In Augmented Apps, we examine how product teams are exploring AI and ML to make their products more intuitive and enhance the user experience.
AI and machine learning (ML) are not just catchy buzzwords; they’re vital to the future of our planet and your business. Datasets have quickly grown too huge, complex, and fast-moving for humans to grapple with. Integrating AI and ML into your product or service is becoming basic table stakes for staying in the market. Doing it right can mean the difference between thriving in the new world of data and disappearing from it.
So what are the high-level steps to incorporate AI and machine learning into new and existing products? AI can add value to your product/service in many ways, including:
- Improved business performance
- Reduced costs
- Increased customer satisfaction
- Improved brand value
- Risk reduction (reduced human error, fraud reduction, spam reduction)
- Improved convenience and accessibility of products
In this article, we’ll dig into the ways AI can help you accomplish these goals, allowing you and your team to envision the future of your product or service.
Improving performance with AI
Improving customer experience and reducing cost in a single step sounds impossible, but this is exactly what correctly implemented AI can achieve. Usually, we have to spend more money to achieve a better customer experience, but AI simultaneously delivers greater accuracy and focus and reduces human capital cost.
One aspect of AI that makes this outcome possible is training using expert domain knowledge. An AI that is trained with expert knowledge in a specific field can perform much better than a chatbot having a conversation on random topics. Visit the Mitsuku chatbot, for example, and ask the simple question, “How does a toaster work?” The reply, “You could spend years researching it,” is not useful. By contrast, have a look at the discussion with the stain removal chatbot in the diagram below.
Notice how the chatbot in this case guides the discussion with questions on a specific topic. This is where expert knowledge AI products can add enormous value to a brand. This is also an important takeaway for teams seeking to implement AI successfully: Start with the key performance indicators (KPIs) you want to measure your AI app’s success with, and see where that dovetails with your expert domain knowledge. Then tailor your approach to leverage your unique data and expertise to excel in those KPI areas.
Example: Products with recommender components
Suppose you’re working with a recommender engine that suggests products to a site visitor. What is the process to improve recommender engines? An obvious mechanical answer is: use relevance as a metric. Look at implicit and explicit feedback. One proven method for success is to refine preferences in real time by inserting new customer choices into the neural net during transactions. This means that the intelligence of the recommender actually improves while a visitor is browsing, even before a transaction happens!
Another important method is to benchmark existing metrics. Know the limitations of your existing dataset and answer these questions: What categories of data are there? What is the specific problem or obstacle to creating accurate recommenders now?
Next, collect root causes to address these questions. If possible in your marketing scope, talk to existing customers and improve their experience by comparing their expected outcomes with current outcomes. Learn the customer’s expectations. Specifically, use an affinity diagram to prioritize and plan the minimum viable product (MVP).
Points to remember:
- Understand the customer problems you want to solve: What are the social and emotional aspects? Can a chatbot help improve relations? What are the business consequences?
- Inclusiveness: Seek to discover and eliminate bias around race, age, and gender. Be sure test cases represent the diversity of app users.
- Recognize user emotions: Emotion is crucial to recommenders. Natural language processing can be used to detect tone and intent in written conversations.
- Build a product that continuously learns: AI systems should adapt to and learn from feedback.
Better data, better AI, better products
Having the right data (and enough of it) to train your AI is vital. Take Grammarly as an example: This popular program checks the grammar, tone, and style of documents. It groups content into one of three styles: professional, moderate, or casual, based on natural language processing. Getting this AI properly trained required a huge learning dataset with countless documents that were tagged according to specific criteria. Accurately prepared data is the base of AI.
As an AI product manager, here are some important data-related questions you should ask yourself:
- What is the problem you’re trying to solve?
- What are the right KPIs and outputs for your product?
- What data transformations are needed from your data scientists to prepare the data?
- What will it take to build your MVP?
Today, we’re developing AI in an era where data is treated as code, or at least as an extension of code, because the code alone cannot achieve deep learning without the data. Moreover, training data must conform to the app design to such an extent that data preparation resembles coding. There are several important steps to take to ensure data integrity for your AI project:
- Correct: Common data errors include Fahrenheit/Celsius, UTC, and a 12/24-hour clock
- Compliant: Data must comply with legal consent and privacy rules
- Current: Data must be relevant in the time scope and not expired
- Consistent: You should validate consistent data types with test cases
- Consolidated: You should determine the output desired, for example, data for regression vs. classification should be distinguished by the ML model in use
The perfect fit
When building your data model, it’s vital to avoid both underfitting and overfitting. This is done by an ML method called validation. Although AI product managers may not be involved at the level of algorithm development, they can benefit from recognizing the symptoms of overfitting in the behavior of the model under development.
An underfitted model is inflexible and doesn’t learn well. An overfitted model learns from all the data, including noise in the data.
The underfitted model looks as if it ignores the data altogether, while the overfitted model at right models the noise in the data! Without getting too technical, we want to mention an example of validation that can be used to resolve these problems: k-fold cross-validation.
Understanding k-fold cross-validation
K-fold cross-validation works by dividing the dataset into multiple subsets called “folds.” This method has the advantage of improving the model without requiring additional datasets. The model with the lowest cross-validation score will model the training data best while achieving a balance between underfitting and overfitting. Here we can begin to see why it is beneficial for AI product managers to have a data science background. The ability to understand ML models at a high level improves our ability to work directly with our team of data scientists to create accurate models.
Sampling the mechanics of AI algorithms as we have just done leads us into the related topics of model selection and model evaluation. As a product manager, you’ll want to do your own research to learn about the best modeling methods for your AI product domain and scenario. Learning this kind of information will aid you in interacting meaningfully with your team of data scientists.
Building smarter products, boldly
The future will be intelligent. Data continues to proliferate, and only AI assistance can help humans make sense of it. These are huge, existential issues but have simple, material applications when it comes to building the next wave of apps, products, and services. Whatever you’re building, data needs to be a part of it. Adding an AI element will help your users get the most out of that data, improve their experience, and keep your company in the market as AI becomes the norm for every aspect of daily life. Envision a better world and a better product, then build it boldly with the right data and AI.
Inna Tokarev-Sela, Sisense’s Head of AI Research, has over 15 years’ experience in the tech industry. She spent the last decade at SAP, driving innovations in cloud architecture, in-memory products, and Machine Learning video analytics. A frequent speaker at industry events like IBC, NAB, WonderlandAI, and Media Festival, Inna holds a BS in Physics and Computer Science, an MBA, and an MS in Information Systems, having written her thesis on Neural Networks.