Document Intelligence

Document Intelligence is Novantra’s AI capability for understanding documents: extracting structured fields from forms, classifying documents by type, summarizing long documents, identifying parties and obligations mentioned in contracts. What it does for any given document is shaped by a document profile — the configured handling for a class of documents.

Like Copilot profiles, document profiles have versions, prompts, retrieval scope, and constraints. Document Intelligence outputs are never silently applied; they appear as suggestions that a reviewer accepts, modifies, or rejects before they affect the workspace.

When you would reach for this

You configure Document Intelligence when:

• A class of documents needs structured extraction (contracts, invoices, certificates, regulatory filings).
• Documents need automatic classification or routing on upload.
• Long documents need summarization to support reviewer workflows.
• Document metadata (parties, dates, obligations) needs to be extracted and proposed for application to the document record.

You don’t reach for this for documents the team handles manually, or for one-off document review. Document Intelligence is for recurring document classes where automation creates leverage.

What lives in a document profile

Two record types:

Document profile is the named handling for a class of documents (“Supplier contract intake,” “Insurance certificate processing,” “Regulatory notice classification”). It carries title, key, kind, owner, and the high-level purpose.

Document profile version is a specific revision: prompts and extraction schema (what fields to extract), classification taxonomy (if applicable), summarization style, output constraints, retrieval (e.g., a glossary of contract terms), and a published/retired status.

A worked example: a real estate management firm processes lease documents at scale

A real estate management firm administers commercial leases across thousands of tenants. New leases, lease amendments, renewal notices, certificates of insurance, and regulatory notices flow in constantly. Document Intelligence turns this from manual data entry into reviewed extraction. The director of operations, Lukas, configures Document Intelligence like this.

Step 1: identify document classes.

• lease-document-intake — full new leases, structured extraction of parties, terms, key dates, rent figures.
• lease-amendment-intake — amendments to existing leases, extraction of changes.
• insurance-certificate-intake — certificates of insurance, structured extraction of coverage, dates, named insureds.
• regulatory-notice-classification — notices from regulatory bodies, classification into action-required vs informational.

Step 2: define each profile and first version. For lease-document-intake:

• Extraction schema: lessor, lessee, premises, term commencement, term expiration, base rent, escalation terms, security deposit, renewal options.
• Output constraints: extracted fields are always proposals (never auto-applied); any field that the model can’t extract with high confidence is marked uncertain.
• Retrieval: the firm’s standard lease term glossary (helpful context for extraction).
• Action policy: this profile can propose extracted field values; it cannot apply them automatically.

Step 3: publish. Versions are reviewed and published.

Step 4: live operation. As lease documents are uploaded into the workspace, the configured profile is invoked. Each invocation produces:

• A document intelligence run record (audit trail of the invocation).
• A set of suggestions: structured extraction outputs proposed for application to the document’s metadata.
• A confidence indication per suggestion.

The operations team reviews suggestions in the in-product reviewer UI. Accepted suggestions apply to the document metadata; rejected suggestions remain in the audit trail with reviewer notes; modified suggestions show the model’s original and the reviewer’s correction.

Step 5: improvement loop. As reviewer corrections accumulate, the operations team revises the extraction schema or prompts in a new profile version. The improved version is published; future invocations use it.

After a year:

• Lease intake is faster and more consistent.
• The audit trail shows every model output and every reviewer decision.
• An auditor reviewing the data quality of lease records can be shown the human-reviewed extraction discipline.

What you’ll see in the product

Document Intelligence lives under Governance → AI Governance → Document Profiles in the workspace, with the reviewer UI accessible from each document’s detail page.

The Profiles list shows every profile with its kind, current version, owner, recent invocation count.

Inside a profile, you see versions with their status, recent runs (link into Runs & Evidence), activity history.

Inside a document (in Document Governance), the Document Intelligence tab shows: recent runs against the document, suggestions and their disposition (accepted, modified, rejected), and the resulting metadata.

Every change is captured in the workspace Audit Log.

Suggestions, not silent application

A core principle: Document Intelligence proposes, humans decide. Even with high model confidence, the workspace does not silently apply extracted values to document metadata. The reviewer sees the suggestion, accepts or modifies or rejects, and the application is captured as a discrete event.

This matters because:

• An auditor can examine the application chain: what the model proposed, what the human decided, what the document record currently shows.
• Model errors don’t silently corrupt data.
• Reviewer corrections feed back into profile improvement.

Common workflows

Configuring a new document profile

1. Document Profiles → New profile. Title, key, kind, owner.
2. Create the first version: extraction schema, prompts, retrieval, constraints, action policy reference, test cases.
3. Route for review.
4. Publish.

Running Document Intelligence on a document

1. Upload the document into Document Governance.
2. The configured profile (per the document’s classification or folder placement) is invoked.
3. The run produces suggestions visible on the document’s DI tab.
4. The reviewer accepts/modifies/rejects each suggestion.
5. Accepted suggestions apply to the document metadata.

Improving a profile

1. Review reviewer corrections accumulated over a period.
2. Revise the extraction schema, prompts, or constraints in a new version.
3. Publish the new version.

Related

• AI Systems - underlying Document Intelligence system registration.
• Document Governance - documents subject to DI live here.
• Action Policies - controls what DI can do.
• Runs & Evidence - every DI run produces evidence.
• Governance Copilot - the sibling capability for drafting and analysis.