I'll admit it - six months ago, I was an AI skeptic.
Not the kind who dismisses technology outright. I'd spent enough time in aerospace to know that innovation matters. But AI in 2020 felt like blockchain in 2017 - every company claiming to be "AI-powered" while doing glorified if-then statements under the hood. I'd seen enough pitch decks with "machine learning" slapped on slide three to be suspicious.
Then I actually had to build something with it.
By summer 2020, I'd been heads down in trade claims for six months. I understood the problem space intimately - retailers submitting millions of claims to consumer goods companies, each one needing validation against promotional agreements. Did the promotion actually run? Was the discount applied correctly? Does the math add up?
At the companies we were talking to, claims analysts were spending 30% of their time just tracking down information. Not analyzing. Not strategizing. Just hunting through systems, cross-referencing spreadsheets, trying to match a deduction to the promotion that caused it.
This wasn't a technology problem in the traditional sense. The data existed. The promotions were documented somewhere. The challenge was connecting the dots at scale - matching patterns across millions of transactions while accounting for regional variations, retailer-specific formats, and the inevitable edge cases.
I remember sitting in a customer workshop thinking: This is exactly the kind of problem that should bore a computer but drives humans insane.
The breakthrough came when our engineering team demonstrated an early machine learning model. Nothing fancy - just a classifier trained on historical claims data, learning which claims matched which promotions and why.
The "why" part hooked me.
Traditional automation would have been rule-based: if claim type equals X and retailer equals Y and date range matches Z, then match to promotion. Brittle. Breaks the moment something unexpected appears.
What the ML model did instead was learn patterns from past decisions. When claims analysts had matched similar deductions before, the model absorbed those signals. When they'd flagged anomalies, it learned those patterns too. The system wasn't following our rules - it was inferring its own rules from thousands of examples.
To put things in perspective: a new claims analyst might take months to recognize that certain retailers always submit claims with a particular format quirk. The ML model picked up on that pattern in its first training run.
From a claims analyst's standpoint, you're suddenly not drowning in routine matches. The system handles the 80% that follow established patterns. You focus on the 20% that actually need human judgment.
From a finance team's perspective, fewer manual touchpoints means fewer errors. Less leakage. Faster resolution.
From my perspective as a PM - and here's where my thinking started to shift - we weren't just automating a task. We were encoding institutional knowledge into software.
Every enterprise has this challenge. Critical decisions get made by experienced employees who've accumulated years of pattern recognition. Those patterns live in their heads. When they leave, the knowledge leaves with them.
I'd seen this at Boeing. Senior structural engineers who could look at a stress analysis and immediately spot what was off - not because they were running calculations faster than anyone else, but because they'd seen a thousand similar analyses and developed intuition. That intuition was incredibly valuable and completely undocumented.
What machine learning offered was a way to capture some of that institutional memory. Every time a claims analyst made a decision, the system learned from it. Approve this match? The model notes the pattern. Reject this claim? It learns what made it suspicious.
Obviously, this isn't a perfect replacement for human expertise. The model can learn patterns but not reasoning. It can recognize that certain combinations of factors usually lead to rejections without understanding the underlying business logic.
But that's fine. The goal isn't to replace the expert - it's to preserve their patterns so the next generation of analysts doesn't start from zero.
I came into SAP Eureka thinking AI was mostly marketing buzz. Six months of actually building AI-powered features changed my mental model entirely.
Here's what I started to see differently:
AI as amplification, not replacement. The best use cases we found weren't about eliminating jobs. They were about eliminating the boring parts of jobs. Let machines handle pattern matching at scale so humans can focus on exceptions, strategy, and relationship management.
Data quality matters more than algorithm sophistication. Our fanciest models performed worse than simple ones when fed garbage data. The unsexy work of cleaning, normalizing, and structuring training data created more value than any algorithmic breakthrough.
Domain expertise is the differentiator. Generic AI capabilities are becoming commoditized. What makes enterprise AI valuable is deep understanding of specific business problems. Knowing what patterns to look for in trade claims specifically - that's the moat.
Iteration beats perfection. Our first model was mediocre. So was the second. By the fourth iteration, with continuous feedback from actual users, it started getting genuinely useful. This felt familiar from startup days - ship, learn, iterate.
We're not done. The December launch will put SAP Intelligent Trade Claims Management in front of real customers at scale. I'm nervous and excited in equal measure.
What I know for certain is that this experience rewired something in how I think about technology problems. I spent years as an aerospace engineer focused on deterministic systems. Given input X, output Y, every time. Predictable. Testable. Safe.
Machine learning operates differently. It's probabilistic. It makes mistakes. It requires trust that the patterns it learned are the right patterns.
That uncertainty used to bother me. Now I see it as the price of capability. You can't teach a rule-based system to recognize patterns it was never explicitly programmed for. ML can - but it trades determinism for adaptability.
Six months ago, I would have said AI is overhyped. Today, I'd say it's underutilized - at least in enterprise software. The problem isn't that ML can't add value. The problem is that most implementations are shallow. Slap a chatbot on it, call it AI, ship it.
The real opportunity is in going deep. Understanding a specific domain, identifying where pattern recognition can substitute for manual labor, then building systems that actually learn from use.
That's what we're trying to do with ITCM. Ask me in a year whether we succeeded.
Cautiously optimistic, for the first time in a while.