Why your product needs an AI product manager, and why it should be you — James Lowe, i.AI

Key Concepts

AI Product Management, AI Evaluation, Feature Experimentation, Pivoting, AI Uncertainty, Human-in-the-Loop, Data Proficiency, Probabilistic Nature of AI, Model Context Protocol.

AI Product Management: A Necessary Role

James Low, Head of AI Engineering at the Incubator for AI (UK government), argues for the critical role of the AI product manager. He builds on Andrew Ing's post highlighting the increasing accessibility of software prototyping due to AI coding agents and AI features, leading to a higher demand for individuals who can strategically decide what to build. The core argument is that AI expertise is essential for product managers in the current landscape.

Product Management Fundamentals

Product management sits at the intersection of three key areas:

• Business: Viability and profitability of the product.
• Technology: Feasibility, including the availability of necessary skills.
• Users: Desirability, focusing on solving user problems.

AI introduces complexities to each of these areas:

• Business: Requires acceptance of higher experimentation rates and potential failures.
• Technology: Demands robust evaluation and monitoring of AI performance.
• Users: Necessitates careful consideration of the probabilistic nature of AI, including guardrails and human-in-the-loop mechanisms.

The central question for AI product managers is determining the feasibility of the AI-driven solution. While existing product management skills remain relevant, data and AI proficiency become paramount. This necessitates upskilling for product managers and highlights the potential for AI engineers to transition into product management roles. The "AI product manager" is more of a mindset than a specific role, emphasizing the need for someone to grapple with the intersection of business, technology, users, and AI feasibility. Brett Taylor's quote from the Latent Space podcast emphasizes the power of combining product and engineering expertise in a few individuals, as great things are rarely created by committee.

Lesson 1: Evaluate AI Early (Consult Project)

The Incubator for AI's "Consult" project illustrates the importance of early AI evaluation. The project aimed to analyze free text responses from public consultations, a process that traditionally takes months and costs millions. Initially, the team rushed into product building using existing NLP techniques like BERT topic modeling. However, user testing revealed inaccuracies and inconsistencies that failed to meet legal thresholds.

The team pivoted to prioritize AI capability first. They gathered data from real users and generated synthetic data to create an evaluation dataset. This dataset was used to optimize their AI model. The output was then tested with real users. This process led to the development of "Themefinder," an open-source package. The key finding was that prioritizing AI capability revealed crucial points in the pipeline where human input was most valuable, leading to a different product than initially envisioned. Themefinder achieved results comparable to humans but was 1000 times faster and 400 times cheaper.

Lesson: Resolve AI uncertainties early on through evaluations and tests with real users. The team published their evaluations, with one even being featured on the BBC front page.

Lesson 2: Go Wide with Features (Minute Project)

The "Minute" project, an AI transcription tool, demonstrates the value of broad feature experimentation. The government has many potential use cases for secure AI transcription and summarization, particularly in reducing administrative burdens on frontline staff. Existing off-the-shelf solutions like AWS and Azure transcription services already existed. The challenge was creating a streamlined user experience.

The team initially explored numerous AI-powered features, leveraging AI coding assistants for rapid development. They then tested these features with different user groups to identify what resonated. After this experimentation phase, they stripped back the features and focused on what worked. The use of AI coding assistants made it easier to remove features without sentimental attachment.

An example is shown of the tool with many features: template selection, agenda input, AI edit button, and AI chat. Users found this overwhelming. Testing with different groups revealed the value of focusing on the probation services use case. The app was streamlined into "Justice Transcribe," built in collaboration with Justice AI (Ministry of Justice). The simplified version removed unnecessary options and merged the AI edit and AI chat features. Early feedback has been positive, and the tool is undergoing evaluation against other solutions.

Lesson: Experiment hard and go wide with features, embracing the uncertainty of what makes a good AI feature, but then cut back and streamline based on user feedback.

Lesson 3: Be Ready to Pivot (Redbox Project)

The "Redbox" project illustrates the need for agility and pivoting in the rapidly evolving AI landscape. The initial concept, born from a hackathon, was to digitize the ministerial red box (containing paperwork and submissions). However, user testing revealed that the most desired feature was secure chat with a large language model, as enterprise solutions were scarce at the time.

The second iteration of Redbox focused on providing easy and cheap secure chat for civil servants. This also presented an opportunity to integrate other Incubator for AI tools, such as Parlex (parliamentary and legislative data), into the chat interface. Within weeks of launching this version within the cabinet office, it gained thousands of users.

However, two key events forced another pivot:

1. Microsoft announced that Copilot Chat (their enterprise version of ChatGPT) would be free for enterprise Microsoft users, a large segment of the government.
2. Claude's model context protocol emerged, providing a standardized way to bring tools and data to models.

It no longer made sense for Redbox to be the primary secure chat interface or the sole access point for the Incubator's tools. The team shifted to investing in the model context protocol to integrate their tools and data into any client, including Redbox, Copilot Chat, and other enterprise AI tools.

Lesson: You'll have to pivot harder and faster than ever before. The AI landscape is constantly changing, requiring continuous adaptation.

Conclusion

While many product management principles remain relevant, AI introduces unique challenges and opportunities. The three lessons emphasize the need for early AI evaluation, broad feature experimentation, and constant readiness to pivot. AI expertise is crucial for navigating this evolving landscape and building successful AI products. The speaker encourages the audience to adopt the AI product manager mindset and highlights that the Incubator for AI is currently hiring.