AI has become the next big innovation that leading enterprises and startups are scrambling to integrate into their organizations. This demand is due to the various benefits that can be gained by adopting this technology and the proven success some organizations have had after integrating AI. While it has become a key strategic differentiator for some companies, others have not seen a return on their investment. Moreover, the field is constantly evolving, making it a complicated process for startup founders to decide on the key areas to begin AI integration. Moreover, when it comes to AI, leaders often have misinformation about the scope of tasks that AI can perform, with the common image of AI being useful only for complex tasks.
In this article, we will discuss some of the most proven and quick wins of AI use cases and how startup founders can integrate them. But before we get into AI use cases, let’s look at what areas AI excels at and what tasks require AI intervention rather than human judgement.
AI vs. Human Decision: What is AI Good at?
Expert systems have been around for a long time, but sometimes all you need is good old fashioned human intuition backed by data. AI doesn't need to solve every problem out there. While AI certainly has its advantages, some problems are better solved by humans, so choosing your use cases wisely is key. You can also rely on the great natural intelligence of well-trained humans. However, there are essential prerequisites that make an AI system stand out. Here are some of them:
- Attributable problems: A few years ago, the job description for a data scientist was something like, “We're collecting loads of data, and we need someone to interpret and leverage it.” Data was thought of as something to gain insights from, and decisions were always made with a view to the “data.” With increased consumer adoption of digital products, improved infrastructure, and an acceptance of experimentation within organizations, data science can produce attributable impact. Attribution is one of the prerequisites for setting up an AI system. If the success of your AI initiative is not attributable and measurable, it's probably best to postpone the experiment. For example, an AI model that drives improvements in a marketing campaign with a well-defined control group is more impactful and attributable than tackling a customer segmentation type problem.
- Complex situations: Unlike systems based on human judgment, AI thrives in rapidly changing environments. The more complex and dynamic a system is, the better it can be leveraged with AI. For example, the challenge of offering the right discounts to the right customers during a flash sale on an e-commerce site can be successfully solved with AI. Customer preferences can change depending on the composition of the discount offer and the duration of the deal itself, and AI can maximize revenue from the sale without degrading the user experience.
- Too many options: Commonly referred to as “choice overload,” this situation causes customers to hesitate to do business with too many options. Humans prefer to make decisions with fewer options. Too many options cause them to lose interest. Recommendation systems powered by ML models can help provide customers with the right options. So, when there are too many product or service options, it is better to consider machine learning models over human judgement. Humans working with spreadsheets cannot make these decisions. ML models can play a key role in providing “meaningful” options to customers.
- Well-measured data: AI is built on data. If your system does not have data, it is better to avoid building complex systems until you collect enough data and establish high scale. For example, if a lending company does not work with customers in different risk segments or observe loan repayments over time, there is no need to rush to start investing in predictive models. The company would be better off focusing on measuring data that will later become the foundation for building AI predictive systems.
- Unstructured data: If your business or product deals with large amounts of unstructured data, hiring humans to sort and process it may not be practical. In this case, AI may be more useful.
- Multiple optimization criteria: AI can optimize multiple constraints simultaneously. For example, consider vehicle route optimization, which has both time and money constraints. This is very difficult to solve using heuristic use cases. Instead, AI can step in and take into account multiple factors such as the best route available, the fastest journey time in terms of distance and time, traffic conditions, etc., and provide the route that takes the least amount of time while minimizing money spent on gas and tolls.
Once you are confident about the AI environment and context described above, you can move on to the next step in integrating AI: identifying the use case.
Low hanging fruit
AI is great at accomplishing a variety of tasks, but what are the best use cases for your organization to tackle first? Popular culture has certainly given people a more complicated idea of AI utility. They're not necessarily wrong, but AI uses range from complex workloads to simpler, easily accomplished tasks. For startups, it's important not to tackle big problems with big solutions right from the get-go. A better way to start integrating AI into your business is to start with the low hanging fruit.
The most successful form of machine learning is supervised learning. A “collect, learn, predict” framework always works. It involves collecting data to use for training. It can then learn from the data and make predictions in the form of probability scores or predicted values. Some of the most common use cases for supervised learning are credit scoring models, constrained campaign targeting, fraud systems, and large-scale predictive systems. In such applications, leveraging AI has proven to be better than relying on human judgment. A key technique for proving impact is to measure how well a model performs against an existing baseline process.
AI is powerful when it comes to making micro-decisions at scale. Imagine you have to present a consumer with the top 6 models from thousands of product categories. AI can select a subset of n categories to recommend to the user in a way that optimizes revenue or clicks. Moreover, the underlying algorithms can learn common patterns, so they can be applied even in cases where the user has never “discovered” a product before.
For example, consider Netflix's recommendation engine, which recommends to users a curated list of movies from the thousands of movies on the platform. This list is personalized and based on the user's past history. It can also learn behavioral patterns to recommend new movies to the viewer. That is, if the engine notices that an English user has watched a few French movies in the drama genre, it might recommend international movies in other languages such as German or Turkish in drama or related genres. The engine does this in microseconds and can personalize for every user on the platform. Moreover, the engine can also fine-tune itself based on the feedback it receives from users.
Another use case for AI is automation. Automation applications that save companies time and costs are a great area for data science teams to embark on. With a growing consumer base, tasks like classifying incoming customer service tickets and manual verification of KYC are impossible to handle manually or without significant investments. This is where AI can be used as a starting point to automate workflows and speed up the pace. AI models have also proven to be a better option when it comes to processing text and images. Leveraging AI to automate tasks can rapidly improve an organization's workflow and drive business growth.
AI can generate insights from structured data, but it needs to be done with caution. When dealing with data insights, it is much harder to explain causality than to actually predict what will happen next. That being said, it is unrealistic to ask data science teams to completely avoid questions that require insightful answers. Business problems like “Why did our sales drop?” or “What are the main drivers of attribution?” are not easily solvable in the first place, and the impact is not a direct cause. In these cases, it is best to spend time establishing the expectations and risks of the project before you start generating insights.
Preparations to make before introducing AI into your organization
We have already discussed areas where organizations need to integrate AI and some basic use cases for this technology. However, there are some starting points that companies need to prepare before adopting AI technologies. Let us briefly discuss them.
Start with a baseline or naive approach. For example, if you use a machine learning model to predict delivery times, a naive approach would be to show the overall average delivery time for all orders. Benchmark your ML model against a naive baseline, offline and in production, to see if your ML model outperforms simple logic.
Second, establish a model governance council or framework. Ensuring that the AI models created are responsible, robust, and unbiased is paramount.
Third, set measurement metrics. The importance of these metrics and guardrails cannot be overstated. Product and data teams should invest in developing metrics and guardrails to measure different aspects of success. One type of intervention can impact other systems. For example, it is common to measure recommendation systems with metrics of relevance, click-through rate, or conversion rate. However, it is also important to measure whether the system is showing a sufficient variety of recommendations. If static recommendations are displayed for a long period of time, users may lose interest in that section. Certain types of recommendations increase demand for certain products, increasing out-of-stock rates, etc.
Fourth, experimentation. A culture of experimentation is what distinguishes organizations that use AI from those that are powered by AI. Experimenting before implementation helps companies understand the impact of the model at hand and derive the most compelling insights. Therefore, investing in a good experimentation infrastructure is crucial.
Fifth, organizations need to invest in setting up infrastructure that facilitates model deployment. Problems often arise at the model deployment stage if the right infrastructure is not available to run the models in production.
Finally, it is important that organizations applying AI to use cases involving images and text invest in their tagging teams up front. Data labeling, or describing data types to the system, is one of the core steps to enable a model to properly understand and interpret the input data. Tagging teams can manually tag data across categories such as images, subjects, topics, etc.
summary
AI is more accessible today than it has ever been. For organizations considering adopting AI, it is important to understand the fundamentals of the technology before integrating AI. Beware of misconceptions about implementing AI and ensure you solve a business problem before integrating AI. The best use cases are the ones that most closely resemble solving a business problem.
AI is a great driver and, if the questions and metrics are set right, it can become a strategic advantage for your business. Start simple, experiment, question complexity, and measure, measure, measure again.
This article was written by a member of the AIM Leaders Council, an invitation-only forum for senior executives in the data science and analytics industry. To see if you're eligible for membership, fill out this form.