AI Product Management 101

Artificial intelligence has evolved from a minor gimmick to an everyday necessity. Now, almost every major organization has joined the AI race, leveraging its power for business gains and expansion.

However, developing and maintaining an AI product has distinct challenges.

Thinking of jumping into an AI product manager role?

Below, we cover the basics of end-to-end AI product management and how to avoid pitfalls while building a reliable application.

Market Landscape

A well-managed AI product can help develop a large customer base and generate substantial revenue.

ChatGPT

Developed by OpenAI, ChatGPT was introduced in December 2021 and broke the record for the fastest-growing application.

This was the first major commercial AI application that breached public consciousness.

It has since become a household name used for professional and personal tasks.

The application reached 200 million weekly users in 2024.

MS Copilot

MS Copilot is a large language model (LLM) based chatbot integrated with Microsoft 365 and offered as an add-on.

It enhances the user experience with MS 365 applications and boosts productivity by summarizing notes, emails, and meetings.

Perplexity AI

Perplexity is an AI-powered search engine valued at billions of dollars.

It searches the Internet like Google and uses AI to generate a natural language response based on authentic sources.

In January 2024, it was reported to have over 10 million monthly users.

It has become popular for everyday use as it saves users from jumping from website to website for a single search query.

Adobe Firefly

Firefly is Adobe’s participation in the AI race.

It is a group of generative AI (GenAI) tools offered by Adobe in their creative suite.

It allows users to edit creatively using generative capabilities, such as adding new graphics or removing objects from existing designs.

It also provides AI enhancements, such as image upscaling and rendering entire designs from scratch. The AI capabilities provide a massive productivity boost and save several hours of manual editing and designing.

Kore.AI

Kore helps build conversational solutions like Chatbots or RAG-based search for industries like healthcare and finance.

To date, they have gathered a large user base, and the company automates 450 million daily interactions for about 200 million consumers and two million enterprise users worldwide.

AI Product Maintenance

The success stories about AI products highlight one critical lesson: create a reliable product that earns and retains consumer trust. 

Most organizations use AI to directly influence business growth and improve customer experiences.

However, these applications need to be updated often to maintain performance.

For example, a sales forecasting model trained two years ago might not recognize current sales patterns. It might generate misleading results that lead to false KPIs and ineffective goals.

AI product maintenance keeps models fair, accurate, and reliable, even as conditions change.

Businesses can safeguard their products, minimize additional development costs, and build for scalability.

AI Reliability

Some of the biggest concerns with AI are its probabilistic nature and lack of transparency regarding internal decision-making.

Results depend solely on the model's learning and the data it observed during training.

AI models are prone to biases and hallucinations caused by a lack of high-quality, noise-free data.

Such errors have drastic consequences in production and can lead to loss of business and customer trust.

Maintaining an AI product also means meeting data compliance and regulatory standards.

Consequences of a Bad AI Product

We have witnessed several past cases related to AI products/ test run failures.

1. In October 2023, New York City launched an AI chatbot to assist citizens with information on starting and operating a business in the city. The bot was soon found to be spreading misinformation and telling citizens to break the law.
2. Air Canada’s virtual assistant once told a customer that he would have up to 90 days to claim a bereavement discount if he purchased a normal-price ticket. When the customer went to process his claim, he was turned down and told that his information was false. The case was concluded in a tribunal, with Air Canada having to pay a total of CA$812.02, including CA$650.88 in damages.
3. A problem with Gemini’s Image Generator put Google in hot waters. The AI was criticized for generating prominent historical figures like the US Founding Fathers and Nazi-era German soldiers as people of color.

While none of these incidents had severe consequences, each did impact the organizations' reputations and damage users' trust.

As we move into the future, the magnitude and consequences of these errors will grow.

Organizations must implement best practices and strategies for AI maintainability early on.

Best Practices for Reliable Development

Here are some best practices organizations can follow for reliable AI development.

Use Diverse Datasets

A reliable AI product must handle diverse scenarios and provide accurate responses.

Datasets must be carefully curated and cover multiple scenarios and use cases. The data must also represent different communities and ethnic groups equally and be free from biases that may introduce unfair decision-making.

Maintain Data Quality

Organizations must develop robust data pipelines for consistent, high-quality data.

The pipelines will clean the data and remove any noise or outliers.

They will also mask critical information.

Implement Regulatory Protocols

Regulatory organizations, such as HIPAA and GDPR, impose strict data collection, storage, and use rules.

Implement a HITL Framework

While AI is intended to automate modern systems, a human-in-the-loop framework is still necessary.

Human supervision keeps AI systems performing as expected. Humans can judge how to deal with complex scenarios or unexpected outputs.

Human feedback also improves the model’s performance. The HITL framework can often be implemented alongside an MLOps pipeline to initiate retraining based on the supervisor's feedback. 

Use Hallucination Detection Frameworks

Hallucinations are one of the key challenges of building a GenAI product.

These occur when the model lacks knowledge regarding a topic but forcefully outputs a response. The hallucinated response consists of made-up facts that can misguide the user.

CRISP-DM: Streamlined Development

Cross-Industry Standard Process for Data Mining (CRISP-DM) is an Industry-Agnostic framework for developing data-related projects. It was first introduced in 1999 as a data mining methodology but is presently used to streamline the development of all data-related projects, such as AI product development.

It provides a step-by-step guide to plan and execute a data science project. 

Problem Definition and Evaluation

The first step to reliable AI development is understanding the domain and the problem.

Every problem requires a unique solution, and an initial assessment helps formulate an execution plan.

It also helps understand the available data sources and any required data and highlights criticalities.

Data Collection and Processing

All data sources listed in the first step must be acquired.

The data should be collated and studied via extensive exploratory data analysis (EDA) techniques. This helps understand the data, highlight missing or anomalous elements, and list down processing methodologies to be applied.

The EDA must also analyze if the data captures diverse scenarios and check for any biases hindering the model’s learning.

Once the EDA is complete, the next step is to build the processing pipelines. These pipelines collect raw data from all sources, process it as required, and deliver a clean, high-quality dataset.

Feature Engineering and Selection

When data pipelines are ready, the next step is to engineer relevant features.

This uses basic mathematical or statistical techniques to extract features from the provided attributes.

These features help the model understand complex, diverse scenarios and improve performance.

Developers may also select a feature subset that reduces computational load without significantly impacting performance.

Model Training and Evaluation

Next, model training can begin.

The training involves experimenting with various models, hyperparameter tuning, and cross-validation to find the optimal settings.

This is usually developed in an automated setup where all the configurations are deployed and evaluated against a test data set using metrics like accuracy, precision, or recall, depending on the type of problem.

The model must be evaluated against an extensive test set, covering multiple scenarios and edge cases.

Model Deployment

Once the model is finalized, it is deployed to production.

The deployment places the model in a practical setting where it is used for real-time predictions. The model is evaluated against real-world data, and the results are used to improve itself.

MLOps

While CRISP-DM streamlines the development process, an MLOps implementation keeps the model performance reliable.

A deployed AI product is prone to problems like data and concept drift and requires constant maintenance and updating.

An MLOps framework continuously monitors the model's performance in production and raises alerts as soon as it starts to degrade. In some implementations, it may also automatically trigger retraining using the latest collected data.

Benefits of MLOps

• Data and Model Versioning: MLOps frameworks retrain models but do not discard the earlier versions. They maintain multiple versions of data assets and trained models for experimentation and easy rollback.
• Drift Monitoring: It monitors the input data stream and model output for data and concept drift.
• Continuous Feedback: The monitoring system notifies developers when the model performance deteriorates.
• Improved Products: The feedback and monitoring loop allows the model to withstand dynamic inputs and maintain long-term performance.
• Improved Customer Trust: MLOps maintains model performance in production, ensuring consistent and accurate results. Accurate results lead to a happy and satisfied customer.