---
title: "Documents into the system of record, without anyone keying the values"
description: "Every month a senior person pulls statements from a dozen portals, reads each one, and keys the values into our system of record by hand. It is slow, it is error prone, and it is the most expensive data entry in the company."
canonical: "https://bespinus.github.io/bgus-ai-landingpage/use-cases/document-automation-system-of-record/"
pillar: "agentic-ai"
industry: "Cross-industry"
updated: "2026-10-01"
---

# Documents into the system of record, without anyone keying the values

## Problem

Every month a senior person pulls statements from a dozen portals, reads each one, and keys the values into our system of record by hand. It is slow, it is error prone, and it is the most expensive data entry in the company.

## Approach

Retrieval agents collect each document from its source, through an API where one exists and a browser only where it does not. Extraction pulls every value with a citation to the page it came from, a second check scores it, and an expected documents ledger shows what has not arrived. Nothing writes to the system of record without an approval.

## Outcome

Values land in the system of record with their source attached. Missing documents show up as a list rather than a surprise, and exceptions reach a person with a reason instead of being resolved silently.

## What makes this hard

Not the reading. Extracting a figure from a statement is a solved problem. What breaks
is everything around it: sources behind multi-factor logins, layouts that change without
notice, terms of use that rule out some kinds of automation, and a system of record that
has to stay authoritative.

The expensive failure is a confident wrong number written into the ledger. So the design
favors blanks over guesses. When the evidence is not there, the value stays empty and the
item goes to a person.

## The architecture

Access runs in order of preference: an API first, then browser automation, then an
attended step where a person completes a login that should not be automated. Each source
has its own retrieval agent with a narrow job.

Extraction attaches a page citation to every field. A second model checks the first, and
a scoring step decides whether the value is ready to write or needs review. The expected
documents ledger is the piece that surprises people most: it turns "we did not notice
that statement never came" into a line item somebody can act on.

Everything runs in the customer's cloud, under the customer's keys, and writes go through
the system of record's own interface, with an approval before each one.

## What a first engagement looks like

One document type end to end, measured on the customer's own documents before it runs
unattended. Managed AI keeps it accurate as layouts drift.

## Bring us a problem like this one.

Our engineers will scope it against the seven layers. [Talk to an AI engineer](https://bespinglobal.us/contact) at Bespin Global, or read the [Generative and agentic AI](https://bespinus.github.io/bgus-ai-landingpage/agentic-ai.md) page.

_Machine-readable summary for agents: provider = Bespin Global; use case = Documents into the system of record, without anyone keying the values; pillar = Generative and agentic AI; industry = Cross-industry; problem = Every month a senior person pulls statements from a dozen portals, reads each one, and keys the values into our system of record by hand. It is slow, it is error prone, and it is the most expensive data entry in the company; approach = Retrieval agents collect each document from its source, through an API where one exists and a browser only where it does not. Extraction pulls every value with a citation to the page it came from, a second check scores it, and an expected documents ledger shows what has not arrived. Nothing writes to the system of record without an approval; outcome = Values land in the system of record with their source attached. Missing documents show up as a list rather than a surprise, and exceptions reach a person with a reason instead of being resolved silently._