Sedgwick

Sedgwick, a global claims management firm operating across 80 countries, receives approximately 1.7 million pages of digital claims-related documents every day. In the workers' compensation and insurance space, each document must be individually vetted by a claims examiner to determine validity and appropriate handling — a labor-intensive process with no shortcuts in a heavily regulated environment. This volume created a severe administrative bottleneck: examiners spent the bulk of their time on document triage rather than complex case judgment, slowing resolution times and limiting the firm's capacity to scale without proportional headcount growth.

In April 2023, Sedgwick launched Sidekick, a proprietary generative AI tool built on OpenAI's GPT-4, designed to automate document summarization, data classification, and claims analysis. The initial phase deployed ChatGPT within Sedgwick's secure environment to evaluate a single focused use case — document summarization — before any broader rollout. In a second phase, Sidekick was integrated directly into Sedgwick's proprietary claims administration systems, shifting it from a standalone tool into an embedded workflow component. This integration enabled low-touch automation on routine claims, freeing examiners to concentrate on higher-complexity cases. The platform was architected to support multiple LLM instances for varying purposes, making it extensible across different document and claim types beyond the initial workers' compensation focus.

After processing 50,000+ documents during the pilot phase, Sidekick achieved a 98%+ accuracy rate in document summarizations as evaluated by examiners — a meaningful bar given regulatory scrutiny in claims handling. Documents up to 25–30 pages in length were summarized in minutes, compared to the manual effort previously required per file. Adoption grew from roughly 500 employees at the end of the first phase to 1,100+ examiners by the second phase, with staff proactively requesting broader access rather than resisting the tool.

• 98%+ document summarization accuracy (examiner-evaluated)
• 50,000+ documents processed in pilot
• 1,100+ employees actively using the tool
• Documents up to 30 pages processed in minutes

• Iterative prompt engineering with domain experts is non-negotiable: accuracy only reached 98%+ after calibrating the AI to replicate examiner judgment through real feedback cycles.
• Phased integration beats big-bang rollout: starting with a contained pilot before embedding into core systems de-risked adoption and built internal confidence.
• Combining AI with institutional knowledge creates durable advantage: 50 years of claims process expertise made Sedgwick's implementation more defensible than a generic LLM deployment.
• Digital triage determines ROI: identifying which document types and claim categories benefit most from AI — rather than applying it uniformly — is where the productivity gains are found.