Agentic Mode: When the Answer Is Not on the Page

The gap it fills

Ordinary field extraction assumes the answer exists somewhere in the content and the work is finding it. A great deal of real document work violates that assumption. Whether an invoice reconciles against its contract, whether a policy meets a checklist of conditions, what a total should be given figures scattered across three tables, how an amendment changes the clause it amends — none of these values is printed anywhere. They have to be constructed.

Agentic mode is the option for that class of problem. It reasons across a document, performs calculations, validates results against conditions, interprets visual elements such as complex tables and figures, and still returns fields that conform to your schema. The output contract does not change; the amount of thinking behind each value does.

Turning it on

Agentic mode is a workflow selection made when you create a document analyzer, using the preview API version. The analyzer config carries a workflow property. Leave it out, or set it to the default, and the service picks a standard or advanced workflow based on how the analyzer is configured. Set it to agentic and you get agentic mode.

There is a subtlety in the response worth understanding, because it explains confusing round-trips. What you send is a creation-time selector; what comes back when you read the analyzer is a resolved, versioned value that names the workflow family and the API version. The value you wrote and the value you read are deliberately different.

The workflow family also determines the billing rate for the contextualisation layer. Standard families bill at the standard rate; agentic and advanced families bill at the advanced rate.

What it costs

This is not a switch to flip out of curiosity. Three costs stack up.

Contextualisation is billed at the advanced rate. Agentic analysis consumes more model tokens than a standard workflow for the same document. And it takes longer, sometimes considerably.

There is also a capacity requirement that catches teams in testing: an agentic analyzer job needs a substantial token-per-minute allowance on the connected model deployment — the guidance names roughly four hundred thousand tokens per minute per job — and without that headroom you meet rate-limit errors rather than results.

The honest way to decide is the way the documentation frames it: run representative documents through both paths and see whether the quality improvement is worth the latency and the money. That is an experiment, not a preference.

What it cannot do yet

The initial preview has real boundaries, and they are examinable precisely because they are surprising.

It is document-only. There is no agentic mode for audio, video or images.

Each analysis request takes one input file. That file may contain several logically related documents — a contract with its appendix and the governing rules, say — and reasoning across those is exactly the intended use. But you cannot pass three separate files and ask for a comparison.

Fields using the extract method are not supported. Since agentic mode exists to construct answers rather than locate them, this is coherent rather than arbitrary, but it means an existing schema full of extract fields cannot simply be pointed at agentic mode.

Improving the analyzer with labeled samples is not supported either.

The retired predecessor

If you find older material describing pro mode, note what happened to it. Pro mode lived only in an earlier preview API version that is now retired, and the replacement is agentic mode in the current preview. The documentation is explicit that this is not a drop-in swap: the configuration is different and both the analyzer and its field schema should be reviewed against the agentic documentation before migrating. Anyone carrying notes from that era needs to update them rather than translate them.

Judgement, which is the actual exam skill

Two boundaries decide most scenario questions.

The first is downward. For straightforward field extraction, use a standard document analyzer. Agentic mode applied to a receipt is slower and more expensive for no benefit. The trigger is reasoning: is a calculation, a cross-reference or a consistency check required to produce the value?

The second is upward. Agentic mode is not a replacement for human review in high-impact scenarios — the documentation says so plainly. Higher reasoning effort narrows the error rate; it does not remove accountability. A system that approves loans or pays invoices without a human in the loop is not made safe by choosing the most capable workflow.

Between those boundaries the mental model is simple. Standard extraction answers where is this value. Agentic mode answers what follows from these values, and returns the answer in the same typed shape as everything else.