineering sophistication. Yet many organizations are struggling with their attempts at so-called digital transformations and are blindsided by the many different ways machine learning can change the way they work and serve their customers. The key reason for such gaps in understanding is the of lack of qualified talent who can enable multi-sided conversations between business stakeholders, customers, and the PHD(s) sitting in your IT department who knows the difference between deep learning, machine learning, and AI.
Different business use cases demand different kind of AI based solutions. While there is a sudden surge in the generalist approach like “give us your data and we will give you all the models you will ever need”, my recommendation is that deal with any such generalization with a healthy degree of skepticism — almost all advanced analytics problems demand bespoke solutions and cookie-cutter approaches should be critically verified.
In many cases, I hear people conflating Machine Learning, Data Science, AI, Deep Learning, and Statistics when all of these are distinctly different and requires specialized training and experiences. For example, an image processing data scientist who is going to help diagnose diseases and recommend personalized treatment plans based on medical scan imagery need to know when to use Deep Learning, how to perform augmentation while someone using NLP to better understand and respond to customer complaints (KLM chatbot for customer service is a great example of this) may bring in experience of word embeddings and sense of the grammar for the language involved. In both of these cases, the key people need to know when to apply supervised learning (leading to classification and regression analysis) vs. unsupervised learning (leading to clustering) or reinforcement learning (to enable automatic decision making).
In many ways, what I’m saying is while it’s important to understand the business use cases, its equally critical to know the technology fundamentals to apply the best techniques and implement solutions in a reasonable timeframe and not blind-trust cookie-cutter systems that promise to get it right every time.
3. Plan and budget for data infrastructure set up
If you work at one of the many companies interested in machine learning, but have not invested time in putting together an advanced data analytics governance process for their own business, then please be mindful of your new ML ask. In order for machine learning programs to function successfully and give you back answers to your meaningful questions, there is usually a need of copious amount of relevant data and the proper insights to know what information should be used to build the models that gives the biggest bang for your bucks.
For example, if you are processing sensitive data worth hundreds of thousands of dollars like healthcare records or you are working with government data then you have figure out how to anonymize data sets so that PII is not being traded and invest in encryption technologies before sharing the data inside and outside the business units.
No advanced data analytics team can estimate accurately the cost of building an enterprise grade infrastructure for predictive analytics without “touching”/”profiling” the actual data. In reality, a big part of such AI projects budget and time is spent on gathering, cleansing, and provisioning the data before proceeding to the modeling stage. How and what you collect, store, and distribute as data across your organization and beyond needs to be well thought through if the goal is to scale the ML based solutions.
Machine learning systems rely significantly on non-machine learning tools has been well-explained by a simple diagram in the seminal paper “Hidden Technical Debt in Machine Learning Systems”
There is a recent crop of venture funded companies (e.g. DataRobot, Sparkbeyond, Algorithimia to name a few) who are claiming that they have successfully set up the computing environment needed to host and run high performant ML algorithms, in a convenient way, helping generate all kind of precious insights that your business needs or didn’t even know, all they are missing is “your” data. Without doubt, such platforms can help facilitate easy deployment of ML models by automating some of the DevOps work needed for customer auth and permissioning, yet none of them can claim to truly provide any meaningful insights without investing in bespoke tuning of the models based on your real data. In many cases, leading companies are more than willing to create their own data technology pipelines and build platforms for continuous experimentation (rather than paying some vendor who have glued together open source tool set to serve the same).
4. Create a team that has experience in implementing end to end solutions
Almost all winning algorithms on Kaggle never reaches to production environment though the brightest engineers with in depth understanding of ML algorithms regularly participate on this site. My guess is you don’t want that when you are investing in an advanced analytics to address an important business challenge in your organization.
While you need specialists MLs to do your super refined algos you will still need full-stack cloud engineers to build data — ML pipelines that allows for smooth deployment and subsequent monitoring. Once the ML model is in place, you will need talented horsepower to integrate that your existing solution and if it’s using Deep Learning then the stuff needs to run on GPUs for performance and scalability. Additionally, once deployed, the solution will change the existing workflows and possibly transform the associated human roles and responsibilities.
If your team is going to concern with writing algorithms only and have no time for making sure it’s running smoothly in production environment, then find someone who is able to take care of end to end pipeline.
Without the data pipeline, your project will possibly be not approved and funded by the CFO. Also, if you have not planned for change management during this AI enablement roll out, you are unlikely to be successful in getting the services adopted by all the stakeholders in your company.
Keeping in mind the end-game, how to reduce the cycle time for making decisions (data->insights->actions), is critical for enterprise adoption success.
5. Create a team with an appropriate mix of product managers, technologists, and academics
Machine Learning is by far most heavily researched field of Computer Science (CS) at the moment. It has brought together world’s sharpest minds — brightest mathematics and CS researchers for good. However, not many ML scientists, with a strong academic research pedigree had opportunities to work on analyzing, solving and successfully deploying an AI-based solution to usher in business success. So it’s strongly recommended to set up team(s) that will approach the problem areas in your company, with a combination of scientific curiosity, deep technical expertise, business dexterity, and sense of ethics.
For example, while real time media buying can be done by complicated predictive algorithms conceptualized by brilliant data scientists and implemented by experienced engineers, you will still need business savvy product managers and/or online-ad SMEs in the team to make sure that the business processes for such transactions are executed in the most optimal way. Also, an ethics review is warranted of such powerful algorithms to ensure that the automated ads buying and selling processes are not biased for or against a particular segment or company, unfavorably over others.
Machine learning algorithms work towards a mathematical goal set by the data scientists and in many cases can have inherent bias in the decision making process based on the data that was used to train the algorithms. More times than often we like to believe, the algorithms may end up ranking people in a way (provide loans to people from a certain background only or harsh court judgements against people of color) that is reflective of the bias in the training process.
Additionally, advanced deep learning systems comes up with answers and recommendations that are sometimes hard to explain. What happens if an algorithm used for automated analysis of radiology reports recommends a certain disease and course of treatment while being unable to explain how it arrived at that recommendation in a way that is well understood by the doctor and the patient. Situation like this needs the presence of academics and related experts who can walk the stakeholders through the process (explainable AI) or help design an adversarial network to prove the validity of the outputs from the machine.
AI has its roots in many different disciplines and so implementing an AI enabled solution in any company needs to be a multidisciplinary exercise implemented by a cross functional team of product managers, technologists, and academics. While there are many good stories (here, here, and here) of how to set up an AI team, remember that your context and budget are the primary factors that will shape your team size and structure.
6. Anticipate the impact and implement change management
ROSS is an AI powered legal assistant: employees in a law firm can query ROSS for any questions related to case details or citations, and ROSS sifts through massive volumes of publicly available law documents to return an answer. However, ROSS is still unable to predict what kind of citations are more valuable than others for its human colleagues as the necessary context is still with the humans in the firm.
Not many people, irrespective of their professions, have yet fully thought through parts of their job that can automated. It so happens that most routine, procedure driven business processes, and structured decision making can be best done by AI enabled machines while more speculative, imaginative, intuition, insights based decision making are best performed by our human colleagues.
Given this, there is paramount misconception in the minds of many of our workforce about their job security. While there is going to be removal of jobs for sure, there is also going to be creation of new roles and jobs. As a senior business executive, embarking on enterprise AI initiatives, it’s imperative that you think through the process of change management including and not limited to new roles, up-skilling, and training for the workforce involved.
If possible, our suggestion is that during the initial phases of AI enablement, focus on a set of initiatives which will not only improve financials for the organization and create competitive advantage but also create positive impact to the workers in your firm. Initiatives that result in upskilling and cross-skilling employees to leverage intelligent machines will go a long way in minimizing anxiety and resistance among employees. At the end of the day, all businesses are people businesses and a workforce weary and fearful of change will not help you in executing your strategy for business transformation using AI.
Moving Forward
Based on the inputs from many and my own reading, I see a future where large enterprises (and not only the tech behemoths and start-ups) will start operationalizing AI like none other, and they will have to conceive, build-buy-borrow tools and products to start automating processes bespoke to their context. Being optimistically cautious about enterprise AI going mainstream soon I hope these 6 principles will provide key markets to guide the journey. Open to hearing thoughts and other point of views.
With a step-change innovative technology like AI, I believe most companies will overestimate the short term effect while underestimate the long term possibilities of collective intelligence — where people and computer be connected so that — collectively — they act more intelligently than any person, group or computer has ever done before.
This future is not going to be an easy one — we are going to witness a war on talent across disciplines like data science, UX, cognitive analysts and many false starts before hitting the jackpot on getting AI to act as a force multiplier, helping enterprise allocate resources effectively and efficiently at scale.
Hope this is useful for senior execs starting to think of implementing AI in their business processes.
This post is part of an ongoing series enabling enterprises outside of Big Tech to get started with their enterprise AI strategy, finding and keeping advanced analytics talent, creating the perfect setting for the right people to work at their best, customer obsession, and nurture world class customer engagement while undergoing digital transformation.
If you like what you read and want to use this content for any presentation or business case or anything that makes sense for you, please let me know how you plan to use it. Open to listening to critical comments and constructive suggestions.
