The answers were in the documents, but finding them took people. Every application meant reading the whole pack, then checking one document against another by hand.
Scanned digital documents in English & Dutch, running to hundreds of pages per case.
The same loan or person appeared in many places, sometimes with conflicting values.
Confirming a loan was really closed meant matching bank statements line by line.
Results lived in notes and spreadsheets, then had to be re-keyed into downstream systems.
Early pilots had no way to budget AI spend as document volume grew.
An AI document service delivered through an API. A pack goes in, Aether runs one orchestrated job across two specialised agents, and a single rule-vetted master JSON comes out.
Aether creates a run for each pack and holds its state. It splits large packs into parallel jobs, assigns token budgets and routes each task to the right model, with retries built in.
Splits the pack into documents and identifies each type. It extracts fields , tables , transactions and registry codings from English and Dutch documents, scanned or digital, and links people, accounts, loans and dates across the pack, surfacing any conflicting values .
Applies business rules to the correlated data, checking loan closures against actual statements, BKR registry codings against their rules and marital status across records. It then scores the risk and writes a schema-valid master JSON.
Rules for loan closures, credit registry, civil status, income and identity are configurable and versioned, so the same checks run the same way on every application.
The master JSON goes straight to under writing and downstream systems, with the source and rule result behind every field. Analysts only look at what was flagged .
Analysts moved from reading every page of every pack to reviewing only the exceptions the agents flag, with structured data flowing straight into underwriting.
Reads Dutch and English documents as they are , such as a scanned BKR Kredietoverzicht, and returns structured fields wit h English keys.
Each rule result points to its source. A loan closure passes only when the final repayment appears in the bank statement on the recorded end date.
Every pack gets a verdict, score and risk level. Cases such as recovered arrears still visible in the registry are routed for manual mortgage review.
Aether routes each task to a suitable model, caps tokens per agent and reports the cost of every pack, so spend stays predictable as volume grows.
Aether Orchestrator
Extraction Agent
Validation Agent
Versioned Rules Library
Model Routing
Token Budgets
Document Service API
Master JSON Schema
