SCRIBE, AI agents that draft credit documents
For Afreximbank, STEPS designed SCRIBE: a supervised multi-agent system that draws on the bank’s systems of record and drafts term sheets, PAMs and credit memoranda in the drafting officer’s style.
SUPERVISED AGENTS
Credit, a business of documents
Every financing granted by the bank goes through a series of standardised documents, drafted by credit officers from data scattered across several systems. Long, repetitive and yet critical work.
Scattered dataDeals, obligors and facilities are stored in different systems that must be checked one by one.
Manual draftingEach term sheet, PAM or memo often starts from an old document, copied and then adapted by hand.
A style to respectEach officer has their own way of writing, and committees expect consistent quality.
A risk of errorA miscopied figure can distort a credit decision.
Hours of drafting before every committee
A credit file draws on scattered data and a demanding drafting style. SCRIBE prepares the first draft so the officer can focus on the analysis.
Systems of record to cross-check: deals, obligors, facilities.
Document types drafted: term sheet, PAM and credit memo.
Of drafts reviewed and approved by a credit officer before use.
Three documents, a first draft in minutes
Term sheet
The key terms of the facility, structured and ready to negotiate.
PAM
The project approval memo, with context, risks and recommendation.
Credit memo
The full analysis of the file, written in the officer's style.
Four agents, one officer in control
The agents work in a chain, each with a precise role. The credit officer reviews, corrects and approves every draft: AI prepares, humans decide.
RetrieveAn agent retrieves deal, obligor and facility data from the bank's systems.
DraftAn agent drafts the document from this data and the bank's templates.
Match styleAn agent adapts the text to the drafting officer's style, based on their past documents.
VerifyAn agent checks every fact against its source before delivering the draft.
What SCRIBE changes for the bank
The gain is not only measured in time: SCRIBE also improves the quality, consistency and traceability of credit files.
Credit officers no longer start from a blank page: they review a complete first draft and spend their time on credit judgement, where their expertise matters.
The first draft is ready in minutes, shortening the path to committee.
Data comes straight from the systems of record, with no copy and paste.
Documents follow the bank's templates and each officer's style.
An auditor can trace the origin of every piece of data cited in a document.
The EKC layer organises the bank's credit knowledge and makes it reusable.
Nothing is sent without a credit officer's approval.
A proof of concept in four steps
Choice of target documents, data sources and success criteria with the credit teams.
Building the EKC layer: connection to the systems of record and indexing of past documents.
Development of the four agents, document templates and style learning.
Tests on real files, review by officers and measurement of the gain before scaling up.
AI designed for a bank
The EKC knowledge layer links every fact in the draft to its source in the bank's systems.
The system is designed to be controlled: scope, logging and human review at every step.
A time-boxed project to measure the real gain before any scale-up.
Under the hood
An orchestrated multi-agent architecture, built on retrieval-augmented generation (RAG) over the bank's knowledge.
- SourcesSystems of record for deals, obligors and facilities
- KnowledgeEKC layer: indexed documents, facts linked to their source
- AgentsRetrieve, Draft, Match style, Verify, orchestrated in a chain
- ControlHuman review, step logging, model risk management
- OutputsTerm sheets, PAMs and credit memos in the bank's format
Frequently asked questions
Does SCRIBE replace the credit officer?
No. It prepares a first draft; the officer reviews, corrects and approves every document.
How are AI errors avoided?
A dedicated agent checks every fact against its source, and the EKC layer keeps track of where every piece of data comes from.
Can it be adapted to other documents?
Yes. The multi-agent architecture extends to other standardised documents: risk reports, committee notes, annual reports.
Do your teams spend hours drafting standardised documents? Let's talk about your AI agents.
Book a scoping session ↗