I Moved from Claims Manager to AI Product Owner. Here's What Surprised Me.
January 23, 2026
Fionna Kossmann
Claims Product Owner
Here's a confession: I used to think AI was magic. A black box. Impossible to understand. Then I spent the past year working as a claims product owner at Friendly, an AI company purpose-built for insurance. What I learnt changed how I see things.
BEHIND THE BLACK BOX—REAL PEOPLE, REAL DATA
Many people think AI is abstract. Magical. Secret. A black box where it's impossible to understand what steps were taken. I used to think this too. I was awe-struck at what was possible—and honestly, at times a bit scared.
But here's the truth: there are people behind the AI. And the AI is only as smart, effective and reliable as the people who created the information it references.
"AI is basically about data. Finding it, understanding it, and connecting it into sequences that make a coherent response."
In my work as a claims product owner, I've gained a new appreciation for structured data. I remember working on IT projects in my previous roles. The project managers would push us to create as many structured data fields as possible. (I wondered if this was just so our actuary colleagues could do better analysis.)
As a claims expert, it was frustrating. Claims assessment is about the connections. The in-between. What's stated and what's not stated. Being forced to think in structured data fields felt unnatural. Like we were misunderstood. Like the complexity of what we do wasn't fully appreciated.
Now, being on the other side—at an AI vendor—I get it. Structured data is the necessary building block for doing the impressive interconnected stuff. The AI first needs to understand the smallest data points. Then it can recognise patterns, spot discrepancies, make connections, and tell a story to the user.
Structured data is the starting point. But it doesn't have to be the end point.
THE UPFRONT INVESTMENT THAT PAYS OFF
Doing the foundational work properly matters. Applying human expertise in classification work is incredibly valuable.
If you teach the AI tool what different documents are, what they look like, where to find different bits of information, and what that data refers to—then you're on your way to magic. You'll get super-fast, reliable and genuinely helpful information out of your AI.
"Once the basics are there, that's when it's possible to do the mind-blowing stuff. Write paragraphs. Answer questions. Reason. Highlight gaps."
This is what makes a difference to claims assessors and managers. This is where AI delivers real impact and time savings.
It goes far beyond summarisation and structuring data. But you have to be willing to put in the upfront effort to get there.
UNDERSTANDING AI'S LIMITATIONS—AND WORKING WITH THEM
Without proper data structure and training, AI can produce unreliable outputs. Here's why: AI generates responses based on patterns in its training data. It doesn't actually understand facts or context the way humans do. When it encounters gaps, it fills them in with whatever seems statistically plausible.
The result? A complete, coherent story that sounds convincing—but might be partly or entirely fabricated. This is commonly known as "hallucination."
It's like when your friend who likes to embellish tells you something. You know bits of it are true. But which parts? It becomes unreliable. Trust breaks down.
"In an industry as regulated and risk-focused as insurance, reliability is critical. We rely on what's disclosed and what can be validated by supporting evidence."
This is why understanding how your AI is built and what safeguards are in place matters.
At Friendly, we address this by using generative AI selectively—as our CEO Natasha likes to say: "We just sprinkle a bit of it on top of what is already robust AI based on proven data science techniques." The generative AI writes paragraphs, for example, but only based on what's actually in the claim file. No external sources. This approach helps limit unreliable outputs while maintaining the power of AI-generated insights.
GENERIC VS. SPECIALISED—WHAT MAKES AI ACTUALLY USEFUL
AI is now embedded in everything. Your music player has an AI DJ. Your inbox has an AI assistant. Your messaging app has an "Ask me anything" feature.
But have you tried them? The answers often leave much to be desired. Entertaining, perhaps. But not always useful.
The challenge is that many of these tools are generic. Built to serve a very general purpose. They're trained on everything—which means they're experts in nothing.
"But it will get better," you might say. Yes, it will. But in the meantime, I'd rather rely on tools that are specialised and can actually help me get the job done faster and better.
"Don't make the task fit the AI. Find the right AI for the task."
When you have complex tasks—like determining the validity of an insurance claim—you need specialised AI if you want to see what's truly possible.
Specialised AI can reduce common pitfalls and unlock real power. But like humans, it requires expert training before it can truly add value to complex work.
Take insurance and claims assessment. It's its own world. Own language, own acronyms, own jargon, processes, regulations, guidelines. If you apply an off-the-shelf generic AI to this, you might be initially impressed—then disappointed when you realise the limitations.
At Friendly, our models have had over five years of specialised training on insurance and medical documents, language and processes. Our AI intrinsically understands insurance. That's why we can deliver output that goes far beyond summarisation—actionable intelligence, decision support and discrepancy detection.
THE REAL TRANSFORMATION—RETHINKING WHAT'S POSSIBLE
I've seen this multiple times now. We're limited by our own thinking about what's possible with AI.
When you embark on technology transformation, don't expect to do things the same way—just faster. You'll end up doing things very differently once you see what's possible. Processes you thought were fixed will become fluid. Questions you never thought to ask will become obvious.
That's the exciting part.
WANT TO LEARN MORE?
If any of this resonates, I'd love to continue the conversation. I've been on both sides—as a claims manager challenged by technology, and now as someone building it.
Or reach out to me directly—I'm always happy to chat about what specialised AI can actually do for your claims process.
LOOKING FORWARD
The gap between claims expertise and AI understanding doesn't have to be wide. In fact, claims professionals are uniquely positioned to shape how AI evolves in our industry—because we understand what actually matters.
These lessons continue to shape how we build at Friendly. Every conversation with a claims professional, every challenge we solve, every "aha moment" feeds back into making our AI more useful, more reliable, more specialised.
"The technology is impressive. But it's the people—on both sides—who make it work."
About the Author
Fionna Kossmann is Claims Product Owner at Friendly Health Technologies, where she develops and optimises AI-driven automation solutions for life and health insurers and reinsurers across global markets. With 13 years of experience in claims management at a leading global reinsurer, Fionna previously led quality assurance initiatives and technology-driven projects to improve claims processes across international markets.
Fionna's unique multidisciplinary background—combining Occupational Therapy with legal expertise from the University of South Africa—gives her a distinctive perspective on insurance technology. At Friendly, she focuses on creating AI solutions that support claims professionals while improving accuracy and decision-making, with a commitment to enhancing both efficiency and the human experience in claims management.