Onboarding a new transaction in asset-based finance is often a time-consuming and complex task. You’re not just dealing with multiple data formats and legal definitions — you’re navigating input from advisors, managing different asset classes, and interpreting hundreds of pages of dense legal language.
Publicly available AI tools aren’t a viable solution either. They often fail to meet data privacy requirements and are prone to hallucinations. At Cardo AI, we’ve built a better way. Our Document Reader tool simplifies the onboarding process by extracting key information directly from credit agreement PDF, .doc, and .docx files — helping you move faster, with more accuracy and confidence.
The models powering this tool are either trained and hosted by Cardo AI on AWS infrastructure or are part of AWS cloud services (like Bedrock), ensuring that your credit agreements and financial data remain confidential and protected at every step.
From document to automation: why we built the Document Reader
- Onboarding takes too long: Understanding and configuring complex deals can take days just to map the relevant events that need to be monitored.
- Junior analyst ramp-up is slow: It takes time for new team members to get up to speed on interpreting legal language and identifying key terms.
- Contracts are complex and interconnected: Many critical terms rely on layered, nested definitions that aren’t easily extracted.
- Manual review is inefficient: Searching across pages to retrieve a single data point is frustrating and error-prone.
Our tool solves all of that. It helps teams save up to 90% of the time typically spent onboarding contracts.
What makes our Document Reader stand out in the market?
Many teams try to use general-purpose AI tools like the ones from OpenAI (or Anthropic, Google, Deepseek, Adobe…) for document analysis, but here are 4 things to take into consideration:
- Model specialization:
Publicly available models are general-purpose. Ours is tailored to extract structured finance information and is trained on real-world credit agreements. - Understanding definitions:
Mainstream models don’t have explicit knowledge of loan agreement definitions. When they face a tough question, they “think harder” with the risk of going off-topic and ultimately hallucinating. Our model “thinks smarter”—by retrieving and using related definitions within the legal document and organizing them with business knowledge, just like a human analyst would. - Fact-checking transparency:
Most publicly available AI interfaces are unable to reference the exact point in a contract where the answer comes. Our reader displays the contract excerpt used in the answer and lets you click to jump directly to the source page in the document. - Hallucinations:
Besides enabling our users to inspect the specific source text, we systematically benchmarked multiple LLMs and LLM providers to select the models with the lowest hallucination rates – those that would not provide answers unless supported by the source documents.
What it does
We built a specific solution that combines SLMs (small language models) and LLMs to get the best performance for the entity extraction task in credit agreements. In a RAG (retrieval augmented generation) setting, we have trained a retrieval model on hundreds of legal documents. Given a user query, the retrieval model extracts the most relevant piece of information from the document.
The pieces of text extracted by the retrieval model are then “read” by an LLM (or generator model), which uses the text as context to answer the user query. This approach allows us to:
- Provide the most relevant context for the question, enabling the LLM to generate a precise answer
- Reduce hallucinations by instructing the LLM to respond using only the pieces of text extracted.
- When the LLM struggles to give a clear answer, it states so by relying on the information passed, pointing out both what is present and what is missing.
Our Document Reader uses a custom-trained AI model to:
- Answer any questions about your credit agreement (e.g. issue date, counterparties, committed amounts…)
- Point to the exact clause used to answer
- Understand and explain nested definitions
- Guide you through follow-up questions
- Extract key information (e.g. counterparties, amounts, dates)
- Automatically highlight and dive deeper into contractual definitions
- Support a more accurate and complete understanding, even for complex queries
Our clients will also soon be able to train custom retrieval models based on the specific semantics of their documents by simply uploading a couple of them and letting the platform do the rest!
How it works
- Upload a credit agreement in PDF, doc* or docx*
- Start document processing, to use the Cardo AI custom language model for document search.
- Get the most relevant elements of the contract (i.e. summarize it in a termsheet)
- Ask a question about the contract (i.e “What is the commitment amount?”):
- Get an answer – plus a clickable link to the clause or definition used (*clickable link available only for PDFs)
- Explore follow-up suggestions, often based on definitions found in the initial answer.
- Drill deeper by increasing the answer’s depth – great for tricky or layered questions.
It’s not only about answering faster – it’s about answering better.
Ready to cut onboarding time by 90%? Try our Document Reader.
Book a demo →