Is your business planning to consume AI services?
(20 minutes read)
1. The What, Why and How of AI?
It will be presumptuous to say if anyone has not heard of the term AI (Artificial Intelligence), and it will be foolhardy of the IT Service Provider to not mention about AI in their conversation with existing and prospective customers. But, for both the Service Providers and the Customers, it is difficult to shrug off the expectations created by Science Fiction and marketing hype; and unlock the potential of AI to deliver practical and reliable solutions for the business.
The domain of AI includes techniques like, Machine Learning, Deep Learning, and Cognitive Computing. All of these are related to the ability of machines to analyze troves of data to support decision making effort – by a human or by another machine. And, sadly, this also is the limitation of AI. Age old computing adage of GIGO (garbage in garbage out) applies here, that is, any inconsistency in the data will be reflected in the final result. Similarly, the more and better the data, the more accurate will be the AI system.
According to IDC, global spending on AI and Cognitive by 2021 will hit $ 52.2B. At the same time, Gartner also predicts that “85 percent AI projects will deliver erroneous outcomes due to bias in data, algorithms, or the teams responsible for managing them.”
Though AI has been successfully integrated into the businesses of technology companies (to name a few, Google, Amazon, Facebook, Uber, Microsoft), but the user industries which lack AI capabilities within the organization (thanks to relentless focus on cost cutting by outsourcing) are lagging, and have to gradually but surely work towards integrating AI into their business processes, products and services. A reliable Partner with proven track record is going to be key to success on this journey.
Here is an attempt to lay out key considerations for Leaders planning to create new value for their business with the help of AI solutions. But, before I get there, let me spend few minutes (words) on helping the Leaders and Managers sitting on the fence to decide why is AI so relevant for the businesses in this Economy?
Why is AI (or Machine Learning – ML, in particular) so relevant for businesses in this economy?
We are moving from an era of “On Demand” consumption of goods and services to an era of “Predictive” consumption of goods and services. And, Data is driving this new, real-time economy, which means that resources in an organization will no longer need to be optimized alone, but directed towards production of goods and services by anticipating demand in advance, and delivering personalized experiences with those goods and services to end consumers.
The role of analytics in an organization, until now, was to optimize the resources so as to deliver greatest value at least cost. The same data that was used to optimize value can now be combined with new sources of data, and with the help of ML, create new value required for the organization to compete in the new economy.
Each industry will create new value in different ways. The retail industry will create a “responsive” supply chain, create trends, build brands, have deeper understanding of consumer preferences to deliver customized products and services by markets and by consumer segments.
The service industry will have better forecasting models to deliver right products and services to the target consumers.
AI combined with Industry 4.0 will help the manufacturing industry reduce downtime in their operations by optimizing the life of components, predicting failures and if failures do happen then proposing supply side adjustments to minimize impact on meeting demand.
There are numerous use cases of AI in healthcare, from diagnosis to drug discovery to patient care to customized insurance. There are several AI solutions already deployed in the healthcare industry, and also there is an ongoing active research in this field.
The agriculture (food grain and livestock) industry is in major need of transformation to feed the growing population of the world. Outside the efforts of tech farming (vertical farming) to solve the problem of food, AI in agriculture can help in deciding the nature and quantity of crops to grow at a particular location. It can make farming efficient, supply chains efficient, increase yield, leading to prosperity for farmers (much needed for developing countries) and sustainable growth.
Similarly, Sports, Media, Entertainment, Telecom, Travel & Transportation industries are deploying AI solutions to create new value for their businesses.
In order for organizations to succeed in this new economy, they need to treat data as an asset, just like human and capital assets in the organizations. How this data will be mined to extract latent insights, patterns and suggested actions will define the competitive edge for that organization.
New sources of data are becoming readily available like never before – sensors from IoT systems, social media, mobile devices, web, and business transaction data in data stores. Data governance in the organization should therefore be centralized and tasked with the mandate of making data readily available from across the organization for creating new value for the organization.
This “new value” from the data can be created with the help of Machine Learning.
Machine Learning – a subset of AI – ingests this massive amount of information to learn from data by identifying patterns and relationships within the data itself. Though ML has been around for decades but with the availability of new and large data sources, and elastic infrastructure (compute and storage) available on the Cloud, it is helping organizations gain competitive advantage in the marketplace. Machine Learning is not magic, and as we will see soon, it is tightly coupled with the analytics landscape.
Back to the “how” part. It starts with a Strategy.
2. The strategy
AI should be viewed as a capability, that is, it is a means to an end, not the end in itself. AI brings the capability of “intelligence” to existing products, services and processes, thereby making them more potent – as a competitive edge for the organization, or as a unique differentiator in the crowded marketplace.
Andrew Ng recommends that AI be adopted as an enterprise-wide decision-making strategy. Therefore, AI should be integrated with existing decision support systems in the organization.
The strategy around using AI should feed into the business strategy and it should take into account how people, process and technology will come together to bring that strategy to fruition. Technology is going to be easy part of the equation. But, transforming an organization to “Data -Wise*” organization will require it moving towards a data driven decision making culture. This shift in culture will deliver long lasting impact and it will aid in adoption and implementation of AI across the organization. The cultural shift happens when people and processes become Data Wise as well.
*Data Wise – you know your data, you understand the importance of data, you treat data as an asset and you are eager to discover new sources of data within or outside the organization to create new value for the organization.
AI is not a revolution but an evolution of analytics landscape. The AI strategy, therefore, should be backward compatible with Analytics platforms.
The traditional application has a user interface, business rules and data. These business rules are coded and rarely change, unless the business changes. However, with AI, business rules and data meld into one. AI analyzes the data and suggests rules based on the data and model it is constructed of. In the traditional approach, you tell the machine what to do. On the other hand, an AI system draws inference from the data and suggests what to do next and learns (new rules) over time. Thus, avoiding the need to explicitly rewrite business rules.
This fundamental difference in how AI applications are different from traditional applications means that staffing, building, deploying and running an AI application has to be different from traditional application build run cycles. This is an important consideration when putting together a core team who will be responsible for executing the AI strategy for the organization.
As with any transformation initiative, there is a gestation period and the investment in AI will not pay off immediately. Be patient.
3. The Business problem
AI is not the right solution to every problem. Many people may disagree with this. But, understanding which problems to be solved with AI is a crucial step towards adopting AI as an enterprise wide decision-making tool.
Focus on simple problems to begin with. The problem should be well defined and the data to solve the problem should exist, though may not be available readily. The standard business logic is not sufficient to solve the problem because there are complex nonlinear patterns in dataset that need to be analyzed to solve the problem.
A bad example of a problem for AI is predicting next year’s sales based on past data when the competitive landscape has changed this year.
A good example of a problem can be related to finding patterns and relationship within a dataset. Like, what factors contribute to a customer abandoning a shopping cart? Or identifying a fraudulent transaction.
The other kind of good problem can be related to predicting what will the customer buy next after buying X?
Also, it is important to understand the margin of error allowed for the chosen problem and accordingly reasonable expectations should be set with the relevant stakeholders.
There are cases where steps taken to arrive at a solution have to be demonstrated to statutory bodies or are part of regulatory filings. Such problems are not a good candidate for AI, because, remember we discussed that business rules are “inferred based on data and model” not “written”.
4. The people
Building an AI capability within the organization to create value for the organization requires an inter-disciplinary team of IT (along with Partners) and Business – Data Scientists, Systems Engineers, Solution Architects, SMEs from Business and Data Stewards. This brings together the right skill sets to build and deploy an AI solution for a business problem.
However, making AI as an enterprise wide capability so that it is embedded (where feasible) into products, services, processes and decision support systems, requires evangelists and champions in each function and geography of the organization. These champions can help in discovering new sources of data lying dormant within the organization, and champion solving a business problem with data. And, what can be a better place to breed and groom these champions than letting them experience the power of AI at workplace. There are several AI powered workplace productivity tools that can be deployed as part of executing an AI strategy for the organization.
Data is an asset, so are the people. You not only need to infuse new AI skills in the team, but also create champions who apply AI to solve problems with data. People are an important part of bringing the required cultural change in the organization to realize AI strategy and in making the organization “AI enabled”.
5. The infrastructure
Democratization has AI has led to ML frameworks from Google, Microsoft, Facebook, Amazon and Uber being made publicly available. Also, AWS and Google Cloud have released ML specific services, including GPUs for machine learning work.
However, there is a challenge associated with interoperability of different ML frameworks, and each framework have their own pros and cons, depending on the business problem to be solved with ML and how wide will AI be adopted in the organization. A careful consideration in choosing the right framework will help in the long run and can avoid the expensive tasks of migration from one framework to the other.
The horsepower (compute and storage clusters) required to make a ML system work is an excellent use case for Cloud computing because of the elastic capabilities of cloud, and reduces the strain on internal system engineers to design infrastructure necessary for ML applications. The horsepower requirement for ML system depends on the learning method and number of ML systems being deployed in the organization – expect the use to grow rapidly after the first pilot is successful.
Adopting a cloud strategy for ML is a good idea but it should take into account how the AI architecture will interoperate with the on prem analytics landscape.
Another important consideration for adopting a cloud strategy for AI is how the output from ML will be consumed – decision making, as an input to other applications or systems, stored for future use, or feed into other machine learning algorithms.
6. The model
The selection of a machine learning model depends on the business problem at hand and the nature of data available (we will see more on data next). The model needs to be tuned for the specific problem being solved, which requires iterations until the results are satisfactory and acceptable.
Humans make mistake, so do these models. The machine learning models will get things right on average but occasionally make mistakes, and in ways you least expected. Therefore, an understanding on the margin of error allowed by the problem is important in choosing the right model to solve that business problem.
As the adoption of AI grows in the organization, more and more ML models will be produced and deployed. So will the need for caring and feeding these models will grow – managing and monitoring, provide traceability and version control, and upgrading the models. The machine learning models mature in their learning and then decay over a period of time. As the environment changes from the conditions used during training the model, due to change in competitive landscape, customer behavior, demographics, the results from the model will be less accurate. Once the results from the model cross the threshold of allowed margin of error, the model needs to be refreshed by either retraining or replacing (new research keeps producing new models).
7. The data
Data fuels AI and AI hogs on data.
Many organizations struggle with data issues, like maintaining a single source of truth on customer. AI does not solve these data issue but amplifies them.
As mentioned earlier, quality of data determines the quality of AI application. Data is an asset for the organization, which can be used to create new value and provide a competitive edge. Data, therefore, should not remain siloed but be made readily available across the organization (except for security and privacy reasons). With AI becoming mainstream, organizations with best data will win.
Though building ML models is seen as glamorous by data scientists, the fact is that more time is spent on preparing data for the models – transformation, normalization, cleansing, instream analytics and preparing training data for supervised learning.
The first step is to identify data sources and where they exist to solve the business problem, which can be: a causal analysis for prediction or correlation for finding patterns and latent insights. The data may be coming from different data stores (technologies), different vendors (syndicated data), different time series, may not be labelled, and may be either streaming, batch processed or real time. Such data integration and processing challenges often turn out to be bigger than anticipated. While combining different data sources, the context in which the original data was generated has to be kept in mind, else it may impact the validity of final results.
If there is any conflicting information, missing or erroneous values in the dataset then it has to be taken care of, preferably at the source if the organization owns the data. Any bias in the data (demographic, spatial, timeseries) will result in biased results. The team working on preparing the dataset for machine learning should be cognizant of this and avoid feeding bias into the models inadvertently.
Secondly, data security is of paramount importance, especially for regulated industries. Any non-compliance here can far outweigh the benefits of sharing such data. A proper consultation is needed and ramification of not sharing the data understood before such data can be fed into ML models.
8. And, finally …
Unlike what happened with internet, AI is not a tide that will lift all the boats. And unlike internet, where late comers to the party did better than early entrants, early movers in AI will have an advantage as their AI capability matures over time, which will provide them with the advantage to compete in the marketplace.
9. References
https://online.maryville.edu/blog/big-data-is-too-big-without-ai/
https://towardsdatascience.com/role-of-data-science-in-artificial-intelligence-950efedd2579
https://www.squiz.net/learn/blog/why-data-is-so-important-when-it-comes-to-ai
https://www.kdnuggets.com/2018/08/top-10-roles-ai-data-science.html
https://www.datameer.com/blog/role-big-data-artificial-intelligence-future/
https://www.dataversity.net/data-architecture-artificial-intelligence-work-together/
https://www.infoworld.com/article/3269873/the-new-data-roles-brought-to-you-by-ai.html
Gartner publication - Preparing and Architecting for Machine Learning
Entrepreneur | Mentor | Investor
5 年I will love to hear your perspective Tarry Singh
Associate Partner, Automation Led Business, IBM Consulting
5 年Nice one Sanjeev.
at the convergence of digital transformation, renewables and sustainability
5 年Top article Sanjeev
Delivery Project Executive at IBM
5 年Good one Sanjeev