Introduction

The transportation industry, like many others, is experiencing a transformative wave driven by artificial intelligence and Large Language Models (LLMs). Document processing, which has traditionally been a manual and time-consuming task, is being revolutionized by LLM-powered solutions.

Consider how DocuSign has transformed contract processing or how Dropbox's AI features now automatically categorize and extract information from documents. These innovations mirror a broader trend: using AI to turn unstructured data into actionable insights. In the trucking industry, rate confirmations represent a perfect candidate for such automation.

Rate confirmations, the formal agreements between carriers and brokers, have long been a bottleneck in operational efficiency. These documents come in various formats, contain complex information, and require careful attention to detail - making them an ideal candidate for LLM-powered automation.

In the following sections, we’re going to dive deeper into how we leveraged LLMs into building an automated rate confirmation processing system that powers almost 70% of our booked loads (aka jobs) per month. This was our very first foray into building LLM backed automations and led to evangelizing and motivating more use cases that have been built and deployed since then.

The Challenge

The traditional rate confirmation process presents multiple challenges:

• Documents arrive in various formats. Majority of brokers share them as PDFs or urls to download from, however there are some other variations that we also encounter now and then.
• Information needs to be accurately extracted and validated.
• Each broker has unique document formats and requirements. There is no industry standard template for a rate confirmation document. However there is a set of data that is lowest common denominator across the industry.
• Real-time communication with multiple stakeholders is required. All parties need to be aware when a load is booked and scheduled to run.
• Compliance and accuracy are absolutely critical. Inaccuracies in this process mean drivers and brokers shook hands on terms they didn’t agree upon.

Our Solution

Before I go into the details let’s set up the visual aid first and then dig into the implementation right after.

Document Retrieval and Processing

When we receive a rate confirmation document, the system creates a support ticket and begins processing. The first step is identifying what kind of document we've received - it could be a PDF, an HTML email, or a special format that needs unique handling. Eventually all these documents are funneled into the system as PDF files, which then get converted to images. Why images? That comes in next section, which is where the actual LLM magic happens.

Technical Implementation

Before we dive into the implementation details, I’d like to give a shoutout to my teammates, Keaton Subjeck and Dan Kras, who have been instrumental in building and maintaining this feature at CloudTrucks.

Now then, lets look into how the rate confirmation documents get processed through several stages:

1. Image Processing
   • Documents are converted to JPEGs because rate confirmations often come in various formats (PDFs, scanned images, HTML) and direct text extraction can miss crucial layout information
   • Multi-page handling as sequential images preserves document structure and page ordering, which is critical for correctly interpreting multi-stop loads
2. Model Selection
   • Primary: GPT-4o
   • First Fallback: Claude 3.5 Sonnet
   • Second Fallback: Gemini 2.0 Flash
3. Input Structure - Careful input formatting is essential because it directly impacts extraction accuracy:
   • System message establishes the context for accurate industry-specific interpretation
   • Detailed prompt includes:
     • Specific formatting requirements to ensure downstream compatibility
     • Industry-specific rules for handling dates, times, and special cases
     • Validation parameters for critical fields
   • Multiple images are processed together to maintain context across pages
4. Output Processing - Robust response handling is implemented because raw model outputs can be inconsistent:
   • JSON validation catches structural errors before they affect downstream systems
   • Multiple error handling paths because different failure types require different recovery strategies
   • Specific error counters enable:
     • Performance monitoring across different document types
     • Identification of systematic issues
     • Cost optimization opportunities
5. Data Extraction - Field extraction is prioritized based on business impact:
   • Load identifiers and rates are primary because they affect billing accuracy
   • Stop details require special handling due to timezone and formatting variations
   • Contact information needs validation against existing broker records
   • Equipment requirements affect load matching and safety compliance
6. Output Format - Structured output is critical because it feeds into automated workflows:
   • Dictionary format enables direct database integration
   • Null returns trigger manual review processes
   • Error logging with specific types enables automated recovery procedures
   • Counter metrics feed into system monitoring and optimization

This structured approach ensures reliable data extraction while maintaining the flexibility to handle variations in document formats and content. The multi-model strategy balances processing costs with accuracy requirements, while comprehensive error handling maintains system reliability.

Finding the Right Match

The system then tries to match this document with the correct job by looking for the load number or other fallback parameters to find a 1-1 match.

If no match is found, the document is safely stored for later processing. If a match is found, the system checks if the load is eligible for processing.

Verification and Upload

For eligible jobs, the system:

• Uploads the document to our secure storage and links to the matched job
• Creates necessary tracking tickets for the CT ops team
• Notifies the driver and broker that rate confirmation for the job has been received

Automated Approval Process

The system then attempts to automatically approve the document by running it through a series of checks. The primary role here is to look for data parity between the original job that was booked and the information that we parsed from the rate confirmation document. Two outcomes are possible:

• Auto Approval: If everything matches between existing load details and processed document, the system automatically approves
• Manual Review: If there are significant discrepancies, the system routes it for manual review. First we expose the discrepancies to the driver on the app for them to review and confirm that the details we received from the doc are correct. If they don’t respond within a pre-set timespan then we escalate this to an internal ops agent.

Final Steps

Once approved (either automatically or manually), the system:

1. Updates the status of the load
2. Sends notifications to all relevant parties
3. Creates a complete record of what happened

This entire process, which traditionally took hours of manual work, now happens in minutes or even seconds, ensuring quick and accurate document processing while maintaining all necessary checks and balances.

So, who reaps the benefits?

• Customer: The automation revolutionizes the driver experience by reducing rate confirmation wait times from 30-60 minutes to just 2-3 minutes for eligible loads. Drivers receive instant notifications, clear communication about any discrepancies, and automated handling of special cases like TQL eRate confirmations. This efficiency boost not only reduces stress but also enables better route planning and potentially allows for an extra load per cycle, directly impacting their earning potential.
• Business: From a business perspective, the automation transforms a 15-20 minute manual process into one that takes mere seconds, translating to significant operational savings. For a mid-sized fleet of 100 trucks, this could mean 25-30 hours saved weekly and approximately $50,000-$75,000 annually in operational costs. The system reduces error rates from 2-5% to less than 0.5%, minimizing billing disputes and payment delays while providing better data quality and compliance tracking. The broader industry benefits from improved broker relations through faster response times and consistent communication, creating a more efficient marketplace.

Conclusion

The integration of LLM technology in document processing represents a significant leap forward in operational efficiency. By combining the power of artificial intelligence with industry-specific knowledge, we're creating systems that not only automate routine tasks but also enhance the quality and reliability of crucial business processes.

As we look to the future, the continued evolution of LLM technology promises even greater improvements in how we handle document processing and business automation. The trucking industry, traditionally run and managed in a highly analog fashion, is proving that with the right application of AI, even the most complex operational challenges can be transformed into streamlined, efficient processes. CloudTrucks is pioneering this transformation and there's more to come, so stay tuned for future updates from our team.

‍