An Agent Needs a Clean Document: Why Extraction Matters More Than the Model

Microsoft puts it plainly: large language models have existed for five years; during that time, the models have become stronger, while the content has become messier.

The authors estimate that the main engineering task is to turn a raw PDF, image, or audio file into verifiable information that the agent can act on.

Content extraction is the layer between the file and the model.

It recognizes text, reconstructs the structure (headings, columns, and tables), extracts fields, and records where each value appears in the source.

The PDF specification alone, ISO 32000-2, spans about a thousand pages. A price list looks like a neat table on screen, but inside the file it is stored as a set of separate text fragments with coordinates.

That same week, Microsoft released version dotnet-1.23.0 of its Agent Framework.

Frequent releases show that vendors have made agent orchestration a standard library for C# / .NET and Python.

A team can build an agent in a day, but it cannot teach it to read company documents correctly in that time.

Consider a typical scenario: an agent receives a supplier invoice or specification and enters the items into 1C.

Errors can occur at the input stage: - in a table with merged cells, the parser mixed up the row order, assigning one item's price to its neighbor; - in a scan, a stamp obscured a digit, and OCR read "8" as "3"; - in an Excel export, the unit of measure appears in the column heading, so "12" without context could mean either pieces or boxes.

A model's reasoning error is usually immediately apparent.

A poorly extracted document produces a plausible answer with an incorrect figure, and the agent confidently enters it into the accounting system.

The result is an incorrect storefront price, inventory discrepancies, and manual reconciliation after the fact.

After a couple of such incidents, the CFO shuts down the pilot, and it is hard to blame them.

### Structure analysis First, the system reconstructs the page layout: reading order, columns, tables, and captions.

For scans, OCR is added to the process.

The step's goal is for a 40-row table to remain a 40-row table with the same relationships between cells. ### Normalization

The system then maps values to the company's reference data: SKUs to the item master, units of measure to base units, and dates to a single format.

This requires master data, so the step's outcome depends on the quality of PIM and master data. ### Linking to

Five years after the start of the LLM era, Microsoft is returning to PDF parsing, and that is an honest signal to the market.

Vendors are quickly turning models and agent frameworks into standard components.

The advantage will go to the company whose agent reads an invoice as accurately as an experienced accountant and shows exactly where it found each figure in the document.

If an AI project's budget goes toward model selection and nothing remains for document analysis, the pilot will end with incorrect figures in the accounting system.