Read a sample: 1. The Extraction Problem and the Worked Project
A business document can be easy for a person to understand and awkward for software to use. An invoice may clearly show the supplier, invoice number, date, and amount due, yet those details may appear in different places on every supplier’s document. A form can present the same question as a typed box, a handwritten answer, or a selected checkbox. Before another system can use those answers, it needs to find them, interpret them, and represent them consistently.
Document extraction addresses that conversion. It takes information from a document and returns a structured representation that software and people can inspect. Instead of requiring someone to retype an invoice total, an extraction workflow might return a field named InvoiceTotal with a value and supporting information about where it was found.
That output is useful, but it is not the same thing as a business decision. A correctly extracted total does not prove that the invoice is legitimate, that the goods were received, or that a payment should be authorized. A plausible-looking answer can also be wrong. The central challenge of this book is therefore not simply to get a model to produce values. It is to build a workflow that can use those values responsibly, identify uncertainty, and learn from errors.
This chapter introduces the fictional project used throughout the book. You will see what the organization wants to extract, who will use the results, and how the team will decide whether the workflow is useful. You will also see how the practical work connects to certification preparation without treating the project as a substitute for current exam objectives.
Northstar Facilities is a fictional company that maintains office buildings. Its small operations team receives invoices from suppliers and internal maintenance request forms from building staff. Today, employees read the documents and enter selected details into separate tracking systems. The work is repetitive, but it also requires judgment: a blurry date, an unfamiliar invoice layout, or an unchecked box can change what an employee should enter.
Northstar wants to test whether Azure AI Document Intelligence can reduce some of the manual transcription. The initial project is deliberately narrow. It will extract a defined set of fields from fictional invoices and maintenance request forms, route questionable results for human review, and compare outcomes with expected answers.
The project does not begin by paying invoices, approving requests, or updating production systems. Those actions would require additional business rules, integrations, permissions, and controls. The prototype’s first job is to produce and evaluate structured candidate data.
Two document streams enter the same general workflow:
- Supplier invoices contain details such as the supplier name, invoice number, invoice date, and total amount.
- Maintenance request forms contain details such as the request identifier, building, reported issue, and urgency selection.
These streams share a need for document intake, extraction, review, and measurement. They do not necessarily share the same model, fields, or review rules. Treating them as one uniform document type would hide important differences.
For example, an invoice from Harbor Office Supply might present an invoice number near the top and a total at the bottom. A different supplier might use a compact table and label the amount as “Balance Due.” A maintenance request form may have a fixed heading for “Building,” while another version uses “Site.” Some requests may contain handwriting or no urgency selection at all. A useful test set must reflect more than the cleanest example.
All names, documents, and values in the project are fictional. Using invented samples lets us concentrate on the extraction process without placing real supplier or employee information into a practice environment.
Northstar’s project has several users, each with a different view of success.
A document intake clerk needs to submit a file and understand whether it has been processed or requires attention. The clerk should not have to interpret a raw service response to discover that a document failed.
A reviewer needs to see the proposed values alongside the source document. When a value is uncertain, missing, or inconsistent with a rule, the reviewer needs enough context to confirm or correct it. A reviewer may also decide that the document itself is unreadable or that a field cannot be determined.
An operations lead needs to know whether the workflow reduces transcription work without allowing unacceptable errors to pass through. A high volume of extracted fields is not, by itself, evidence that the process is working.
A developer or AI practitioner needs to select an appropriate analysis approach, handle responses and failures, protect access to the service, and record results in a form that can be tested. That person also needs to understand which errors originate in extraction and which arise later, such as an incorrect normalization rule or a reviewer’s correction.
These roles may overlap in a small prototype. Naming them separately is still useful because it exposes competing requirements. A reviewer may want every document displayed for inspection, while an operations lead may want fewer items in the queue. The appropriate balance depends on the field, the observed errors, and the consequences of a mistake—not on a confidence score alone.
Before running a model, Northstar needs to define what it wants back. “Extract the invoice” sounds straightforward until two people disagree about which date or total belongs in the result. A field list turns a broad ambition into something the team can test.
For the initial invoice…