× STEPS
Case study | Generative AI and legal documents, Afreximbank

Loan Term Agreements, AI that reads loan contracts

For Afreximbank, STEPS built an AI platform that reads loan agreements, automatically extracts their covenants and turns them into structured, queryable data.

Generative AI
Document intelligence
12 months of delivery
01 - The context

Commitments buried in hundreds of pages

Every loan agreement contains covenants: financial ratios to maintain, documents to deliver, operations requiring consent. Tracking them means knowing them, contract by contract.

!

Long contractsLoan agreements often run to dozens of pages, with schedules and amendments.

!

Manual readingLawyers and monitoring teams used to record clauses by hand in spreadsheets.

!

Hard to monitorWithout an overview, a deadline or a threshold can go unnoticed.

!

A compliance riskAn unmonitored covenant is an undetected credit risk.

02 - The principle

From contract to data, in three steps

Dozens of pages per contract, scattered obligations: the platform makes covenants visible and trackable.

INGESTIONThe contracts

Loan agreements are uploaded to the platform, as PDF or Word, with their related facility.

EXTRACTIONThe clauses

AI identifies covenants, conditions and obligations, and classifies them by type.

STRUCTURINGThe data

Each clause becomes structured data that teams can filter and query.

03 - The platform

Upload, extract, review

Lawyers upload a contract, follow the extraction and find each covenant with its original excerpt. Legal review focuses on validation.

12months of delivery, from design to go-live
04 - The results

All covenants, on a single screen

Each contract produces a results sheet: covenants by category, processing status and direct access to the source text.

05 - The benefits

What the platform changes for the bank

Covenants become monitored data, no longer text to reread.

Main benefitReliable covenant monitoring

The bank has a structured view of all its borrowers' commitments, strengthening credit risk monitoring throughout the life of the loan.

TimeFaster legal review

Lawyers validate an extraction instead of rereading everything.

CompletenessNo clause missed

All obligations are identified and classified automatically.

SearchQueryable data

Finding every contract with a DSCR ratio takes seconds.

EvidenceThe source text at hand

Each covenant is linked to the contract excerpt it comes from.

RiskAlerts made possible

Structured data feeds deadline and threshold monitoring.

ScaleA whole portfolio

The same method applies to hundreds of contracts.

06 - Our work

12 months, from design to go-live

PHASE 01Scoping

Covenant typology, contract formats and needs of the legal and credit teams.

PHASE 02Models

Design of generative AI extraction chains and tests on real contracts.

PHASE 03Platform

Development of the ingestion, results and search interface.

PHASE 04Go-live

Validation with the lawyers, adjustments and production deployment.

07 - Architecture

Under the hood

A document intelligence chain combining file reading, generative AI and a structured database.

Generative AILLMNLPOCRPythonAPI
  • InputPDF and Word contracts, with their related facility
  • ReadingText extraction and splitting by clause
  • AnalysisGenerative AI that identifies and classifies covenants
  • StorageStructured database, filterable and queryable
  • InterfaceWeb platform for ingestion, review and search
FAQ

Frequently asked questions

Which documents does the platform process?

Loan agreements and their amendments, as PDF or Word.

Can what the AI extracted be checked?

Yes. Each covenant is linked to the contract excerpt it comes from, for quick validation by lawyers.

Can it be extended to other contracts?

Yes. The same principle applies to guarantees, financing agreements or supplier contracts.

Do your teams still read contracts line by line? Let's talk about your document AI.

Book a scoping session ↗