AI product managers - a critical but challenging role

"Just give me an AI engineer and an API endpoint - that's all I need to build a product."

I have heard this sentiment often since the release of ChatGPT. And it's true - you can now create impressive demos in hours that would have seemed impossible just a year ago. The reality is that building demos and building useful products are fundamentally different challenges. And if you want to turn your ideas into products that really work, we have found it invaluable to have an “AI product manager”.

What is a product manager?

A product manager owns both vision and execution - they're the person who ensures a product delivers real value to users while remaining technically and commercially viable. Think of them as the conductor of an orchestra: they don't play every instrument, but they ensure all parts work in harmony. They translate between user researchers, designers, and engineers while keeping everyone focused on the end goal. 

We did not invent this role. The value of product managers has long been understood in the private sector (e.g. Silicon Valley Product Group) and public sector (e.g. elsewhere in the Government Digital Service). At i.AI, we take a flexible approach to this role. While each product has a designated product manager, it's more about mindset and responsibilities than a job title. On smaller incubation projects, you'll often find our lead AI engineers wearing both technical and product hats. What matters is that someone is actively bridging the gap between user needs and technical possibilities. To be good at this role you need to understand enough about the technical details to be able to drive a deliverable strategy, not just something that sounds impressive.

What is an AI product manager? 

No prizes for working out that an AI product manager is a product manager for an AI product. Compared to traditional product managers, there are two main differences: 

1. A new and fast moving technology to get to grips with 
2. The extra uncertainty that comes with not knowing what is possible until you experiment and build

Even simple AI products require extensive experimentation to determine technical feasibility and optimal user interactions. Take an automated email triager: key decisions include whether to use a real-time team directory with LLM-based forwarding or train on historical data. The system must also balance automated decisions versus human review to maintain user trust while delivering value. These core product choices emerge through hands-on testing rather than upfront planning and show how coupled the AI and product decisions are.

It is extremely difficult to build and adjust the product vision without a good understanding of the AI capabilities driving it, especially in the early stages of a product build. Similarly, it is incredibly difficult to pull everything together to create the product without a product mindset. 

This is what makes the role so difficult but so important at the same time. In this post I will illustrate the importance of this role with a real example from the incubator. 

Consult 

Let me illustrate this with Consult, one of our products. The challenge is significant: Government spends millions of pounds and countless hours running public consultations. These consultations generate thousands of responses that need careful analysis, often slowing down critical policy decisions.

Our initial approach seemed logical. We knew data scientists in the No10 team had successfully used BERTopic (a natural language processing technique) to cluster similar consultation responses and label these clusters. Why not turn this into a product? 

While we had run some major recruitment rounds for AI engineers, not many had started yet, and so this product was assigned to a traditional software team with a traditional product manager with the assumption that the AI part (BERTopic) had already been validated. 

Because of this, the team focused on the familiar product elements - UI, data upload, analysis pipeline, results visualisation, but skipped the crucial experiments that would have surfaced fundamental questions about the AI capability:

• Could the auto-generated theme labels actually help analysts make sense of responses?
• Does the process produce useful themes?
• How would we handle opposing viewpoints being clustered together? ("plastic bottles cause pollution" vs "plastic bottles don't cause pollution")
• What about lengthy responses that covered multiple themes?
• Could analysts customise themes to match their domain expertise?
• How would the system adapt to technical language specific to each consultation?

When the real outputs were tested with real users, the results were sobering. The theme labels were inconsistent and hard to trust, making the product difficult to use in practice. The original assumption, that the AI part had been validated, was wrong. It was wrong because it is very different for a data scientist to use AI as part of their analysis, and be able to interpret, iterate and react to the results, vs it being built into a product. 

Given the circumstances, the team did well to build a strong foundation of data (they collected real consultation data from 8 departments) and a comprehensive amount of user research. This meant that when an AI product manager did join the team, they were able to move quickly and they brought about the following changes. 

With an AI product manager, the team:

1. Went back to basics and tested new AI capabilities from scratch
2. Embedded ourselves with consultation teams to truly understand their process
3. Developed a new vision built around LLM capabilities
4. Established rigorous testing protocols

Now, with concrete evaluation results proving our approach, are we building the surrounding product features. The key lesson? Having someone who understands both AI capabilities and product development principles helps you ask the right questions before investing in the wrong solutions.

You can read more here about our development journey with Consult.

Analogy - clearing the fog 

This isn’t just an isolated example - it represents a fundamental challenge in AI product development. I will finish with an analogy to drive the point home. Imagine a team lost in dense fog.

This team has all the right skills: builders ready to construct, architects eager to design, and explorers keen to scout ahead. But in the fog, each group pulls in different directions. The explorers want to investigate every path. The builders are anxious to start working with whatever materials they find. The architects insist on perfect plans before taking action.

This is exactly how early-stage AI products feel. The fog represents our uncertainty: about what's technically possible, what users truly need, and how to bridge that gap. Different team members, with their valuable but distinct perspectives, naturally pull in different directions.

This is where an AI product manager shines. Like an expert navigator in the fog, they:

• Keep the team moving together, balancing exploration with progress
• Update the map as the fog clears, adjusting the vision as they learn
• Help the whole team advance in parallel - because waiting for perfect visibility means never moving forward
• Focus on finding a path through (solving the problem) rather than committing to a single solution (like building a boat)

With AI products, the fog is particularly thick. Your tools and materials - the AI capabilities and data - might produce something extraordinary or lead to dead ends. You need someone who can guide the team through this uncertainty, someone who understands both the technical landscape and the destination you're trying to reach.

If you find your team caught in the fog of AI development, get yourself an AI product manager. Or better yet, become one yourself. 

Notes from the author

If you want to learn more about what it takes to be a great product manager, then I can’t recommend the book “Inspired” enough. 

If you are already a product manager and want to get more AI experience, the best advice I have is to try and use AI as much as possible and to pair up with an AI engineer or data scientist to understand what is possible and how you go about evaluating the performance of AI.