For most legacy enterprises, the dream of integrating generative AI often crashes against the reality of decades-old infrastructure. The challenge isn’t usually the AI itself, but the “plumbing”—the rigid, document-heavy workflows that keep global industries running. For Sedgwick, a global risk and claims administration firm based in Memphis, Tenn., the hurdle was a mountain of medical records and case documentation that required human examiners to sift through thousands of pages under tight deadlines.
Rather than attempting a “rip-and-replace” overhaul of its core systems, the company has found success by scaling AI into legacy workflows through a strategic layering approach. By deploying a proprietary platform called Sidekick+, Sedgwick is now using OpenAI’s GPT-4 technology to act as an intelligent interface—a “wrapper,” as the company describes it—that sits atop existing systems to summarize massive volumes of data without disrupting the underlying claims infrastructure.
The strategy is a blueprint for how document-intensive organizations can modernize. With roughly 30,000 employees globally, Sedgwick operates at a scale where even a minor disruption to core operations can have significant ripple effects for clients, and claimants. By focusing on a “human-in-the-loop” model, the company has shifted the role of the claims examiner from a manual searcher to a high-level decision-maker.
The Architecture of a ‘Wrapper’ Strategy
The technical friction in most AI rollouts stems from the attempt to force LLMs to interact directly with legacy databases, often resulting in fragile “screen scraping” or costly system migrations. Sedgwick avoided this by leveraging its existing ecosystem of services and APIs. Because the company had already built a foundation of APIs to move data—such as notes and diaries—between platforms, the AI could be integrated as a complementary layer rather than a replacement.
Sean Safieh, the company’s CIO of global platforms and digital solutions, explains that Sidekick is designed to simplify the user experience by hiding the complexity of prompt engineering. Instead of requiring examiners to learn how to “talk” to an AI, the system provides a simple button that triggers a pre-defined, optimized prompt in the background.
This architectural choice allows the company to maintain “observability,” ensuring that the AI queries do not overtax legacy platforms. By monitoring performance in real-time, the team can ensure that the AI tools enhance speed without breaking the systems they rely on. This is particularly critical given the company’s use of other automation tools, such as optical character recognition (OCR) and robotic process automation (RPA), which already handle the initial scanning and sorting of documents.
Efficiency Gains and the Audit Shift
The primary value proposition for Sidekick is the drastic reduction in “hunting and reading” time. In a typical claims workflow, an examiner might spend 10 to 15 minutes reading a complex document; Sidekick can provide a comprehensive summary in one to two minutes. This efficiency does not replace the human; rather, it frees the examiner to focus on the claimant’s needs and the final decision.
Beyond individual productivity, the AI has fundamentally changed how Sedgwick handles quality control. In a traditional environment, auditing is a sampling exercise—reviewing a tiny subset of claims to infer the health of the overall process. With the integration of AI, the company has moved toward a model of total oversight.
Safieh notes that the company is now auditing every claim, ensuring that the information provided and the resulting decisions are acceptable across the board. This shift from representative sampling to universal auditing significantly reduces the risk of inconsistent decision-making in high-stakes claims environments.
| Metric | Traditional Workflow | AI-Enhanced Workflow (Sidekick+) |
|---|---|---|
| Document Review Time | 10–15 Minutes | 1–2 Minutes (Summary) |
| Audit Scope | Subset of claims | Every claim |
| Accuracy Rate | Human Baseline | 98% to 99% (Parallel Testing) |
| Decision Maker | Human Examiner | Human Examiner (AI-supported) |
Building Guardrails for Sensitive Data
In the insurance and risk sector, data privacy is not a feature—We see a legal requirement. Scaling AI into legacy workflows requires rigorous guardrails to prevent sensitive claimant data from leaking into public models or being used for third-party training. Sedgwick addressed this by deploying its models through Azure OpenAI Service, ensuring the data remains within the company’s secure cloud ecosystem.
To combat the industry-wide problem of “hallucinations”—where AI confidently presents false information as fact—the company implemented a dual-layer defense. First, the AI is fed only “clean” data sets. Second, the system is programmed to admit ignorance; if the AI cannot interpret a piece of information, such as a messy handwritten note, it is required to explicitly define that inability in its output.
the company has prioritized “provenance.” If a claim decision is ever challenged, Sedgwick must be able to reconstruct the logic used to reach that conclusion. The Sidekick platform stores the original document, the AI-generated summary, and the final recommendation made by the human examiner, creating a transparent audit trail.
Lessons for the Enterprise CIO
The rollout of Sidekick highlights a critical tension in current tech leadership: the urge to chase the newest tool versus the need for operational stability. Safieh cautions other executives against “ripping and replacing” tools too frequently, arguing that constant switching actually slows down enterprise velocity.
One of the most significant takeaways from the Sedgwick experience is the importance of trust. The company did not simply deploy the tool; it ran extensive pilots and “parallel testing,” where AI outputs were compared directly against human outputs. This process revealed a 98% to 99% accuracy rate in document summarization, which provided the internal evidence needed to win over skeptical users.
Safieh admits that if he were to do it again, he would focus on user adoption even earlier. The technical implementation—which took about three months for the first use case—was faster than the cultural process of getting people to trust and embrace the tool.
As the company looks forward, the focus is shifting toward “agentic AI”—systems that can not only summarize data but execute sequential steps in a workflow. This will require new decisions regarding orchestration and whether to move toward the Model Context Protocol (MCP) to better string together complex services. For now, the company remains committed to a flexible architecture that avoids vendor lock-in, ensuring they can switch or blend LLM providers as the market evolves.
Disclaimer: This article is for informational purposes only and does not constitute legal or financial advice regarding insurance claims or AI implementation.
The next phase of Sedgwick’s AI evolution will likely center on the orchestration of these agentic capabilities across more complex, multi-step claims processes. We will continue to monitor how these shifts affect the balance between automated efficiency and human oversight in the risk management industry.
Do you think “human-in-the-loop” is a permanent necessity for AI in legacy industries, or just a stepping stone? Let us know in the comments.
