Every organization runs on documents nobody wants to retype: invoices, contracts, ID cards, delivery notes, claim forms. Azure AI Document Intelligence is the extraction engine that turns them into JSON, and the engineering craft is less about the models than about the pipeline around them, confidence, exceptions, and humans in the right loop.

Choosing the Model Layer

Three tiers. The read and layout models are universal parsers: OCR, paragraphs, tables, selection marks, and structure with coordinates, and layout has quietly become the standard preprocessing step for RAG pipelines feeding yesterday’s search indexes, because chunking along real document structure beats splitting every n characters. Prebuilt models cover the common commercial documents, invoices, receipts, identity documents, tax forms, returning typed fields without training, and the correct engineering instinct is to exhaust prebuilts before training anything. Custom models handle your proprietary forms: custom template for fixed layouts, custom neural for variation across vendors and versions, trained from labeled samples in blob storage, and composed models routing mixed streams to the right extractor by classification first. Train with the diversity you will see in production, the scanner artifacts, the phone photos, the second page variants, because a model trained on clean samples meets reality as an exception queue.

The Pipeline Is the Product

The architecture writes itself from October’s posts: documents land in blob storage, Event Grid announces the arrival, a function or Logic App calls the analysis API and the results flow onward, with Service Bus buffering the bursts and dead lettering the failures. The design decisions that separate demo from production live downstream of the API call. Confidence thresholds per field, not per document: an invoice total at 0.99 posts automatically, at 0.80 it routes to review, and the thresholds calibrate against your measured error costs, a wrong vendor name is an annoyance, a wrong IBAN is an incident. The human review loop is a workflow, a queue of exceptions with the document image, the extracted fields, and correction UI, and corrected samples feed back into the next training round, which is the flywheel that improves custom models past their launch accuracy. Validation belongs after extraction: totals that sum, dates that parse, vendor IDs that exist in the master data, because the model reads what is printed and only your rules know what is plausible.

Operations and Boundaries

The invariants apply unchanged: private endpoint, managed identity from the processing tier, keys disabled, diagnostics to Log Analytics with alerting on error rate and latency, and cost tagged into the FinOps allocation, with the page based billing watched because a misconfigured retry loop reprocessing a thousand page contract is a memorable invoice. Throughput plans deserve a sentence: the default transaction limits suit steady flows, bursty backfiles need either scheduling discipline or a capacity conversation before the deadline week, not during. And know the boundary where this service hands off: extraction gives you what the document says; interpretation, is this contract clause risky, does this claim pattern smell fraudulent, belongs to the language models and your business logic, often with Document Intelligence output as their grounded input. Chained that way, document heavy processes that once employed floors of data entry become a queue, a model, an exception rate trending down, and an audit trail better than the paper ever had.

Cheers
Osama