Document Intelligence

AI-assisted educational manuscript. The clauses and extraction records are fictional. No document-upload or model service is claimed.

Learning objectives

• Map an extracted value to its source passage, version and scope.
• Preserve unknown, conflicting and corrected values.
• Separate machine proposals from authorized document interpretation.

A field needs more than a value

Document intelligence is useful only if an extracted field can be inspected and corrected. A value such as “fee: 0.5%” is incomplete without the fee type, base, currency, effective period and source passage. A document may describe an illustration rather than an actual charge. The system must preserve those distinctions.

NIST's AI RMF provides a voluntary framework for governing, mapping, measuring and managing AI risks. In this proposed educational workflow, that means defining which documents may be used, what the output may claim, who reviews it and how errors are corrected. It is not evidence of a deployed pipeline or a certification. NIST: Core

Worked example: one fee, two versions

Imagine fictional Terms v1 says a subscription charge is 1% of the accepted amount. Terms v2 says 0.5% for subscriptions accepted from an explicitly stated future date. A candidate extractor merges both into “subscription fee: 0.5%” for every transaction.

For a teaching subscription of 2,000 before the v2 effective date, the assumed v1 charge is 20. For the same amount after that date under v2, it is 10. The percentage is not wrong in isolation; applying it to the wrong date is wrong. Preserve both versions and select the applicable rule using evidence rather than overwrite history.

Suppose the source has no ATM fee because it describes a fund, not a payment card. The correct field is not applicable or out of scope, depending on the schema. It is not zero. A blank field, an unknown value, a contradictory value and a value of zero have different meanings.

Document revision → passage location → candidate field
  ├→ value and unit
  ├→ effective date and scope
  ├→ evidence status: supported / unknown / conflicting
  └→ review decision and correction history

The documents and amounts are invented. No real commercial fee is asserted.

Define the correction lifecycle

A candidate extraction starts as proposed. A reviewer may confirm, correct, reject or request another source. Keep the machine proposal and reviewer change distinguishable. Do not silently replace a number and then attribute the corrected answer to the original extraction.

Record a document checksum or other revision identifier so a later update can invalidate affected fields. Passage references need enough context to remain inspectable: section, page where available, text excerpt and document identity. PDF layout or OCR uncertainty should be represented rather than hidden by a precise-looking decimal.

A source document can contain irrelevant or malicious instructions. It remains untrusted input to the task, even when it is official evidence about a financial term. Its financial assertions can be researched; it cannot authorize the assistant to change policy, run commands, transmit secrets or publish content. Separate document text from control instructions in any future implementation.

Evaluate normal extraction, absent fields, conflicting versions, tables, footnotes, conditional fees and units. Measure whether the system abstains correctly and whether a reviewer can reproduce the result from the cited revision. Do not claim that schema validity proves semantic truth.

Common misconceptions

“Valid JSON means correct extraction.” It proves structure, not source support or interpretation.

“Missing means zero.” An absent fee is not evidence of a free service.

“The newest document governs every past event.” Effective dates and contractual scope matter.

Risks and limits

This is a design exercise, not a production document-intelligence feature. It stores no private documents and makes no accuracy claim. Real implementation requires data permissions, retention rules, access control, source licensing and evaluation appropriate to the intended documents. Human legal interpretation is not replaced by a field extractor.

Exercise and answer

Exercise: A 2,000 subscription predates the 0.5% rule and is governed by the fictional 1% version. What fee applies under those assumptions, and should v1 be deleted?

Answer: The charge is 20. Preserve v1 and its scope; v2 does not erase the evidence for earlier accepted transactions.

Related learning

Continue in AI and future finance with compliance assistance, and revisit source versioning in the engineering phase.

Source map

nist-ai-rmf-core: NIST's voluntary AI-risk framework, directly read on 2026-10-18. The field schema, clauses and correction process are original teaching designs. No extraction provider, private document or live fee schedule was tested.