The company is an online, direct-to-consumer auto finance intermediary. It helps consumers refinance existing auto loans to lower their interest rate or monthly payment. Customers apply and complete their loans online, and the company works with banks, credit unions, and institutional investors to fund them. 

Before an auto loan can be funded, the company’s funders review the customer’s supporting documents, such as a driver’s license and bank statements. That document review is where this project began. 

The company did not start this project simply because AI was available. It started because of how loan documents reached its funders. Customers sent documents through whichever channel was easiest, and funders then had to open each one, identify what it was and check it against the loan request by hand. Customers also received no automatic email when a status change meant documents were due. On top of that, legacy fax queues and upload screens added yet more channels, none of which could read a document. 

Before launch, an application averaged 20.2 days (about 3 weeks) from Pending Sales to Sold. This process needed to change to speed up document processing and to handle the extra workload as ORL rapidly scales. 

From January 2025 to early February 2026, only 0.6% of the company’s documents came through the customer portal. The rest arrived by LOS to upload, email, text message, fax, and a handful of other routes. Email and text message alone carried 48.8%. 

A model can only read what is shown. Most documents arrived through channels where they could not be automatically read at the point of uploading. Funders then checked each document against the loan request manually. 

So, the AI-driven Document Intake had to do three things together: get customers upload through the portal, read and classify each document as it arrives, and put everything a funder needs to decide in one place. 

Insight’s approach: make the intake experience part of the solution 

The goal was not simply to add AI to document handling. The workflow needed to encourage customers to use a channel where documents could be processed, classify and read each upload as it arrived, and bring the information funders needed into one review experience.

The solution we proposed would use Microsoft’s AI based system Azure Document Intelligence at its core. ADI Microsoft’s service for classifying documents and extracting fields from them. Around it, we built the pieces that make it usable day to day. 

• Upload with instant feedback. Customers drag and drop JPG, JPEG, PNG, PDF, or HEIC files up to 10 MB. A progress bar runs while the document is read, then a plain-language message says whether the document matched their application, matched but could not be verified, or needs a second look. 

• AI document classification and data extraction. Each upload is sorted by document type, and its fields are extracted. Anyone can override the AI’s choice, which triggers a fresh extraction because the fields to pull depend on the type. 

• Automatic document splitting. If a customer scans a license and two bank statements into a single file, the system separates it into individual files by document type. 

• One funder review queue. Funders see the extracted fields, a confidence score for each and the application data alongside them, flagged as a match, a mismatch or for information. They approve or reject, and the next document loads automatically. A rejected document triggers an email to the customer with the funder’s reason, sent once per document. 
• Resilience when the AI is down. A health-check service pings Azure Document Intelligence every four hours, and every hour once it detects an outage. Documents uploaded during an outage are saved without being read. A recall service picks them up once the AI is back, along with anything still pending to review. 

Upload and Review Flow

Flow showing the ADI health check, the AI classification path, the recall service when ADI is down, and funder review.

When the AI is available, a document is read at the point of uploading. When it is down, the file is saved, and the recall service reprocesses it later. Either way, funders see matches and mismatches against the application. 

Intake figures are shares of all documents received in each period, so the two periods can be compared directly. Pre-launch intake covers January 2025 to 9 February 2026; pre-launch cycle times cover 1 November 2025 to 1 February 2026. 

• Early signal, not a settled trend. The results cover only seven weeks. They show direction, and a longer window is needed before we can call them a trend. 

• AI is not the only driver of the faster cycle. The post-launch cycle time counts only applications that had reached the next status by 5 April, so slower ones are excluded, and the average may look shorter than it is. Phase 1 also changed how documents arrive and how funders review them, so the improvement cannot be credited to AI alone. 

• Accuracy applies to document type, not individual fields. The 95.5% compares the AI’s first document type with the funder’s final type, and only approved documents. It does not show whether each extracted field was correct. 

• Rejections do not measure AI errors. About 64.5% of rejections include documents from channels where AI was not used, so the rejection rate cannot be read as the AI’s error rate.