AI in Practice: Insights for the Road Ahead
There’s one topic we simply can’t ignore right now, and that’s the role of artificial intelligence (AI) in our work. AI has the power to transform not just the products we build, but the way we go about building them.
In true Product at Heart fashion, this was a topic we wanted to approach with curiosity and practicality (and minimal hype). That’s why we invited speakers who could share their lived experiences with AI—what have they learned so far and how might this impact your own AI journey?
In case you missed it, this year’s Product at Heart lineup featured four themed sessions:
AI in Practice: Insights for the Road Ahead
The AI in Practice themed session was designed to provide practical examples of how real product people are using AI to transform existing products or build new AI-first products.
This themed session featured three 20-minute talks. We heard from:
Zamina Ahmad, CEO, shades&contrast
Dominik Faber, Co-founder and Chief of Product, Paul’s Job
Jonathan Evens, Product Lead at Google DeepMind
In this post, we’ll share some highlights from each talk. If you’d like to explore any of the content in more detail, make sure you check out the recordings from each session.
Dominik’s talk: Lessons from Agentic AI-First Product Development
Jonathan’s talk: AI Isn’t Magic—It’s About Building Transformative Features
Zamina Ahmad: The AI Experimentation Era Is Over
More than two-thirds of German companies are already using AI tools—and this is just the beginning.
“We need a reality check,” AI Strategy Coach Zamina Ahmad told us at the beginning of her talk. Here’s the truth: Using tools like Grammarly and ChatGPT doesn’t make you an AI company.
Yes, the majority of companies are using AI tools, but they’re focused on digitizing existing inefficiencies rather than reimagining workflows. And this approach doesn’t give you a competitive advantage.
Remember: Simply using AI tools does NOT make you an AI-first company.
In fact, there’s an exponential gap between the companies that are rebuilding entire customer journeys and the ones that are just optimizing emails or other admin processes. Remember: The latter are simply individual productivity gains and not organizational transformation.
If you’re interested in tapping into those game-changing aspects of AI (rather than making incremental improvements to your productivity), Zamina outlined a three-stage AI transformation process:
Step 1: Trying to replace humans with AI (this represents tool thinking)
Step 2: Discovering AI’s contextual limitations (realizing that it can’t completely replace humans)
Step 3: Designing hybrid workflows that leverage both AI efficiency and human judgment
AI transformations typically follow these three phases. Your goal is to move beyond the first and second stages so you can begin to really tap into AI’s potential.
Watch Zamina’s full talk to learn how the real-life example of Klarna illustrates the difference between toolplay and transformation and what this means for you. And when Zamina challenges you to decide whether you’d like to be known as an efficient company or an innovative one, you should have a much clearer vision of how to answer.
Dominik Faber: Lessons from Agentic AI-First Product Development
In a quick poll of the audience, Dominik Faber, founder of Paul’s Job, revealed an embarrassing truth: The majority of us are clueless about the future of AI. But don’t feel too bad—even Dominik himself said he often feels this way, despite spending the past 18 months building AI products.
In his themed session, Dominik focused on the role of agentic AI in product development by walking us through his journey with his company, Paul’s Job. The idea behind Paul’s Job is automating high-volume recruiting with many skilled agents.
A quick introduction to Paul’s Job and all the AI agents it employs.
For example, the slide above shows an overview of the different agents Paul’s Job employs, including data collection agents, scheduling agents, job applications agents, etc.
But let’s take a step back for a moment. How did Dominik know that this was a problem that could be solved with AI agents? And how did he know which tasks to assign to different agents?
Dominik shared the CAIR metric, which he originally encountered on the LangChain blog. CAIR stands for Confidence in AI Results, and they define each of the components of the equation in the following way:
Value: The benefit users get when AI succeeds
Risk: The consequence if the AI makes an error
Correction: The effort required to fix AI mistakes
The CAIR (Confidence in AI Results) metric, introduced by LangChain, can help you determine whether something is a good use case for AI or AI agents.
Using this calculation, Dominik felt confident that recruiting automation was a perfect use case for agents because the value was high but the risk and correction would be low.
Dominik offered the following advice for determining your own AI or AI agent use cases: “The value should be high, the risk should be low, and correction should be easy when things go wrong—and they will go wrong!”
Tune in to Dominik’s full talk for a closer look at how to overcome technical challenges, align cross-functional teams, and craft AI-first strategies that drive innovation.
Jonathan Evens: AI Isn’t Magic—It’s About Building Transformative Features
What’s the difference between a gimmick feature and one that’s truly transformative? Jonathan Evens, Product Lead at Google DeepMind, shared a simple way to tell. Just answer the following questions:
Is it useful?
Does it feel out of place?
Is it trustworthy?
These questions are especially meaningful right now, because many product people are being asked (or told, to be more precise!) to build AI features. And when the need to build AI features is coming from executives rather than because of an actual customer need, you run the risk of building things nobody wants—those dreaded gimmick features.
Luckily, Jonathan reminds us, there’s a pattern that was used in the past to develop machine learning products. The slide below illustrates the deep learning development cycle.
The Deep Learning Development Cycle that was used for machine learning products in the past can be a useful guideline for developing AI products today.
It’s important to bring this process and rigor to any product, even an AI one.
Jonathan shared this slide from his colleague Timo Wagenblatt, which warns of the risk of falling into the “mobile first” trap. Too many people are treating it as simply “adding an AI feature here” or creating an AI strategy document here.”
We can’t just think of AI as a new feature—we need to think of it as a new operating system. If we simply layer AI onto old paradigms, it’s like trying to put a jet engine on a horse-drawn carriage and hoping that it will make it gallop faster.
Remember: The true AI opportunity is about transformation.
Working on a product like Google search also provides a useful perspective since Google search went through transformation from desktop to mobile and can apply those learnings to the transformation from mobile to AI.
Be sure to watch Jonathan’s full talk to learn more about Google search’s approach to launching transformative AI features and get his transferable learnings you can bring to your own product and organization.
Want to dive deeper into any of the topics from Product at Heart? Make sure you check out the other posts on our blog and dig into the video archive!