# Design a human-reviewed AI document intake workflow

**trustWF workflow pack** · IT & Business Automation

## Outcome
A controlled intake design that classifies documents, extracts fields with provenance, routes exceptions and retains human approval before consequential writes.

## Who this is for
Teams considering AI-assisted intake of forms, invoices or requests

## What to provide
Document types; allowed data classes; required fields; source systems; confidence or exception policy; human reviewer roles; retention and access rules. Use synthetic examples only.

## Steps
1. Define allowed document types, required fields and prohibited data.
2. Map extraction fields to source locations and specify uncertainty handling.
3. Design exception queues, reviewer actions and audit logging.
4. Set test cases, access controls and human approval boundaries.

## Copy-ready instructions
Stage 1 — Prepare and diagnose
When to use: Before doing the work, use this stage to check source quality, definitions, scope and critical missing information.
Copy and paste:
You are a careful, tool-neutral assistant helping teams considering ai-assisted intake of forms, invoices or requests. Help me design a human-reviewed ai document intake workflow. Use only the information I provide. Separate confirmed facts, assumptions and suggestions. Never invent figures, source claims, commitments or policy. Ask no more than three essential questions if missing facts would change the result; otherwise use clearly marked placeholders. Keep language plain. Remove or mask personal and confidential information before use.

Context: [organization, role and situation]
Source material: [paste only information you are permitted to share]
Constraints or policy: [format, time, rules and approvals]

Stage 1 — Prepare and diagnose
Confirm the document purpose, data sensitivity, approved AI environment, field definitions and consequences of an error.

Return: A brief input check, missing-information list and any essential questions. Do not hide uncertainty. Mark missing facts as questions or placeholders.
Expected output: A brief input check, missing-information list and any essential questions.
Check before continuing: Check that the period, units, source and decision boundary are correct.

Stage 2 — Analyze and plan
When to use: Use this stage after the inputs are ready. Carry forward the Stage 1 input check and supplied material.
Copy and paste:
You are a careful, tool-neutral assistant helping teams considering ai-assisted intake of forms, invoices or requests. Help me design a human-reviewed ai document intake workflow. Use only the information I provide. Separate confirmed facts, assumptions and suggestions. Never invent figures, source claims, commitments or policy. Ask no more than three essential questions if missing facts would change the result; otherwise use clearly marked placeholders. Keep language plain. Remove or mask personal and confidential information before use.

Context: [organization, role and situation]
Source material: [paste only information you are permitted to share]
Constraints or policy: [format, time, rules and approvals]

Stage 2 — Analyze and plan
Design extraction with a source reference for each field, explicit unknown values, exception criteria and a human review step. Describe accuracy measures as targets to test, not promises.

Return: A traceable analysis or plan with calculations, rules and uncertainty stated. Do not hide uncertainty. Mark missing facts as questions or placeholders.
Expected output: A traceable analysis or plan with calculations, rules and uncertainty stated.
Check before continuing: Check formulas, definitions, priorities and evidence before using the draft.

Stage 3 — Produce and review
When to use: Use this stage after you accept the analysis. Carry forward prior outputs and correct any facts before finalizing.
Copy and paste:
You are a careful, tool-neutral assistant helping teams considering ai-assisted intake of forms, invoices or requests. Help me design a human-reviewed ai document intake workflow. Use only the information I provide. Separate confirmed facts, assumptions and suggestions. Never invent figures, source claims, commitments or policy. Ask no more than three essential questions if missing facts would change the result; otherwise use clearly marked placeholders. Keep language plain. Remove or mask personal and confidential information before use.

Context: [organization, role and situation]
Source material: [paste only information you are permitted to share]
Constraints or policy: [format, time, rules and approvals]

Stage 3 — Produce and review
Return a workflow map, field schema, exception rules, reviewer interface needs, audit events, data lifecycle notes, synthetic test set plan and launch gate.

Return: The requested draft plus assumptions, unresolved questions, a review checklist and one next action. Do not hide uncertainty. Mark missing facts as questions or placeholders.
Expected output: The requested draft plus assumptions, unresolved questions, a review checklist and one next action.
Check before continuing: Verify facts, policy, calculations and required human approval before acting.

## Review before using
Do not use real confidential documents in an unapproved AI tool. Do not auto-approve payments, employment actions or other consequential decisions. Validate privacy, security, retention and accuracy with accountable owners.

---
Use the result as a draft. Review it for accuracy, context and any required approvals.