(Image credit: Dan Woods, CEO, Socotra.)
Dan Woods, CEO, Socotra, does not lack for conviction when it comes to the insurance industry’s current fixation on artificial intelligence. In a recent conversation with Insurance Innovation Reporter, he offered a view that cuts against the prevailing momentum: much of what carriers are seeing today is not production-grade capability, but performance.
Woods illustrates the point with a metaphor he created for a recent presentation—one that frames the dynamic he sees playing out between vendors and carriers. A bullfighter, labeled “IT vendor,” stands poised with a red cape emblazoned “Vibe-Coded AI Demo,” as a charging bull labeled “Insurance Carrier” bears down.
At the center of Woods’ critique is a distinction he returns to repeatedly: the difference between AI demonstrations and what he calls “production-grade AI.”
From Demos to Production
Woods by no means disputes AI’s potential. On the contrary, he breaks its value into three clear domains: improving IT velocity and engineering efficiency, enhancing business decision-making, and increasing productivity and accuracy for insurance workers. The third category—AI assisting underwriters, claims handlers, and operations staff—has received the most attention in the industry, but Woods argues that the first is the most consequential.
Dan Woods, Founder and CEO, Socotra.
Without modern engineering foundations, he insists, AI cannot safely or effectively scale.
This is where the industry’s enthusiasm runs ahead of its capabilities, in Woods’ view. Demonstrations are easy to produce. They can be compelling, even dazzling. But they do not necessarily translate into systems that can be deployed, maintained, and evolved in a live insurance environment.
“The difference between a simple app and a production system is like microwaving a burrito versus a three Michelin star meal,” Woods says.
The concern is not that vendors are being disingenuous, but that the gap between demonstration and reality is not always clear to buyers under pressure to act.
Legacy Systems Don’t Disappear
Much of that gap, Woods argues, comes down to legacy systems and the technical debt they carry.
He is particularly critical of the notion that AI can effectively “wrap” or modernize legacy environments without addressing their underlying structure. Encasing a legacy system in APIs or layering AI on top may create the appearance of modernization, but does not eliminate the underlying complexity, he warns. When a carrier needs to make a substantive change—launch a product, respond to regulation, adjust underwriting logic—that complexity resurfaces.
Similarly, translating legacy code into a modern programming language does not solve the problem if the underlying logic remains opaque.
“Why is it better if I translate it to a massive Java codebase understood by zero people?” Woods asks.
In his telling, AI does not erase technical debt. It makes its consequences more immediate.
Woods characterizes much of the current landscape as “legacy as a service,” a formulation that reflects his broader critique of how some core system providers have approached modernization—migrating existing architectures to the cloud without fundamentally rethinking them.
What AI Actually Requires
If AI cannot fix legacy systems, what does it require?
Woods points to a set of characteristics that make software environments “AI-compatible”: true modularity, open languages and formats, accessible documentation, and real-time, queryable data. These are not new ideas. In fact, he emphasizes that they are the same principles that have long defined sound engineering practice.
“What’s good for engineers has always been good for AI,” he says.
The difference now is that AI amplifies the consequences of ignoring those principles. Systems that are tightly coupled, poorly documented, or dependent on proprietary constructs are difficult not only for humans to work with, but for AI to operate within.
This has implications for how carriers evaluate vendors. Woods contrasts what he sees as “mature software” companies—where products are standardized, documentation is sufficient for independent use, and customization is constrained—with the more fragmented, heavily customized environments common in the insurance industry.
In mature environments, he notes, customers do not run bespoke versions of software. They adopt a product and operate within its framework. That model, he suggests, is far more conducive to continuous improvement, including the integration of AI.
The Risk of Overbuilding AI
Woods also cautions carriers against responding to AI’s rise with large, internally driven development efforts.
His advice is direct: avoid massive AI projects, at least for now.
The technology is evolving too quickly, he argues, to justify large, long-term investments in bespoke solutions. He likens the current moment to the mid-1990s, when hardware and software cycles were so rapid that systems could become obsolete almost as soon as they were deployed.
Instead, he advocates for modular approaches that allow components to be swapped out as models improve, and for careful consideration of where to build versus where to buy. Systems that rely on interchangeable AI models—rather than hardwiring a single approach—are better positioned to adapt.
At the same time, Woods suggests that the industry is not paying enough attention to a critical dimension of AI: user experience.
An Unfinished Interface
If the technology itself is advancing rapidly, the way users interact with it remains unsettled.
Woods compares the current state of AI interfaces to the early days of the web or mobile computing, when conventions were still forming and design patterns were in flux. Over time, he expects AI to become less visible as a distinct feature and more embedded within workflows.
“In 10 years, AI is just going to be the way an interface is made,” he says.
That shift has implications for how vendors position their offerings today. The proliferation of branded “assistants” and standalone AI features may, in retrospect, appear transitional—analogous to earlier eras of software design that emphasized novelty over integration.
Woods expects a move toward AI that is less conspicuous but more effective: embedded, context-aware, and seamlessly integrated into everyday tasks.
From Features to Infrastructure
For Woods, the defining characteristic of meaningful AI adoption is not novelty but deployment.
Production-grade AI, in his view, is not something demonstrated on a stage or previewed in a roadmap. It is something that ships, integrates, and operates as part of a system. He points to familiar examples—such as AI features appearing within widely used applications—as a model for how enterprise software should evolve.
That perspective extends to emerging architectural elements such as Model Context Protocol (MCP), which Woods describes as a new layer of interaction between AI agents and enterprise systems.
“User interfaces are how software talks to humans. APIs are how it talks to software. MCP servers are how it talks to AI,” he explains.
If that framing holds, insurers will need to think not only about data and APIs, but about how their systems expose context and functionality to AI itself.
A More Measured View of AI
Woods’ argument is not that insurers should slow down on AI. Rather, it is that they should be more deliberate about how they engage with it.
The risk is not in adopting AI too early, but in misunderstanding what is being offered—confusing demonstrations with deployable capability, and assuming that new layers of technology can compensate for unresolved structural issues.
Woods’ bullfighter metaphor is deliberately provocative, but it reflects a genuine concern: that in the rush to embrace AI, carriers may find themselves reacting to signals that are more theatrical than substantive.
For Woods, the path forward is less about chasing the latest capability and more about ensuring that the underlying system is ready to support it. Only then, he suggests, can AI move securely from performance to production.
Socotra CEO Dan Woods on Innovation and Insurance’s Technology Gap