Futur Labs
← BlogOctober 28, 2026

Document Quality Assurance: Rules, AI Flags and Release Gates

Build document QA around deterministic checks, evidence-linked AI review, exception handling and approval records for the exact released version.

A magnifying glass over document tables while a reviewer marks an orange flag with a pencil.
Date
October 28, 2026
Category
AI Workflows
Reading
10 min
Bryce Choquer
Bryce Choquer
Founder, Futur Labs

Document quality assurance checks an artifact against defined requirements before release. Use deterministic rules for required fields and record comparisons, AI to flag possible meaning-level problems, and authorized people to resolve exceptions. Preserve the exact reviewed version. Passing automated checks means those checks passed; it does not establish overall document correctness.

This guide covers administrative document control and internal operational reporting. It does not certify engineering work, establish safety or provide a compliance opinion.

What is the difference between a rule, an AI flag and approval?

Separate these three outputs in the interface and in the audit record. A green badge that conflates them invites the wrong decision.

Layer Appropriate work Output Limitation
Deterministic validation Required fields, allowed formats, revision matching and recalculated totals Pass, fail or not evaluated for a named rule Depends on correct extraction, rules and reference records
AI-assisted review Possible contradictions, ambiguous wording or unsupported summaries Candidate issue with evidence locations Can miss problems or flag acceptable wording
Human disposition Resolve conflicts, assess context and accept or reject corrections Recorded decision for the artifact version Requires an authorized reviewer with relevant competence

A required heading can be present while its content is empty. Two numbers can agree because both copied the same erroneous source. An AI flag can identify a concern without establishing that it is a defect. Define what each check proves before presenting its result.

NIST's Generative AI Profile, published July 26, 2024, identifies risks from confident but incorrect output. That is relevant to a reviewing model as well as a drafting model: its reassuring assessment is still an output to evaluate. See NIST AI 600-1, section 2.2.

What inputs does a document QA run need?

Start with an explicit acceptance specification for a single document type. “Check this PDF” leaves both scope and authority undefined.

Input Required detail
Candidate artifact Document ID, revision, file digest and readable source file if available
Reference set Exact revisions of source records the document must agree with
Rule set Stable rule IDs, version, severity, applicability and release behavior
Extraction record Parser version, pages processed, missing tables and unreadable regions
Review policy Who resolves each issue type and who authorizes release
Distribution scope Intended audience and permitted content

Preserve both the extracted text and original rendering where policy permits. Headers, footnotes, table columns and symbols can lose meaning during extraction. If a page is unreadable, mark its checks as not evaluated. Do not report “no issues found” across material the checker never processed.

Cross-system reference matching requires stable identities and permissions. AI integration is the relevant commercial scope when approved revisions live in one system and the candidate report lives in another.

Which document checks should be deterministic?

Use rules when the expected result can be specified precisely. Document each rule's scope and test it against both compliant and deliberately defective fixtures.

Check Rule to define Failure handling
Identity Header ID equals the requested package ID Block wrong-document review
Revision Cover, footer and manifest revision values agree Route to document controller
Required content Each required section exists and has substantive content under the defined rule Return missing section locations
References Each listed source resolves to an allowed manifest revision Flag unavailable or stale references
Quantities Displayed subtotals reconcile to typed source rows using the defined formula Block the affected calculation
Dates Reporting interval and cutoff follow the contract Identify inconsistent fields
Placeholders Known draft tokens remain absent Block unfinished output
Review status Required decisions refer to this artifact digest Reject approval copied from an older version

“Substantive content” needs a document-specific definition; a character count alone cannot judge adequacy. Likewise, automated unit checks require an explicit quantity model. Matching the text “m” does not establish that two values describe the same physical quantity. Defer interpretation beyond the rule's scope to an appropriate reviewer.

What should AI-assisted review return?

Ask for candidate issues with bounded categories, rather than an unrestricted verdict that the document is “correct.” A useful finding includes:

  • Issue category and explanation in plain language.
  • Candidate document revision and precise location.
  • The relevant excerpt and any conflicting source location.
  • Why the comparison may matter under a named requirement.
  • Suggested reviewer action, without automatically applying it.
  • An explicit status of “candidate issue — unresolved.”

The reviewer can check whether a summary says every request was resolved while a source table includes open requests. The model may flag that contradiction; the source owner determines whether the narrative, source table or reporting scope needs correction.

Keep “no candidate issue returned” separate from “human reviewed.” Do not convert the model's self-reported confidence into an acceptance threshold without evaluating how that score relates to actual errors in the intended workflow.

What should an exception register record?

An exception register should connect each finding to the artifact, evidence and person responsible for resolving it. Keep issue status separate from document release status: resolving one finding does not clear the others.

Register field Required content Purpose
Artifact identity Document ID, revision and digest Prevent findings or approvals attaching to the wrong version
Check identity Rule ID/version or AI review category Distinguish a reproducible rule failure from a candidate concern
Evidence location Candidate location and permitted reference location Let a reviewer inspect the actual discrepancy
Evaluation coverage Processed, unreadable or not evaluated Expose checks that could not run
Release effect Blocking, advisory or pending a policy decision Make consequences explicit without an aggregate quality score
Disposition Assigned owner, decision, rationale and decision time Preserve accountability for corrections and accepted limitations
Recheck record New artifact version and repeated check outcomes Confirm that a correction was evaluated before release

Use a release summary that names unresolved blocking issues and incomplete checks. Avoid replacing those records with a reassuring overall percentage. A revision mismatch requires correction and another check; unreadable material requires a readable replacement or recorded manual review. An AI-flagged contradiction remains a candidate until the reviewer compares the source and document scope.

When an author changes the artifact, preserve the earlier findings and run applicable checks against the new version. Approval belongs to an exact reviewed artifact; it does not automatically transfer to its replacement.

How should versioning and release gates work?

Use a state model with distinct transitions:

  1. Received: identify the artifact and expected source package.
  2. Checked: store rule outcomes, extraction coverage and candidate AI findings.
  3. In review: assign unresolved issues to named roles.
  4. Ready for release: required issues are resolved and the designated reviewer has accepted the exact version.
  5. Released: record the authorized artifact and distribution event.
  6. Superseded or withdrawn: retain the earlier release record and reason for the change.

Store the rule version, model identifier, prompt version and input manifest with the run. These records help explain what produced a result; they do not guarantee an identical future model response. Keep sensitive source text out of broadly accessible logs, and set retention according to the organization's obligations.

A waiver needs an authorized person, rationale, scope and the rule permitting the waiver. Some requirements cannot be waived by an internal process. Do not let the model invent that authority or clear its own findings.

When a source changes, identify dependent checks and invalidate affected decisions. If dependency tracking cannot identify the scope safely, require a full recheck. Preserve the old report so a later reviewer can distinguish what was known at release from what changed afterward.

Where is the professional review boundary?

Document control can help a professional locate missing sections, references and unresolved issues. It cannot decide whether engineering analysis is adequate or whether a professional obligation has been fulfilled.

Engineers and Geoscientists BC explains the professional responsibility associated with authentication in its Practice Advisor FAQs, “Sealing Documents”, checked October 28, 2026. Its current quality management guide index separately lists documented checks and authentication guidance. Apply current requirements with the responsible professional; an automated QA result does not substitute for those processes.

These BC sources describe a jurisdiction-specific professional boundary. Confirm the requirements applicable to the work with the responsible professional.

How can a team evaluate document QA before relying on it?

Build an authorized reference set containing representative clean documents, known defects, unusual layouts and incomplete inputs. Label expected rule outcomes and candidate issues with the document owner. Include acceptable wording that resembles a defect so the test exposes false alarms.

Measure missed known defects and incorrect flags separately, grouped by issue type and severity. Also measure extraction coverage, reviewer corrections and unresolved exceptions. A system that finds formatting problems can still miss unsupported conclusions; one aggregate score hides that distinction.

Repeat the relevant evaluation when changing the parser, rule set, prompt, model or document template. Compare against a manual or rules-only baseline before attributing any improvement to AI. No measured accuracy, time saving or financial return is established by this article.

Treat uploaded documents as untrusted content. OWASP's LLM01:2025 guidance, checked October 28, 2026, describes indirect prompt injection through external files. Keep the checking step unable to approve releases or alter source records. Test hostile document instructions as well as ordinary mistakes; these controls reduce exposure without establishing immunity.

Common questions

  • An application can implement a narrowly defined automatic acceptance policy for suitable low-consequence checks. That policy needs an authorized owner and explicit scope. A model's “looks good” response is not an approval record or professional signoff.

  • Proofreading addresses wording and presentation. Document QA can also check identity, revisions, required evidence, numerical consistency and release decisions. Define which of those checks are in scope for each document type.

  • It can provide another candidate review, but agreement does not prove correctness. Both may rely on the same incomplete evidence or share a failure mode. Validate against sources, explicit rules and competent human review where required.

  • Record the affected checks as incomplete or conflicting and route them to the source owner. The release policy determines whether that blocks the document. Never resolve a missing record by inventing a replacement value.

What should the first implementation include?

Choose one document type, a bounded rule set and an issue queue with a clear release owner. Start with deterministic checks where the expected answer is explicit, then evaluate whether AI flags add useful coverage. AI workflow automation covers intake, checks, review routing and report handoffs; AI integration covers the permissioned reference connections and version identity behind them.

For assembling the initial report, use the AI report generation guide. General program planning remains in the AI implementation roadmap.

Subscribe

Field notes on AI agents, ERP, and shipping software — straight to your inbox. No spam, unsubscribe anytime.

The latest in AI, in your inbox.