Microsoft adds agentic mode to Azure Content Understanding for multi-pass document analysis
Azure Content Understanding now has an agentic preview mode that lets a language model reason iteratively over a document—navigating tables, checking visual evidence, and recomputing—for extraction tasks a single pass cannot reliably handle. In Microsoft's example, it calculates a U.S. Treasury monthly outlay figure from a 137-page bulletin by tracking a table across a page break and applying exclusions. The mode uses more tokens and takes longer, so Microsoft positions it alongside standard mode on a price-performance curve.

Microsoft adds agentic mode to Azure Content Understanding for multi-pass document analysis
Microsoft has introduced agentic mode, a preview feature in Azure Content Understanding within Foundry Tools, for document-analysis tasks that cannot be completed in a single efficient pass. The mode wraps a language model in a reasoning loop with document-specific tools, allowing it to navigate text and tables, inspect visual evidence, calculate, and re-examine material when needed.
Standard mode, the default, reads a document once, applies a user-defined schema, and returns structured fields such as invoice totals or policy numbers. Microsoft recommends it as the starting point for most workloads; it is faster, more predictable, and cheaper. Agentic mode is intended for cases where scattered evidence, exclusions, arithmetic, or complex layouts make a single pass unreliable.
Microsoft's Principal Product Manager illustrates this with a U.S. Treasury example. The task is to calculate January 2003 agency outlays from a 137-page bulletin, with specific agencies included and entries like receipts excluded. The relevant table spans two pages, and values must remain tied to their column headings across the page break. A revision mark must be stripped from a value while a minus sign must be kept. The agent selects, normalises, and totals the numbers, producing 180,681 million dollars.
In the Treasury case, the agent plans the task, uses document tools to track down the right table, validates selected values, and runs a final calculation. A single model pass would have to handle page relationships, notation, exclusions, and arithmetic simultaneously—the failure mode agentic mode addresses.
Two further examples support the pattern. Reading four economic charts in a 20-page excerpt requires interpreting subtitles, axes, and surrounding prose rather than merely identifying plotted values. The agent can revisit the source when a chart is ambiguous. In a network log, a table continues across 19 pages; the agent must consolidate the pages into one table, ignore repeated headings, and handle conditions in the right order. Inferring event count from first and last event IDs fails because the IDs contain gaps.
Agentic mode consumes more model tokens and typically takes longer for the same document. Existing analyzers can be tested against it in Content Understanding Studio, with API version 2026-06-01-preview and "workflow": "agentic" in the analyzer configuration.
Microsoft frames the two modes as points on a price-performance curve rather than as an old-and-new replacement, and recommends comparing both on representative files. Whether the improvement in accuracy on difficult documents justifies the added cost and latency remains for prospective users to evaluate.
Read the original at devblogs.microsoft.com →