In a 9 November 2026 Microsoft Foundry Blog post, Microsoft described agentic mode in Azure Content Understanding, a preview workflow for document analysis when the answer must be built from evidence across a file. The feature keeps the analyzer schema and structured output model, but adds a reasoning loop that can navigate text and tables, inspect visual evidence, calculate and validate results.
Azure Content Understanding normally reads content with a schema in one pass and fills extracted or generated fields. Microsoft says that remains the right design for most documents because it is faster, more predictable and lower cost than agentic mode.
What agentic mode adds
Agentic mode is aimed at document tasks where the value is not located in one place. The service can connect evidence from different pages, apply field instructions, normalise selected values, perform arithmetic and check the result before returning fields that match the schema.
Microsoft uses a U.S. Treasury example to show the difference. The task asks for January 2003 agency outlays with specific exclusions, using a table in a 137-page bulletin that spans two pages. The workflow has to follow the January row across column bands, keep each amount tied to its agency heading, remove a revision marker, retain a minus sign where it changes the value, and calculate a final result of 180,681 million dollars.
That is not the same kind of extraction as finding an invoice total or a signature date. A single pass has to handle page relationships, exclusions, notation and arithmetic at once. Agentic mode separates that work into steps and uses document-specific tools to revisit evidence where needed.
Where the preview may help
The preview is most relevant where standard extraction fails to meet accuracy requirements on representative files. Microsoft lists multistep reasoning, calculations, validation, visual analysis and structured output as the main use cases.
The announcement gives two further examples. One reads values and dates from four economic charts in a 20-page excerpt and returns seven fields, using axes, subtitles and surrounding prose to interpret the plotted values. Another consolidates a network log table that continues across 19 pages, filters rows by event type and sums durations. In that example, using first and last event IDs as a shortcut would produce the wrong count because the IDs have gaps.
These examples matter for teams building document automation around financial reports, operational logs, contracts, insurance records or other long files where the answer depends on selection rules. They also show why the feature is not a general replacement for standard extraction.
Preview limits and production boundary
Azure Content Understanding is generally available with the 2025-11-01 API version. Agentic mode is available through API version 2026-06-01-preview, which Microsoft identifies as a public preview without a service-level agreement and not recommended for production workloads.
The initial preview supports one input file per analysis request. That file can contain one document or multiple logically related documents, such as a contract and appendix. Agentic mode supports document analyzers only, does not support fields that use the extract method, and does not support using labelled samples to improve the analyzer.
Microsoft also states that agentic mode is not a replacement for human review in high-impact scenarios. That constraint is relevant where automated results affect regulated decisions, financial outcomes or legal obligations.
Cost, latency and capacity
Agentic mode uses more model tokens than a standard workflow and typically takes longer for the same document. Microsoft says teams should compare both modes on representative files and use agentic mode only where the quality gain justifies the additional cost and latency.
The resolved agentic.* workflow uses the advanced contextualization rate. Microsoft also notes that agentic mode typically requires about 400,000 tokens per minute per analyzer job on the Foundry model deployment, and recommends configuring at least 400,000 tokens per minute capacity to reduce the risk of 429 rate-limit errors.
Pricing for agentic use is therefore not just a feature switch question. It depends on the analyzer, the document shape, the chosen Foundry model deployment, contextualization charges, token use and the latency tolerance of the downstream process.
Configuration path
In Content Understanding Studio, the preview path starts with a new project using API version 2026-06-01 (Preview). Under Workflow kind, users choose Agentic Extraction, define the fields to return, and test against their own documents.
In code, the analyzer definition uses API version 2026-06-01-preview and sets config.workflow to "agentic". After creation, the service returns a resolved versioned value such as "agentic.2026-06-01-preview". Microsoft also says the completion model value should use the supported model’s catalog name, not its deployment name.
Teams migrating from pro mode need to review their analyzer and field schema rather than treating the change as a direct replacement. Pro mode was available only in the 2025-05-01-preview API, which Microsoft says is now retired. Workflows that depended on multiple input documents or reference data need particular checking against the agentic mode limits.
Document analyzers that plan, inspect charts, calculate and validate fields raise the same design questions that ECS's Core AI topic covers across agentic AI, Azure AI Foundry, the EU AI Act and responsible AI. Sessions and workshops give attendees a place to compare preview features, model capacity, grounding, review thresholds and operational controls with people building similar systems. Bring representative document scenarios and the cost, latency and governance questions that decide whether agentic extraction should stay in evaluation or move into a service design.
This article was written by the ECS editorial team with the assistance of Amal Agent, run.events' AI assistant, which supported the research, drafting and proofreading.