A founder's perspective on what it actually takes to build agentic platforms that work in production
Yesterday, my co-founder Parvind's former colleague Saumik Banerji shared some sharp insights about the insurance industry's AI transformation on LinkedIn. His point about insurers needing to rethink operating models—shifting from project-based to product-aligned teams—really resonated with both of us. But as we work through migrating our entire Sage Agentic Landscape from AWS to Azure, I'm reminded that there's another story we need to tell about AI transformation in insurance.
The unglamorous one.
Most posts about AI agents and insurance ecosystems highlight what's possible. SageCMO revolutionizing marketing workflows. SageInsure transforming underwriting. SageRetail reimagining customer experiences. And yes, these capabilities are real—but few talk about the painful, unglamorous work it takes to make them actually run in production.
Right now, our team is deep in the weeds of rewiring our entire infrastructure while migrating from AWS to Azure. Why? Simple economics—we've burned through AWS credits, and Microsoft's Startup program offers $50K in Azure credits. A practical move for a startup, but far from a simple one.
Here's the reality nobody talks about:
What breaks during migration? Almost everything. Agents, knowledge bases, S3 storage, EventBridge events, databases, Route53 DNS, authentication flows. Each component seems like it should be "lift-and-shift" until it's not.
Why is it so painful? Because despite similar primitives across clouds, each comes with different defaults, quirks, and gotchas. The "experience" is never quite the same.
The time sink? Entire days lost to debugging something that worked perfectly on AWS but behaves differently on Azure. Not because the architecture is wrong, but because the ecosystem wiring is subtly different.
This connects directly to Saumik's point about operating model transformation. When insurers talk about deploying AI, they often focus on the sexy stuff—the algorithms, the insights, the customer experiences. But the real transformation happens in the messy middle: the integration work, the data plumbing, the endless debugging sessions that make intelligent agents actually work with legacy systems.
Coming from two decades in maritime operations—where system failures can mean life-or-death consequences—I've learned that operational excellence isn't optional. At sea, you can't just restart a server when something breaks. The same principle applies to insurance AI: reliability trumps features every time.
At MapleSage, this maritime discipline shapes how we approach the Sage Agentic Landscape:
Mission-critical systems demand maritime-grade reliability. You can have the smartest agentic AI in the world, but if it can't reliably integrate with your core systems under pressure, it's just an expensive demo.
Infrastructure decisions compound—plan for scale from day one. In maritime operations, you don't get to retrofit safety systems after something goes wrong. The same applies to AI platforms: that "quick fix" becomes the foundation everything else builds on.
Operational discipline beats algorithmic cleverness. The most sophisticated AI is worthless if your operations team can't deploy, monitor, and maintain it reliably across multiple client environments.
This isn't my first experience managing complex, interconnected systems. Over two decades of maritime operations taught me that the real challenge isn't in individual components—it's in the orchestration. Whether you're coordinating vessel operations across multiple time zones or orchestrating AI agents across insurance workflows, success comes down to obsessive attention to system reliability and failover mechanisms.
The parallels are striking:
In both cases, partial failures aren't acceptable. When a marine insurer needs to process a major casualty claim, our SageInsure platform can't be "mostly working"—it needs to be as reliable as the navigation systems we depend on at sea.
Building agentic AI systems requires more than just algorithms. Our tool augmented AI system approach combines multiple AI capabilities with the operational discipline I learned at sea—where system integration isn't optional, it's survival.
The Architecture Behind Our Agentic AI System
Most companies build their first AI agent today as a standalone microservice. We architect ours as distributed, event-driven components within a fault-tolerant ecosystem. When SageInsure processes a marine casualty claim, it doesn't work in isolation:
The Maritime Engineering Approach to Code Agentic AI Apps
In maritime operations, redundant systems aren't luxury—they're mandatory. Engine failure protocols, backup navigation, emergency communications. We apply this same fault-tolerance when we code agentic AI apps:
Maritime System: Primary GPS → Backup GPS → Dead ReckoningAgentic AI System: Primary LLM → Backup Model → Rule-Based Fallback
Technical Differentiation from Other AI Systems
While other AI capabilities focus on individual generative AI use cases, we engineer for production reality. Our tool augmented AI system handles the technical challenges every agentic AI tutorial skips:
Beyond Simple Generative AI Use Cases
This isn't prompt engineering at scale. Our agentic AI system performs actual new content generation while maintaining audit trails, compliance validation, and regulatory reporting—simultaneously.
Example workflow:
This maritime-grade engineering approach means our agentic AI system doesn't just work in demos—it scales under the operational pressure that sinks other AI capabilities.
If you're leading AI transformation at an insurer, here's my blunt advice based on living through multiple migrations and system rewrites:
1. Budget for the Hidden Costs That AI platform you're evaluating? Double your timeline and budget estimates for integration work. The vendor demos never show the three weeks you'll spend figuring out authentication flows with your legacy policy admin system.
2. Invest in Your Orchestration Layer Don't just buy AI tools—build the infrastructure that lets them work together reliably. At Sage, our Agentic Landscape isn't just a collection of smart agents; it's an orchestration platform that manages how they collaborate, share context, and handle failures.
3. Build for Marine-Grade Reliability Your AI platform needs to handle the unexpected with the same rigor as maritime safety systems. At MapleSage, every agent in our Agentic Landscape is designed with multiple fallback mechanisms—because in insurance, like at sea, there's no room for single points of failure.
4. Start Small, Scale with Discipline Begin with one critical workflow and nail the integration completely before expanding. Maritime operations teach you that you can't shortcut safety protocols—the same applies to AI deployment patterns.
5. Prepare for Culture Change Saumik is spot-on about the shift from project-based to product-aligned teams. But make sure your people understand this isn't just about reporting structures—it's about taking ownership of outcomes, not just deliverables.
Why share all this operational reality? Because every difficult migration, every system integration challenge, every "why doesn't this work the same way on Azure?" moment is also a chance to rethink architecture, improve resilience, and tighten orchestration.
This Azure migration—painful as it is—is forcing us to examine every assumption we made about how Sage agents communicate, share state, and recover from failures. The result will be a more robust platform that can scale with our insurance partners' needs, built with the same operational discipline that keeps vessels safe at sea.
The insurance industry is indeed at a turning point, as Saumik noted. AI, especially GenAI and Agentic AI, is driving real transformation. But that transformation doesn't happen in PowerPoint slides or vendor demos—it happens in the operational trenches, in the integration work, in the unglamorous engineering that makes intelligent agents actually work with real insurance workflows.
At MapleSage, we're not just building AI tools for insurance—we're building the operational infrastructure that lets insurers rethink how they operate. SageCMO doesn't just generate marketing content; it orchestrates entire campaign workflows with maritime-grade reliability. SageInsure doesn't just assess risk; it coordinates underwriting processes across multiple systems and stakeholders with the same precision we demand from navigation systems.
This is what the future of insurance looks like: not just smarter tools, but smarter orchestration built with operational discipline. Not just AI features, but AI-native operating models that can handle the unexpected.
And yes, it's messy. Yes, it requires operational discipline that most tech companies skip. Yes, there will be more late-night debugging sessions.
But for insurers willing to do the hard work—the integration work, the culture change, the fundamental rethinking of operating models with maritime-grade reliability standards—the payoff is transformation that goes far beyond what any individual AI tool can deliver.
From our Dubai headquarters, Parvind and I are building something different: an AI platform where operational excellence isn't an afterthought—it's the foundation.
Captain Chris Thor Illum is COO and Co-Founder of MapleSage FZCO, bringing over 20 years of maritime operations experience to building the Sage Agentic Landscape. Connect with him on LinkedIn to continue the conversation about operational excellence in insurance AI transformation.