Microsoft has added an agentic mode to Azure Content Understanding for document tasks that cannot be solved reliably in one extraction pass.
The preview wraps a language model in a loop with document-specific tools. It can navigate text and tables, inspect visual evidence, perform calculations and revisit earlier evidence before returning structured fields defined by a schema. Microsoft demonstrates the approach on a 137-page U.S. Treasury bulletin where values span pages and require exclusions, normalization and arithmetic.
Standard mode remains the default for extracting items such as invoice totals, policy numbers or signature dates. It reads analyzed content once and fills requested fields, which makes it faster, more predictable and less expensive. Agentic mode spends additional model tokens and compute on tasks involving cross-document synthesis, reconciliation, exhaustive review or complicated visual structures.
The new mode does not remove the need to test accuracy. Microsoft explicitly frames the two options as different points on a price-performance curve, not a replacement of the old system. Its guidance is to begin with standard mode and move difficult workloads to the agentic path only when the added reasoning materially improves results.