Relationship managers at a corporate bank were reading 50 or more contracts, amendments, guarantees and security agreements before every client meeting. The folder for a single client grows over years. Terms get amended, facilities get added, collateral gets swapped. Before a meeting, the manager had to reconstruct the current position from the whole pile.
The process was slow, and it was risky. A missed amendment, an outdated margin or a misquoted limit could surface in front of the client. Nothing in the folder told the manager which document held the current answer.
We deployed EigenPal inside the bank's environment, on an on-prem model, and turned the whole folder into one consolidated, cited summary. That summary now powers a chatbot the managers use to prepare for meetings.
What the workflow does
The manager drops every client document into one folder, in any format. From there the workflow runs without human interaction.
1. Classify each document
Every file is parsed and tagged by the information it contains: facility details, interest and fees, maturity and repayment schedules, covenants, ongoing obligations, collateral and guarantors, amendments.
2. Consolidate in parallel
Matching topics are merged across all documents. Amendments are ordered chronologically and conflicts are resolved facility by facility, so the summary reflects the terms currently in force rather than the terms first signed. Gaps in the data are flagged automatically instead of being papered over.
3. Build the cited summary
The output is one JSON document per client: a document tree, the merged terms for every facility, ongoing obligations and a list of data gaps. Every leaf value carries its own citation in the form {value, source_document, source_page, source_paragraph}. The schema is audit-ready, not just readable.
4. Answer questions
The summary feeds a chatbot. A manager asks a plain-language question and gets a structured answer with the source attached. Typical questions:
- "Prepare me for the meeting."
- "What is our collateral position?"
- "What are the main risks for this client?"
The same summary answers different kinds of questions. One surfaces facility-by-facility deal terms for the meeting. Another surfaces cross-contract risk patterns. Every figure in the answer, whether a margin, a limit or a risk driver, links back to the exact contract, clause and page.
Designing the taxonomy was the hard part
With 50 documents per client, extraction was not the difficult step. The difficult step was deciding what the summary should contain.
We worked with the bank's team to design the taxonomy itself. The result was 17 output fields covering client identity, every facility's terms (interest, maturity, covenants, collateral) and a set of quick-answer fields built specifically for what the chatbot needs to return. Designing those fields up front is what makes the answers consistent from one client to the next.
A proof point on covenants
Before this project, covenants were found one-off by risk teams, re-reading the contracts every time they needed them.
From just 10 uploaded documents, our builder agent learned the bank's definition of a covenant, drafted the extraction workflow, ran it on the examples and reported its accuracy back. The workflow follows the bank's existing business logic, took minimal engineering time and little of the business owner's time, and passed the bank's compliance review. Covenants now flow straight into the CRM.
Why it holds up in a bank
Every answer is cited. Nothing is taken on faith. The manager can open the source page for any value in seconds.
Amendments are resolved, not listed. The summary states the current terms and shows the chain of changes that produced them.
Data gaps are explicit. If the folder does not contain the answer, the summary says so rather than guessing.
It runs inside the bank. Documents never leave the bank's environment. The model runs on-prem.
Results
- 70% reduction in meeting preparation time for relationship managers
- 50+ files per client sorted, classified and summarized with no human interaction
- One consolidated JSON per client, every value cited to document, page and clause
- Covenant extraction built from 10 example documents and passed through compliance review