Skip to content

Assistant Document Analysis

Thorsten Sommer edited this page Oct 31, 2026 · 2 revisions

Deutsch: Assistent Dokumentenanalyse

The "Document Analysis" assistant analyzes documents along rules you define once and reuse: for example, checking applications against the criteria of a call, extracting the key figures of reports, or reviewing contracts for certain clauses. Such a set of rules is called a policy.

Policies

A policy describes how the AI analyzes the documents and how it presents the result. You can keep as many policies as you like and switch between them. AI Studio saves every change at once.

Under "Document analysis policies":

  • "Add policy" creates a new policy. Click it in the list to edit it.
  • "Import policy" adds a policy that a colleague shared with you as text.
  • "Delete this policy" removes the selected policy after asking you.

Policies with a building icon come from your organization. You can use them, but not change or delete them. Some don't even show their rules, only their description.

Setting up a policy

Expand "Policy definition" and fill in:

  • "Policy name": 6 to 60 characters, unique among your policies.
  • "Policy description": 32 to 512 characters about what the policy does.
  • "Select a minimum confidence level": only providers at this level or higher can analyze the documents.
  • "Tools this policy permits": tools the AI may use during the analysis, for example, Web Search.
  • "Preselected provider" and "Preselect a profile": what the analysis starts with.
  • "Analysis rules": how the AI should analyze the documents, for example, which criteria to check.
  • "Output rules": how the result should look, for example, a table with one row per criterion, and in which language.

Both rule fields can also be loaded from a document.

"Would you like to protect this policy so that you cannot accidentally edit or delete it?" locks the policy against changes. You can still choose its profile.

Analyzing documents

  1. Open "Assistants" → "Document Analysis", and click a policy in the list.
  2. Expand "Document selection", and add the documents under "Documents for the analysis", with "Add file" or by dropping them. Audio and video files are transcribed first; images go to the model if it can read them.
  3. Choose a provider, and click "Analyze the documents based on your chosen policy".

All documents of a run go to the model together, so it can compare them. The model sticks to the documents and the policy. When information is missing and the output rules say nothing about it, the answer starts with INSUFFICIENT_INFORMATION and lists what's missing.

The result

The analysis appears below the form. "Export result" suggests the name of the policy as the file name. "Send to ..." → "New Chat" opens a chat with your documents and the result, so you can ask follow-up questions.

Tips

  • Write the rules as precisely as you would brief a colleague. Name the criteria and the format you expect.
  • Test a new policy with a few documents and refine the rules before you use it widely.
  • All documents together must fit into what the model can handle at once. For many documents, use Batch Processing, which analyzes each document on its own with the same policy.

For administrators

With "Show administration settings?" switched on in "Settings" → "App Options", "Export policy as configuration section" copies the policy, so that your IT department can roll it out to everyone. "Hide the policy definition when distributed via configuration plugin?" then hides the rules from your colleagues. See Assistants and Templates.

Where to go next

Clone this wiki locally