Essential Features in AI Products

In this article, we will dive into AI product characteristics, the most important features, and how they are related to various data product types.
Compared to other data products such as analytical, AI products have differences such as how the product is getting built, how it is used, its non-deterministic behavior, and the need for continuous monitoring considering data and model drift. In addition, some data products have included Generative AI features, which makes them different from traditional ML and AI products. For now, let’s not make this differentiation and focus on the common characteristic among AI products:

• User centricity: This is a common step for any product design to identify the “persona” of the product. Persona presents the target audience with specific needs, behaviors, and challenges. During the product design, we should always keep the end-user in mind, what does he/she expect from the product, is it rapid access to the information? Personalization? Fast response?

  Considering other types of products, in an AI-based product, this feature is getting more important, since these types of products don’t have a linear behavior because they are based on probabilistic models that need to provide transparency or explainability to the end user. Imagine in the case of medical image analysis, the results should be trustable and interpretable for the final users else it will bring life-threatening risks to the patients.
  The main objective varies by use cases, however, the product should fulfill at least one main expectation of the user: why should he/she use our product and what problem are we solving?

• Explainability: To ensure that the final output is easily understandable to end users, it's crucial to present evidence demonstrating our advanced model's effectiveness. This is particularly important when trying to convince business users who might view a predictive model as a complex and unfamiliar "black box." By providing clear proof, we aim to build their trust in the model's capabilities. For instance, in the case of a chatbot that allows users to interact with their data, users may not require a detailed explanation of how the underlying model functions. Instead, what they need is the reassurance of the system's reliability and trustworthiness, which we will elaborate on in the next section.
• Trustworthiness & Reliability:
  For example, consider a chatbot powered by large language models (LLMs) that retrieves information from your data. To trust the chatbot, users would want access to the documents it referenced to verify the information for themselves. If there's any uncertainty about the chatbot's output, it's important to clarify that the response is AI-generated and may require human verification for accuracy.

  Additionally, the performance of AI models can vary, leading to instances where they might not predict data patterns accurately. This variability can be due to changes in the data sources (known as data drift) or changes in the relationships between variables, including external factors that affect the output (referred to as model drift). In such cases, it's necessary to continuously monitor and update the models to ensure they maintain stable performance.

• Data privacy: Data privacy involves the methods, rules, and technology used to safeguard personal and sensitive information that artificial intelligence systems collect, process, store, and share. Considering the vast amounts of data these systems deal with, including personal details, preferences, and behavior patterns, ensuring privacy is both vital and complex.

  Ensuring the security of user data and compliance with regulations like GDPR is crucial for all products, not just those based on artificial intelligence.
  In the case of a product based on a Large language model (LLM), it's particularly important to prevent the transfer of sensitive information to third-party LLMs over which you have no control. This is because such data could potentially be used for model retraining, leading to unintentional disclosure.

• Adaptability: considering the nature of AI-based products which are based on the Data, the product should be adapted to the changes in the environments and new data which can keep up its performance and accuracy over time.

  This requirement for adaptability is not just about handling new types of data or larger volumes but also about responding to shifts in user behavior and market trends. Moreover, adaptability also means the product's ability to detect and correct for biases or errors that may arise over time, ensuring fairness and reliability in its outputs.

• Fairness: This feature addresses the need for various types of users to have a universal and inclusive experience. AI products can inherit biases from the data they are trained on, leading to unfair or discriminatory outcomes against certain groups. There are multiple examples in systems with facial recognition, that some ethnicities could not be identified and recognized, and it is bringing biases and increasing discrimination.
  As AI is advancing fast and is integrating more into our daily lives, it is crucial to consider this aspect of AI and ensure that AI systems are transparent and fair, especially in critical areas like healthcare and justice.

Now, we have to see how these characteristics can be related to types of AI products, before that, let’s dive into various AI product types:

• Descriptive product: a product based on a descriptive model that works by analyzing and interpreting patterns and insights from past data. For example, it identifies the most influential features in the data that significantly impact predictions.These types of products can be beneficial for Data or Business analysts, which helps them to understand the behavior of past data to do decision-making.
• Predictive product: a product based on a predictive model, that can predict a pattern or trend, for example, the next price of the stock market. This type of product can be used by marketing professionals, financial institutions, retailers, or healthcare providers.
• Prescriptive product: a product that can suggest an action to optimize a specific output. The results can be used to improve the decision-making process. For example, an AI advisor can predict the best strategy for investment to minimize the risk or maximize the profit. This product can be beneficial for operational managers, business strategists, financial analysts, and healthcare providers.
• Generative (Gen) AI-based product: refers to the products targeting generating text/image/videos/audio (digital content) which relies on generative AI technology. The most known product from this category is chatGPT, which is a virtual assistant that can interact with users through text, image, or voice. The users of these products are normally content creators, designers, artists, educators, and programmers.

Let’s look at how the mentioned AI product feature can fit into these types of products, this differentiation is important because each product addresses different aspects, which can be more significant in one, and less important in another one.

Consider all of the characteristics are relatively important for all of these products, and maybe their weights are changing just based on the final goals and functionality.

Descriptive AI Products

In descriptive AI, the accuracy of data insights and interpretation is important, so trustworthiness and data privacy are the most important features to ensure conclusions are accurate.

• Trustworthiness & Reliability: the outputs must accurately reflect the data insights.
• Data Privacy: it is essential when handling and analyzing potentially sensitive information.
• User centricity: Insights and visualizations should be relevant and actionable for the user and enable him/her to make informed decisions.
• Explainability: Users need to understand how conclusions were drawn from the data.
• Fairness: Ensures the analysis does not misrepresent the data and is fair.
• Adaptability: Since the focus is mostly on historical data, adaptability is not an essential feature.

Predictive AI Products
Predictive AI relies on data accuracy and model reliability, to empower final users.

• Trustworthiness & Reliability: it is a critical aspect because predictions must be accurate and reliable for decision-making purposes.
• Data Privacy: Crucial because the data used for making predictions is often sensitive. We have to be careful also to mask personal data to not violate users' privacy.
• User centricity: Ensures the predictions serve the user’s specific needs and purpose.
• Explainability: It is important for users to understand how predictions are made to trust and act upon them.
• Adaptability: Predictive models must adapt to new data to stay accurate over time.
• Fairness: Ensures predictions do not create biases or inequalities.

Prescriptive AI Products
For prescriptive AI, the emphasis on user-centricity and explainability reflects the importance of actionable and understandable recommendations for users.

• User centricity: Recommendations must align with user needs and contexts.
• Trustworthiness & Reliability: Prescriptions should be based on accurate, reliable predictions and analysis..
• Explainability: Users need to understand the rationale behind recommendations to follow them.
• Fairness: Ensures the model does not recommend actions that could be biased.
• Data Privacy: the application of the prescriptive model is on the recommendation of an action based on the analyzed data, rather than focusing on the sensitivity of the data itself.
  Adaptability: The system should evolve with changing data and conditions to provide relevant patterns.

Generative AI Products
In Generative AI user-centricity is fundamental to meeting objectives, with a strong emphasis on privacy and the ability to adapt to user feedback for improving outputs.

• User-Centricity: The generated content must meet the user's creative needs and preferences.
• Data Privacy: Especially important as generative AI might use datasets that could include personal information.
• Adaptability: Needs to continually learn from feedback to improve content generation.
• Trustworthiness & Reliability: Generated content should be accurate and appropriate for the context.
• Explainability: Users still benefit from understanding how inputs influence generated outputs, but it is less crucial compared to predictive AI
• Fairness: Ensures the AI does not create content that brings biases.

In this article, we had an overview what are the most important characteristics of AI products, which are User centricity, Explainability, Trustworthiness & Reliability, Data Privacy, Adaptability, and Fairness, and why they are important. Then consider various AI product types which are
Descriptive, Predictive, Prescriptive, and Generative (Gen) AI-based, and how these characteristics relate to AI product types, and what is the order of their importance.
This exploration emphasizes that the application context can influence the prioritization and relevance of these categories and characteristics.