The Challenge
A regional insurance carrier was processing hundreds of inbound claims submissions every day, arriving as a mix of plain-text emails, scanned PDFs, and Word documents. Adjusters were spending the majority of their shifts on manual data entry—opening each message, reading the attachment, and re-typing claimant names, policy numbers, incident dates, coverage codes, and loss amounts into the CRM one field at a time.
The consequences were tangible: transcription errors introduced downstream billing disputes, slow intake meant claimants waited days for acknowledgement, and experienced adjusters—whose judgment is irreplaceable—were buried in clerical work instead of assessing complex claims.
Our Solution
XY Space designed and delivered a fully custom AI workflow tailored to the carrier's specific document formats, CRM schema, and regulatory obligations.
Intelligent document ingestion — The system monitors the carrier's dedicated claims intake mailbox in real time. When a new email arrives, the pipeline fetches every attachment regardless of format: PDFs (both machine-readable and scanned), DOCX files, and inline message bodies are all handled. Optical character recognition is applied where needed before the AI extraction layer runs.
Structured field extraction — A fine-tuned extraction model, trained on the carrier's own historical claims documents and insurance terminology, identifies and normalizes the fields the CRM requires: claimant identity, policy number, date of loss, peril type, estimated loss value, contact details, and supporting documentation references. Confidence scores are generated for every extracted value.
Direct CRM integration — Validated records are written to the carrier's CRM via API in real time, with full audit trail metadata attached to each submission so compliance teams can trace every value back to its source document.
The Quarantine Workflow
The most critical design decision in this engagement was what happens when the AI is not certain. We built a dedicated quarantine workflow from the ground up, purpose-built for this client's operations:
- •Any field extraction that falls below the confidence threshold is held in a quarantine queue rather than written to the CRM.
- •Adjusters receive a structured review task in their existing dashboard, showing the source document side-by-side with the AI's proposed values and the specific fields it flagged.
- •The reviewer can approve, correct, or escalate each record with a single action—no re-keying the fields the AI got right.
- •Every human decision is logged, and those corrections feed back into the model's improvement cycle on a rolling basis.
This quarantine layer is not a fallback—it is a deliberate part of the workflow. The goal was never to remove humans from claims intake entirely, but to focus their attention exclusively on the cases where their expertise actually matters.
Human-in-the-Loop as a Design Principle
Throughout the engagement, XY Space treated human oversight as a first-class requirement, not an afterthought. Insurance is a domain where a single incorrect field can affect claim validity, fraud detection flags, or regulatory reporting. No confidence score eliminates that risk entirely.
By building the quarantine workflow before going live, the carrier's team could observe the AI's behavior, calibrate the confidence thresholds to match their risk tolerance, and build institutional trust in the system incrementally. Automation coverage increased as confidence grew—not the other way around.
Results
- •50% reduction in manual data entry labor within the first quarter of deployment
- •Average claims intake time dropped from hours to minutes
- •Data entry error rate reduced significantly, with every automated write traceable to a source document
- •Adjusters redirected from data entry to higher-complexity claims assessment
- •Quarantine review rate declined steadily as the model improved on the carrier's real-world document mix
Technology
- •Custom AI extraction model trained on the carrier's document corpus
- •OCR pipeline for scanned PDFs and image-based attachments
- •Real-time email monitoring and attachment routing
- •Confidence scoring and threshold-based quarantine queue
- •CRM API integration with full audit trail
- •Human review dashboard embedded in existing adjuster workflows