The problem
AEC Benefits manages health and dental benefits for employees at construction and engineering firms. They receive thousands of documents annually: enrollment forms, claim submissions, address changes, beneficiary updates, benefit change requests.
Every document arrives in different formats. Handwritten, PDF, email attachment. Data needs to be extracted, entered into their system, validated for completeness, and stored for compliance. A single missing field can block processing.
Document intake required repeated data entry and validation. The opportunity was to create a more consistent process while preserving human review for decisions and exceptions.
The solution
We built an AI document processing system that works like this:
1. Documents arrive via email or upload. 2. AI reads the document and extracts key data fields (name, coverage type, dates, dependents). 3. AI validates the extracted data against their requirements. 4. If complete, it auto-loads into their system. 5. If incomplete, it flags the missing field and routes back to the member with a specific request. 6. High-confidence extractions bypass manual review. Low-confidence extractions go to a human for 10-second validation.
The system uses Claude AI for document understanding and OpenAI for image processing. All data is encrypted and stored in a secure, compliant database.
Results
The process became measurable
Human review remains part of the system
Team retention improved
Compliance maintained
Why this worked
This project worked because three things aligned:
Clear problem
Repeatable workflow
Measurable outcome
What didn't work at first
Initially, we tried to automate 100% of documents. Some documents were handwritten and difficult to read. Some had unusual formats. The accuracy dropped when we aimed for zero human touch.
The solution was to set a confidence threshold. High-confidence extractions process automatically. Medium and low-confidence go to human review. This "human in the loop" approach gave them the speed benefits without sacrificing accuracy.
FAQ
Can this be replicated for other document types?
How much did this cost?
What happens when documents are malformed or in a new format?
If your business processes documents and you're wondering if AI automation could help, let's talk about what's possible.
