The role of the AI Product Manager isn’t a new one, but with the proliferation of AI products and services the role has taken on new significance. AI and Machine Learning have been around for decades, but the main shift that we’re in the midst of is a shift from retrieval to generative computing.
It’s important to make a distinction between the PM who specialises in AI (the AI Product Manager) and the PM using the AI to be more effective in their role. In this post, we’ll look at the ‘AI Product Manager’ and what it means to manage products based on AI technology. When I talk about AI Product Managers, I think of those PMs that work on any of the following AI enabled product categories:
Applied AI products — AI powers part of your product, think about specific features such as Notion AI and Legal Zoom’s Doc Assist.
AI platforms — such as Vertex AI, RapidMiner and SageMaker facilitate development and deployment of machine learning models.
AI services — There are a whole host of AI-based services that solve for a specific use case or problem. Services like Azure Databricks and Amazon Rekognition address specific use cases around data management & governance and computer vision respectively.
Traditional software development is based on rule-based instructions, programmed by developers. These rules are predeterministic which means that the software will always produce the same output based on coded rules. Whilst AI systems can be rule-based, these systems are designed to learn from data. AI is an umbrella term which covers distinct techniques such as Machine Learning (ML) and Natural Language Processing (NLP). ML systems, for example, use training data for the machine to learn, with a ‘model’ being the training output. Deep learning is a subset of ML, training the machine on complex, unstructured data like images and text).
Generative AI is a subset of deep learning, using ML models to generate new content based on existing data. Whether the product is based on machine or deep learning, it’s important to realise that AI products are different to traditional software products. These are the main areas where there’s a clear difference:
- Development stages — Traditional software typically starts with a functional spec, and involves design, coding, testing and release phases. The focus with AI based products is often on performance accuracy, and development involves many experiments and iterations to train and tune the AI system (e.g. algorithms, libraries and frameworks) that power the product.
- People involved in development — Traditional software products are developed by cross-functional teams consisting of a PM, a designer and software engineers. With AI based products, a PM also works closely with data scientists, ML engineers and data engineers.
- Data dependency — AI based products rely heavily on large amounts of data to train AI models. The data ultimately determines the product’s functionality and user experience.
- User Experience — Traditional software has predefined user interfaces and workflows. AI (applied) products like chatbots, tax advisory or slide generators offer user experiences that adjust based on a user’s behaviour, preferences and inputs. Customer satisfaction with AI products tends to be driven by the perceived accuracy and trustworthiness of the data.
- Interpretability — Traditional software products are generally more explicit and interpretable, meaning that users can easily understand cause and effect of a particular action. Compare that to AI products — particularly if they’re based on deep learning models — which can be more difficult to explain.
- Testing — If you’re testing a traditional software product, you’ll be using a predetermined testing plan and test against set inputs and outputs. When testing AI products, you’ll often be evaluating performance based on unseen data, edge cases and any bias in the data. Businesses will have their own evaluation metrics to test the robustness of an AI product or system.
- Higher risk — AI systems are inherently riskier due to non-deterministic behaviours. Think about algorithms using different paths to arrive at an outcome, which can pose a risk from a regulatory and public perception point of view.
- Adaptability — AI based products are usually designed to learn from new data and user interaction, creating new content or recommendations in the process. Compare that to traditional software products that can remain fairly static once deployed.
Whilst there are differences between traditional and AI based products, even as an ‘AI Product Manager’ the four cornerstone elements of product management aren’t changing:
- Why? — Why is the problem worth solving? Why do we prioritise solving one problem over another?
- Who? — Who has a problem worth solving? Who do we need to solve the problem for the customer?
- What? — What solution will solve the customer problem? What value will it deliver?
- How? — How will we build the solution? How will we take it to market?
Within these four key elements of product managements there are aspects that are particularly relevant to AI / ML products:
Why? — “What’s the customer problem to solve?” “Why is it worth solving?” “Do we need AI / ML to solve this problem?” These are the foundational questions that each PM should have an answer for, before considering AI as a means to solving the customer problem.
There are already several areas where AI is delivering tangible customer and business value, like content generation and personalised recommendations. Don’t, however, treat AI as a magic bullet or gloss over potential negative ramifications of using AI in your product. The necessity of starting with the problem and understanding the ‘why’ won’t change if you’re developing a product leveraging AI. Strong product managers will have a clear and compelling rationale of the ‘why’ and differentiated value of an AI based product.
Who? — Understanding who we’re building a product for won’t change. Whether AI is your product (e.g. virtual assistants) or is integrated in an existing UX, PMs will continue working with customers and stakeholders to solve the right problem in the right way. User centric design isn’t going to go away!
When I worked on chatbots at Intercom I learned that when developing them we still needed to think about human to human conversations, as well as the user intent and the mental models of the person interacting with the chatbot. AI first products like Copilot, Spotify AI DJ and Harvey are good examples of AI first products that are designed based on traditional human centred design principles.
Data determines the agenda for AI PMs when engaging with customers and stakeholders. The AI PMs that I know are highly data literate and possess varying levels of technical depth with respect to data management and data science.
Customers will want to know how their personal data is used and stakeholders will want to be on top of data compliance. AI is such a fast evolving technology, stakeholders will want to be kept abreast continuously about new developments and potential risks. The probabilistic nature of AI bears an inherent amount of risk and you can’t confidently assure your stakeholders that something “will never happen.” Instead, the AI PM needs to have detailed knowledge and close involvement in managing stakeholder comms around information and model risks.
What? — Exploring potential solutions to solve a problem won’t change. The AI PM will continue to work closely with customers, stakeholders and cross-functional stakeholders to define the right solution.
Learning from PMs currently working on AI-first products, it’s clear that they need to be proficient in data analysis, understanding the data lifecycle and have a good grasp of algorithms.
→ Data collection: If you want to create a feature that is data driven, you’ll need to understand what data types (numerical vs descriptive) you want and are allowed to use. You’ll need to figure out how and where you’ll get the data from. For instance, do you use real world data or synthetic data?
→ Data analysis: Once you’ve established data availability, you need to figure out how to clean the data, typically involving the removal of redundant data and making sure that the relevant data is stored logically and efficiently in the database. Rationalising the data involves handling missing values and dealing with outliers in the data.
From a UX perspective, data is the product and PMs will need to think about the data as UX:
- Is the data accurate?
- Is the data complete?
- Is the data consistent?
- Is the data up to date?
The answers to these questions culminate in the relevance of the responses generated by the AI. AI based products rely on generating user input and feedback to improve the accuracy of responses. Making sure that responses are accurate isn’t a one off exercise.
Most companies that use AI as part of their products or services have human annotation practices in place to constantly reinforce machine learning models. Humans will review and label data such as images, text, video or audio, making sure that the data is classified correctly.
How? — Every product manager needs to have a good understanding of the technology underpinning their product and be closely involved in the ‘how’, shipping the product. With the rise of AI based products and features, PMs will need to have a good grasp of data analysis, machine learning and algorithms.
I’m currently learning about the opportunities and constraints of the different approaches to AI conversation design, and I realise how PMs need to have a good understanding (and ask the right questions) about these more technical concepts:
Intent classification — Intent classification is a Natural Language Processing task that determines the underlying goal behind a text or spoken voice input. Each user intent represents a specific outcome or action that a user wants to accomplish. The algorithm analyses, tags and assigns content to the relevant category based on intent.
Disambiguation logic — Disambiguation logic covers the approaches used to determine the intended meaning of an ambiguous word or sentence based on its context. It applies logical reasoning and analysis to remove ambiguities in natural language.
Take as a simple example a user asking a chatbot to recommend a comfy chair:
User: Can you recommend a desk chair?
At this point the chatbot could go in a certain direction, based on assumed user intent. Instead it could ask the user to clarify their input based on context. The chatbot can provide a few options, with the user selecting the most relevant option.
Chatbot: I see you’re working from home. For your desk chair, would you prefer an ergonomic chair that adapts to your needs and movements, or would you prefer a stackable chair that you can easily store away to free up space?
Explainability — Explainability is about providing clear explanations to users about the AI’s decision-making process and being transparent about the AI’s data sources and algorithms used. In our desk chair example, the chatbot needs to be transparent about the data that was used to provide a recommendation so that the user can verify the validity of the recommendation.
Main learning point: The AI Product Manager is a specialised role, just like there are PMs that specialise in payments or healthcare but the key product management principles still apply. The heavy focus on data is what differentiates the AI PM role: collecting, analysing and deploying data at scale is critical to building any AI product. It will be interesting to see how the expectations of these specialised AI Product Managers will evolve.
A huge thank you to Marco Pfrang for their input into this blog. Thank you for reviewing my drafts and providing your invaluable feedback!
For further learning:
What is ‘AI Product Management’? by Polly Allen
The Rise of the AI Product Manager by Liat Ben-Zur
Strategies and Surprises:A Peek into the World of an AI Product Manager by Diogo Marta
