At Futurify, we strive to integrate cutting-edge technology into our solutions to provide the best service to our clients. One of our notable clients, Beniplus, relies on us to handle both medical and non-medical claims efficiently. For several years, we utilized Taggun as our receipt scanner. It has been a robust tool, reliably extracting crucial details like the amount, merchant name, and date from receipts. These details allowed us to partially automate the claim form completion process for Beniplus's customers, significantly speeding up claim processing.

‍

The Limitations of Taggun for Medical Claims

While Taggun has been effective for basic receipt data extraction, it fell short in meeting the specific needs of medical claims. Medical receipts often require additional information such as the patient’s name and insurance co-pay amounts. Taggun did not provide this data by default. We faced a choice: either customize Taggun’s engine to handle these specifics or explore alternative solutions.

‍

Exploring OpenAI Vision

We decided to test OpenAI Vision to see if it could meet our needs more effectively. Our approach involved integrating directly with OpenAI APIs. By providing a prompt and the receipt image, we could specify the context, the required data, and the expected format. OpenAI Vision delivered highly accurate results, extracting all necessary details including the patient’s name, the amount paid after insurance co-pay, merchant name, and purchase date. Impressively, it could also detect image blurriness, a common reason for claim rejections.

Here is an example response from OpenAI Vision:

‍

1{
2  "businessName": "Shoppers Drug Mart",
3  "businessAddress": "123 Main Street, Toronto, ON, M1A2B3, CA",
4  "amountPaidByClient": 20.00
5  "clientName": "John Doe",
6  "clientAddress": "456 Elm Street, Toronto, ON, M1C2D4, CA",
7  "lineItems": [
8    {
9      "description": "Prescription Medication",
10      "price": 20.00
11    }
12  ],
13  "taxAmount": 0.00,
14  "businessNumber": "N/A",
15  "categoryOfMedicalExpense": "Prescription",
16  "dateOfExpense": "2024/05/10"
17}
18

‍

Comparing Taggun and OpenAI Vision

We conducted a detailed comparison between Taggun and OpenAI Vision by processing the same set of receipts through both tools. Remarkably, both systems achieved a 94% accuracy rate across various types of receipts, including both medical and non-medical.

‍

Cost Efficiency with OpenAI Vision

One of the most significant advantages of using OpenAI Vision is the cost efficiency. The cost per scan is approximately 5 cents for about 1300 input tokens and 200 output tokens. We have not yet applied any optimizations, which we believe could further reduce costs. Additionally, with the recent announcement of GPT-4o, the cost could potentially drop to 2 cent per receipt.

‍

Addressing Challenges

One area where OpenAI Vision did not consistently perform well was in categorizing the claim. We provided a list of categories in the prompt and asked OpenAI to make the best guess for each claim. The results were not consistently accurate. To address this, we are developing our own Convolutional Neural Network (CNN) using historical claim data to determine categories more accurately.