There is surely a lot of buzz about artificial intelligence (AI), machine learning (ML), and data science (DS). Views vary widely from what people are doing with AI now and what could be done in the future. The practitioners of data science (data scientists) are in very high demand, as they are the “unicorns” who will help your company implement AI and ML. All the buzz has lead to a lot of misunderstanding of how to actually operationalize AI and the level of difficulty of getting something useful out of it for your business.
To be frank, implementing AI is not simple. But, the reasons for the difficulties may not be what you think. In this post, I’ll discuss a holistic approach of implementing AI in your business.
What Is AI, Machine Learning, and Data Science?
First and foremost, what is AI, ML, and DS? Well, there isn’t a single definition. So, I’ll define them according to a couple of sources:
From the detailed definitions above, we can simply say that:
Artificial intelligence (AI) is a computer science field that focuses on making computers behave more like humans. Machine learning (ML) is a subset of AI where computers use data and statistics to learn and be able to handle new situations without explicitly being programmed for each situation. Finally, data science (DS) is the scientific process for harnessing knowledge from raw data.
What Can You Do with AI?
Theoretically, AI can do anything that a human can do and more, since computers are able to perform calculations faster than humans and can be scaled up. This is a scary thought for many people! It also doesn’t help that the media sensationalizes the capability of AI and how it will take away all of our jobs.
In practice though, computers are pretty far from being able to completely act like humans. Humans are incredibly great at accomplishing certain tasks that computers are not great at and vice versa. So, AI is actually used for much more specialized use cases (narrow AI) where they enhance or automate tasks that would not be very efficient for humans to perform (such as providing recommendations from huge amounts of data, recognizing objects in an automated fashion using images/video, or performing tasks at a high volume such as detecting fraud from credit card transactions).
We could list hundreds of use cases where AI has been commercially applied. However, we can organize these applications in roughly three categories:
Process Automation: This is where capabilities such as text analysis and image recognition can take a look at in input (such as an email or a video) and automatically perform the next task (such as sending the email to a spam folder, updating a customer’s record, or flagging equipment as likely to need maintenance).
Automated Insights: This is where large amounts of data can be analyzed to determine the next best action (such as which product to recommend, which ads are best suited for which audience, or which combination of drugs are likely to be most effective to treat certain types of cancers).
Human Engagement: This is where speech recognition or natural language processing capabilities can be used to interact with humans (such as chat bots, smart speakers, or personal robotics).
Some mature AI solutions blend a combination of many of these capabilities into a holistic system (such as a self-driving car).
Understanding AI Implementation
Determining what to do with AI is where the difficulty with implementation begins. As a business, you must first begin with the question of what are you trying to accomplish. You must determine if using AI is the right tool to help you accomplish your goal. And you must assess the costs and resources necessary to use AI to accomplish the goal.
Too often, companies begin with the technology or the method and spend a lot of money implementing it, only to find that it doesn’t really accomplish any valuable goal. This is a completely backwards way to go about it! You have to start with a goal in mind, and AI is only a means to accomplish the goal.
Sounds simple enough, right? Well, not really! We could have a chicken and an egg problem here. If you don’t understand what AI is able to do, how can you tell if the goal you are trying to accomplish should use AI? Also, AI may offer a solution to a problem you didn’t really know you had until you learned how someone else used AI to solve it.
So, the bottom line is that you have to intimately understand your current goals or problems, plus the capabilities of AI and how they intersect. Specially, you need to look at how AI can be used to solve your problems or identify problems that you may not have thought you had. Keep in mind that is this not a one-time process. Business needs change constantly, and so does AI technology.
Getting Started with AI
Generating a list of solid use cases is an essential first step. But, in order to get there, you will need a multi-disciplinary team that not only understands your business, but also understands the capabilities of AI. The concept of unicorn data scientists who are able to do it all is pretty much science fiction. You will need to involve not only data scientists, but also business leaders, technology leaders, and get buy in from the operational staff that will be impacted by the process changes introduced with AI.
Collaboration and coordination is a key requirement to success. If you are just starting out, creating a center of excellence (COE) is a good first step. But, the ultimate goal has to be to spread out the capabilities beyond the COE throughout your organization since AI applications are far reaching.
Once you’ve determined some solid use cases for AI, you have to prioritize which ones you want to tackle first based on many factors, including cost and impacts to your organization. We highly recommend not boiling the ocean and attempting to accomplish too many things at once or creating project plans that span many months or years.
AI is best implemented in an iterative fashion. Research from the Harvard Business Review has shown that companies are much more successful when they implement AI in smaller chunks vs. very large revolutionary projects. There is often a lot to learn in the process of going from ideation to production. Also, measuring the performance of AI is very important, and adjustments will likely need to be made based on results once the system is running in the real world. So, it is very advantageous to approach AI in an agile fashion.
This is not to say that you shouldn’t strategically think about how all your different use cases fit together. You should have a road map of many AI use cases and think about an overarching strategy and platform for embedding AI into as many business processes as possible. But, this cannot be at the expense of actually implementing some real-world applications and gaining the valuable knowledge that comes with that experience.
The Connection Between AI and Data
A key component of any AI project is data! Without great data, the current crop of ML and AI tools and techniques will fail miserably. Computers are able to learn based on observations, and if the observations are flawed, the results will surely be flawed as well. Some of the key considerations in assessing your data are:
Availability: Is the data that you need available? The key word here is need. It is likely that you won’t have all the data you want and/or have lots of data available that is not needed.
Accessibility: Are you able to access the data needed? How easy is the access? Can the data be accessed at the frequency needed? Do you have the technical infrastructure to access the needed data?
Representation: Is the data representative enough to accurately drive the AI model? Are there any missing records or categories of data? Is there potential for bias when using the data set?
Time Horizon: Is the data available for the time period needed. Do you need historical data? How far back do you need to go to be effective? Do you need data in real time?
Quality: Have you assessed the quality of the data? How was the data collected? Were appropriate controls in place to ensure quality? Do you have a process to ensure the quality of future data sets?
A big percentage of time is typically spent preparing data for AI modeling. Lack of foresight, infrastructure, process, and strategy typically lengthens the process. Investing in getting your data right is foundational and a huge enabler for applying AI to your business.
Creating the AI Models
Once you have solid use cases and good data, it’s time to start creating AI models. AI models use statistical and computational methods to take inputs (data) and turn them it into outputs that lead to a decision (whether automated or manual confirmed by people). The model further learns with new data and should ideally have a feedback loop that tells it how it’s doing (integrating star ratings or thumbs up/thumbs down buttons are examples of providing input to an AI model).
AI modeling is a very deep subject and beyond the scope of this post, so, we will not go into all of the technical details. Data scientists are the specialists who write, adjust and maintain AI models and need to have a special blend of skills as described by Drew Conway’s Venn Diagram that include computer science/hacking skills, math/statistics knowledge, and domain knowledge/expertise. As we’ve been discussing in this post though, data scientists are not the only ones involved in the process, so, collaboration and coordination with other stakeholders is very important.
Based on our experience implementing AI over the years, we’ve seen the modeling process run into several critical issues. Here are a few very important things to keep in mind during the modeling stage:
- Ensure your data scientists thoroughly understand the business use cases and are involved as early as possible in use case discussions, selection and prioritization. A lack of understanding of the business objective could easily lead to models that work from a technical perspective but not from a business perspective because they are optimizing the wrong metrics.
- Allow time for exploration and experimentation. In most cases there isn’t a silver bullet modeling solution that fits a given business problem. Data scientists must be given time to explore and try out different models and iterate, measure and tune the models to arrive at a great solution.
- Employ rigorous peer and business review processes. Regardless of how talented a single data scientist is, it is imperative to have others review his/her work. The work should be reviewed from a technical and a business perspective to ensure the model is appropriate and will adequately meet the business objectives.
- Allocate support and resources for systems integration. It is rare that a single data scientist can fully integrate AI models into all the systems needed to go from data collection to serving decisions. Involving the right people (such as Data Engineers, Application Developers, and Infrastructure Engineers) to provide infrastructure and application support is key to have a truly production-grade AI solution.
- Ensure AI models are continuously reviewed after being deployed to production. The modeling process does not stop once a model gets deployed to production. The results must be evaluated based on real-world data and the models often need to be tweaked and adjusted over time.
Measurability, Transparency, and Agility Matters More than Ever
Since AI systems have the capability to make decisions in an automated fashion, it is imperative to ensure that the solution is of high quality at all times. A key requirement in determining quality is to be able to measure and monitor the performance of the AI system. Furthermore, you want to put controls in place to prevent unintended consequence or damage from automated AI decisions. You want to make sure you are agile enough and can plan on being able to deactivate or adjust the AI system over time. Inevitably, something will go wrong and that is where transparency comes in. You want to be able to trace how and why the AI solution made a decision and have a plan to make adjustments in an expeditious fashion.
Conclusion: Putting it All Together
In this post we discussed several steps to applying AI to your business. We began with emphasizing the need to come up with business use cases that are well suited for AI to solve. We then discussed the need to prioritize these use cases based on several factors while being keenly aware of the probability of the solution getting to production. We discussed the critical nature of high-quality data. We touched on the AI modeling process and how to avoid common pitfalls. Last, but certainly not least, we discussed the need for AI models to be continuously measured and adjusted over time and the need to have the infrastructure in place to quickly and transparently make such adjustments. These steps are not sequential and there is a good amount of iteration involved. The diagram below puts all these concepts together and illustrates what can be pursued simultaneously and iteratively.
At Cognetik, we help companies ensure their data is of the highest quality at every stage from collection to usage, making the complexity of implementing AI into your practice much easier. Our expertise in ML and DS, combined with our products and services, can help your organization use AI to its full advantage, all while making sure AI is the best option to help you achieve your goals. Contact us today to see how we can make AI work for your business.