Building a Custom Document Analyzer in Three Steps

▶ Watch (06:23)

Azure AI Content Understanding builds analyzers through a three-step flow: define a schema, test against sample files, then deploy a REST endpoint. Vinod Kurpad walked through the invoice analyzer demo (06:23) live, starting with the pre-built invoice template and extending it by adding a custom email address field. Extraction methods go beyond simple lookup: fields can be extracted, generated (summaries), or classified (hazard flags on line items). Every output includes a confidence score, which lets teams set a threshold below which a human reviews the value instead of passing it downstream automatically.

Multi-File Input and Reasoning Across Documents

▶ Watch (15:19)

Standard extraction sends one document at a time. Pro mode takes multiple documents in a single request and reasons across them. In the invoice-PO mismatch demo (15:19), Kurpad loaded an invoice and a matching purchase order as two separate PDFs. A “PO match” field on each line item asked the model to compare quantities semantically, not just as string equality. The model flagged one item where the invoice listed two units and the PO listed four, with a plain-language explanation of the inconsistency.

Reference Data: Grounding Analysis in Organizational Knowledge

▶ Watch (23:10)

Multi-file input compares documents against each other. Reference data goes further: it feeds static organizational knowledge, such as contracts or SOPs, into the same analysis pass. In the contract reference data demo (24:32), Kurpad added four to five contract documents alongside the invoice and PO. The model caught a payment-term risk: the contract required milestone-based billing, but the invoice demanded the full amount upfront.

“You’re getting invoiced for that full 29,900 right up front.” — Vinod Kurpad

Neither discrepancy is visible from the invoice alone.

Compliance Recording at ASC Technologies

▶ Watch (35:05)

ASC Technologies records and analyzes communications for financial services firms across 50+ countries. Without AI, compliance teams could manually review only 3% of calls,

“leaving 97% compliance risk.” — Tobias Fengler

After integrating Content Understanding into Recording Insights, ASC replaced six separate OpenAI services with one, cut R&D effort by 30%, and now monitors 100% of Microsoft 365 communications at five times the previous speed. Jannik Adam showed how analysts configure extraction fields and policies through a no-code interface, with Content Understanding creating the analyzer templates silently in the background.

Sports Highlight Generation at MediaKind

▶ Watch (51:15)

MediaKind’s MK.IO platform reaches 2 billion viewers daily. Traditional sports highlight workflows relied on manual tagging, taking hours or days per event. After embedding Content Understanding in the MK.IO AI pipeline, the system detects goals, fouls, player identities, and crowd reactions as the game runs, then packages clips into personalized reels,

“or play types delivered seconds after the live action.” — Feras Siam

In the MediaKind sports highlight demo (57:15), Siam queried by team (Real Madrid) and player (Mbappe), assembled custom reels on the spot, and inserted personalized ads between clips. The pipeline cuts production time by up to 80%.

Notable Quotes

extraction is not what you’re trying to do. Vinod Kurpad · ▶ 13:25

You’re getting invoiced for that full 29,900 right up front. Vinod Kurpad · ▶ 30:44

leaving 97% compliance risk. Tobias Fengler · ▶ 37:30

or play types delivered seconds after the live action. Feras Siam · ▶ 56:50

Key Takeaways

  • Pro mode lets a single API call reason across multiple documents and a reference knowledge base simultaneously.
  • ASC replaced six OpenAI services with Content Understanding, cutting R&D effort 30% and speeding results five times.
  • MediaKind’s MK.IO cuts highlight reel production by 80%, delivering fan-personalized reels seconds after live action.