On one of the hottest evenings on record in Dublin, Cubic3 hosted ProductTank Dublin for an evening centred on a question every product team is working through right now: how do you build AI products that actually make a difference – not just demos that impress, but tools that stick? Attendees from the world of product marketing came to share tips and views and to debate how to build AI products that matter with some great insights below.
Stop building before you’ve asked the right questions
Cathal O Riain, Head of AI Product at Cubic3, opened the evening. He’s spent 25 years working in data and AI across Accenture, Aon, Flutter Entertainment and Cubic3, and he came with a clear argument: most AI initiatives don’t fail because of the model. They fail because nobody asked the right questions before building it.
Cathal’s framework was built around the confusion matrix – a tool most data scientists use to measure model accuracy – but he applied it in a way most of the room hadn’t seen before. The model itself, he said, is only the middle of the problem. What matters as a product manager is what feeds into it and what happens with its outputs.
He walked through three examples from his career: credit card fraud detection, insurance pricing elasticity and gambling harm prevention. Each one showed that false positives and false negatives carry very different real-world consequences depending on context. A fraud detection system that declines a card when a customer is trying to impress investors at dinner is a different kind of failure from one that lets a fraudulent transaction through. The model score might look the same. The customer impact couldn’t be more different. The lesson: design the process around the consequences of the errors, not just the accuracy of the model.
On data quality, Cathal was direct. Product managers own this. Not just the data team. You don’t need to know which table lives in which database, but you do need to understand where your data comes from, what might be missing, and how it could break. Unstructured data is almost never as bad as teams assume – it’s usually just ungoverned. Sort the governance and you have something to work with.
He finished with two practical tests for any AI product initiative. Do you understand your data – its source, its gaps, its limitations? And do you understand your process – what happens when the model is wrong, how errors are handled, how the whole system holds up when something upstream goes down? If you can’t answer both, you’re not ready to build.
The panel: from model to market
The second half of the evening brought together two more voices from inside Cubic3 for a panel on the go-to-market and product discovery side of AI.
Vlad Shevchenko, AI Product Manager at Cubic3, has worked in AI product management since 2019 – well before ChatGPT made it a mainstream conversation. His diagnosis of what changed after 2023: billions of people suddenly understood what AI could do, which was good. But it also generated a wave of products built to solve problems that don’t exist, by people who hadn’t thought about distribution. A lot of wasted code, in his words.
Vlad’s advice for managing changing requirements was characteristically practical: front-load discovery, run fast prototypes before committing engineering time, and – his favourite line of the evening – if you’re working in enterprise, put a clause in the contract that charges for change requests. If the customer still wants the change, it’s worth building.
His point on anomaly detection at Cubic3 landed well with the room. Training one model across all vehicle manufacturers sounds efficient. But a car manufacturer and an agricultural vehicle manufacturer have completely different normal traffic patterns. For an agri fleet, traffic spikes every Monday morning when the combines go back into the fields. A model that doesn’t know that will flag it as an anomaly. Customers will stop trusting the dashboard. Context matters as much as the model.
Visishta Karasani (Vish), Senior Product Marketing Manager at Cubic3, brought a different angle: how do you take an AI product to market without either underselling what it does or making promises the product can’t keep? Her answer was to stop leading with AI at all. B2B buyers – procurement heads, finance directors, fleet managers – don’t buy AI. They buy outcomes. If you can tell them that your product improves a specific metric by a specific amount, and a peer in their industry can validate that claim, the conversation changes.
Vish also made the case for product demos as a tool for credibility, not just sales. In long enterprise cycles, showing that a feature works is worth more than any amount of messaging. The room responded to that one – it’s a tension a lot of product teams feel.
What the room took away
The Q&A covered a lot of ground: how to fine-tune models when false positives are unacceptable, how to balance data privacy with the need for personalised diagnostics, what makes an AI feature ready to become a product. A few threads ran through all of it.
First: the pressure to ship AI quickly is real, but it’s usually the wrong pressure. Speed of experimentation is valuable. Speed to production, without the foundations, is expensive.
Second: AI as a feature and AI as an outcome are different things. A spell checker that uses AI is a feature. A tool that reads six months of customer conversations and tells a team what to do next is an outcome. Only one of those is worth building a go-to-market strategy around.
Third: the people in the room are not short of ideas or ambition. What they’re navigating is the discipline to slow down, ask the hard questions early, and build on solid ground.
Thanks to everyone who made it to The Hive, despite the heat, and to ProductTank Dublin for another excellent evening. We’re already looking forward to the next one.
See how Cubic3 turns vehicle data into decisions with Explore3
