Copied to clipboard!
Module 4 For Government intermediate 40 min

Planning & Document Workflows with AI

Reusable templates and Intelligent Document Processing

What you'll be able to do

  • Use AI to capture institutional knowledge inside a reusable workflow template
  • Build an Intelligent Document Processing pipeline that extracts and flags but never approves
  • Bake a human-review checkpoint into every consequential step

Intelligent Document Processing Workflow

Permits, applications, and records pile up, and reading and routing them by hand eats staff time. This exercise builds a repeatable workflow where AI reads a de-identified document, extracts the key fields into a structured table, flags what is missing instead of guessing, and produces a ready or not-ready determination. The approval decision stays with a human. Operationalizes Ascend Pilot 1 and its cognitive delegation zones.

Recommended tool

claude

Claude is strong at structured extraction and at flagging gaps rather than inventing values. Keep the whole workflow in one thread so the extraction rules and the completeness check stay consistent. If you want it grounded in your actual permit or records forms, upload de-identified samples to NotebookLM so it cites the source.

Bring to the exercise

  • The document type you process most (permit application, license, records request, intake form)
  • The key fields you need extracted from every one of them
  • What "complete" looks like: the required fields and attachments
  • Who makes the final approval decision (the human in the loop)
  • A de-identified or fictional sample document you can paste safely
Step 1

Set the Delegation Boundary

Decide what is AI Zone, Human Zone, and Collaboration Zone before any document is processed. Uses the Ascend Pilot 1 zone table.

Confidentiality Guardrail: Read Before Pasting
  • Use role labels, not real names (for example: the client, a partner, the team).
  • Strip identifying details, signed dollar values, account numbers, and anything that re-identifies a person or organization.
  • Confirm the AI tool you are pasting into is approved for this kind of information.
  • You, not the AI, own the final product. AI cannot invent facts.

Use the PII Redactor tool to scrub before pasting.

I am building an Intelligent Document Processing workflow for government records. Before we process anything, confirm these cognitive delegation boundaries and hold to them:

- AI Zone (AI may do this): document classification, data extraction from forms, completeness checking, status tracking, regulatory citation lookup.
- Human Zone (only a human decides): final approval decisions, complex zoning interpretations, variance determinations, constituent communication, policy judgment calls.
- Collaboration Zone (AI assists, human confirms): AI flags incomplete applications for staff review, AI drafts compliance checklists for staff validation, AI summarizes reports and staff confirms.

Confirm you understand these zones. For this workflow you will extract and flag, but you will NEVER issue an approval or denial. A human makes every consequential decision. Acknowledge before we proceed.
What good output looks like
  • AI restates the three zones in its own words.
  • AI explicitly commits to never issuing an approval or denial.

Verify before using AI's output

Step 2

Classify and Extract from a Sample Document

Extract the key fields into a structured table, flagging missing or illegible fields instead of guessing.

Confidentiality Guardrail: Read Before Pasting
  • Use role labels, not real names (for example: the client, a partner, the team).
  • Strip identifying details, signed dollar values, account numbers, and anything that re-identifies a person or organization.
  • Confirm the AI tool you are pasting into is approved for this kind of information.
  • You, not the AI, own the final product. AI cannot invent facts.

Use the PII Redactor tool to scrub before pasting.

What to substitute before pasting

  • [PASTE DE-IDENTIFIED DOCUMENT] A fictional or fully redacted sample. NEVER paste a real applicant name, address, or identifier. Use the PII Redactor tool first.
Here is a de-identified sample document. It contains NO real names, addresses, case numbers, or identifiers.

[PASTE DE-IDENTIFIED DOCUMENT]

Do the following:
1. Classify the document type.
2. Extract the key fields into a structured table: field name, extracted value, source location in the document.
3. For any field that is missing, blank, or illegible, mark it "MISSING" or "UNREADABLE". Do NOT guess or fill it in.
4. List any field whose value looks internally inconsistent and explain why.

Extract only what is actually present. If you are unsure, say so.
What good output looks like
  • A clean field table with a source location for each extracted value.
  • Missing and unreadable fields are flagged, not invented.
  • Inconsistencies are surfaced rather than silently normalized.

Verify before using AI's output

Step 3

Completeness Check and Staff-Review Flag

Produce a ready or not-ready determination with the specific missing items, routed to a human for the approval decision.

Confidentiality Guardrail: Read Before Pasting
  • Use role labels, not real names (for example: the client, a partner, the team).
  • Strip identifying details, signed dollar values, account numbers, and anything that re-identifies a person or organization.
  • Confirm the AI tool you are pasting into is approved for this kind of information.
  • You, not the AI, own the final product. AI cannot invent facts.

Use the PII Redactor tool to scrub before pasting.

What to substitute before pasting

  • [LIST YOUR REQUIRED FIELDS AND ATTACHMENTS] The fields and attachments every complete submission must include.
Using the extracted fields, run a completeness check against this requirement list:

[LIST YOUR REQUIRED FIELDS AND ATTACHMENTS]

Produce:
1. A "READY for staff review" or "NOT READY" status.
2. If NOT READY, the exact list of missing or deficient items, in plain language a resident could understand.
3. A one-paragraph summary a staff member can read in 15 seconds.
4. A clear handoff line stating that the final approval or denial decision belongs to the assigned staff member, not the AI.

Remember: you flag completeness. You do not approve.
What good output looks like
  • A ready/not-ready status with a specific, plain-language deficiency list.
  • An explicit handoff to a human for the approval decision.

Verify before using AI's output

Step 4

Build It Into a Repeatable Checklist

Turn the one-off run into a reusable workflow with the human-approval checkpoint named.

Confidentiality Guardrail: Read Before Pasting
  • Use role labels, not real names (for example: the client, a partner, the team).
  • Strip identifying details, signed dollar values, account numbers, and anything that re-identifies a person or organization.
  • Confirm the AI tool you are pasting into is approved for this kind of information.
  • You, not the AI, own the final product. AI cannot invent facts.

Use the PII Redactor tool to scrub before pasting.

Turn what we just did into a reusable Intelligent Document Processing checklist my team can run on every document of this type. Include:
1. The classification and extraction steps,
2. The completeness-check requirement list,
3. The MISSING/UNREADABLE flagging rule (never guess),
4. A required human-approval checkpoint before any decision is communicated to a resident,
5. A privacy reminder to de-identify every document before pasting it into AI,
6. A target metric to track (e.g., reduction in average document processing time of 50 to 75 percent),
7. A version date and a named workflow owner.
What good output looks like
  • A saveable checklist with the no-guess rule and a hard human-approval gate.
  • A tracked metric and a named owner so the workflow stays maintained.

Verify before using AI's output

Defining Your Workflow: The Court Clerk's AI Delegation

Right now, the quality of document intake, docketing, and case-file summaries depends entirely on which clerk does the work. When an experienced clerk leaves, their knowledge walks out the door. This exercise uses AI to capture that institutional knowledge inside a reusable template, so any clerk, including someone brand new, can run a consistent, accurate workflow. You are not using AI to do the clerk's job. You are using AI to BUILD a template that encodes the job, with a human-review checkpoint baked into every consequential step, because in government AI recommends and humans decide.

Recommended tool

claude

Building a reusable template is an iterative back-and-forth: brief the AI, get a draft, test it against a real scenario, find gaps, refine. Claude holds the whole build in one long thread and produces a clean, structured final template you can save and share. If you anchor it to your court's local rules, upload them to NotebookLM so the template stays grounded in that exact source.

Bring to the exercise

  • The clerk workflow you want to template (document intake, docket/scheduling, or case-file summarization for handoff)
  • The non-negotiable steps every case must go through
  • The questions filers and the public ask most often that clerks repeat all day
  • The points where cases most commonly get mis-filed, mis-scheduled, or bounced
  • Which steps are purely administrative vs. which require a clerk or judge judgment
  • (Optional) Your court's published local rules or filing-procedure documents to upload for grounding
Step 1

Stage 1: Map What the Template Must Know (No AI Yet, 10 min)

Surface the institutional knowledge before writing a prompt. If you struggle to answer, that gap is itself the data. It shows where the workflow is undocumented.

Write your answers on paper before opening any AI:

1. What must happen to EVERY document before it is accepted into the record? (Non-negotiables: required fields, signatures, fees, filing deadlines, jurisdiction check.)
2. What do filers and the public ask clerks over and over? (The repeated questions that eat your day.)
3. Where do cases most often go wrong: mis-filed, mis-scheduled, sent back for correction? (The friction points.)
4. How does the workflow differ for a self-represented filer vs. an attorney e-filing? (Variation by profile.)
5. Which steps are purely administrative (a new clerk could do with a checklist) vs. which require clerk or judicial judgment (deadline calculations, conflicts, sealing, anything consequential)?
What good output looks like
  • Concrete non-negotiables, not vague categories.
  • At least two friction points named from real experience.
  • A clear line drawn between "checklist-able" steps and "needs human judgment" steps.

Verify before using AI's output

Step 2

Stage 2: Brief the AI on How Your Court Works

You are not asking for help yet. You are teaching the AI your process so it can co-build the template. Let it ask clarifying questions. Gaps it surfaces are gaps the template must flag.

Confidentiality Guardrail: Read Before Pasting
  • Use role labels, not real names (for example: the client, a partner, the team).
  • Strip identifying details, signed dollar values, account numbers, and anything that re-identifies a person or organization.
  • Confirm the AI tool you are pasting into is approved for this kind of information.
  • You, not the AI, own the final product. AI cannot invent facts.

Use the PII Redactor tool to scrub before pasting.

What to substitute before pasting

  • [INTAKE / DOCKETING & SCHEDULING / CASE-FILE SUMMARIZATION] Pick ONE workflow to template first. Intake is the easiest starting point.
  • [DESCRIBE THE STEPS FROM YOUR STAGE 1 NOTES] Walk through your actual current process, step by step. Specificity here equals a useful template.
I am a court clerk. I want to build a reusable AI prompt template that any clerk, including someone brand new, can use to run [INTAKE / DOCKETING & SCHEDULING / CASE-FILE SUMMARIZATION] consistently and accurately.

I'm going to walk you through everything you need to know about how this workflow works in our court. Please ask me follow-up questions as I go, and flag anything that seems important to capture in a reusable template. Once we've covered everything, help me structure this into a master template prompt.

Important: this is a government records process. Some steps are consequential: deadline calculations, scheduling that affects a party's rights, anything touching sealed or confidential records. For every consequential step, the template must require a human clerk to review and approve before anything is finalized. AI recommends; the clerk decides.

Here is how the workflow runs today: [DESCRIBE THE STEPS FROM YOUR STAGE 1 NOTES].
What good output looks like
  • The AI asks clarifying questions instead of immediately producing a template.
  • It flags steps that sound consequential and asks where human review belongs.
  • It does not assume your local rules. It asks.

Verify before using AI's output

Step 3

Stage 3: Build, Test, and Find the Gaps

Build the template, then immediately stress-test it against a real scenario and ask what a 10-year veteran clerk would know that it missed.

Confidentiality Guardrail: Read Before Pasting
  • Use role labels, not real names (for example: the client, a partner, the team).
  • Strip identifying details, signed dollar values, account numbers, and anything that re-identifies a person or organization.
  • Confirm the AI tool you are pasting into is approved for this kind of information.
  • You, not the AI, own the final product. AI cannot invent facts.

Use the PII Redactor tool to scrub before pasting.

What to substitute before pasting

  • [A REALISTIC SAMPLE SCENARIO, use a fictional case, no real identifiers] Invent a case. NEVER paste a real case number, party name, or filing. Use "Case 1, a small-claims filing by a self-represented plaintiff."
Based on everything I've shared, draft a reusable master prompt template for clerks. The template should:
(1) Ask the clerk to enter the key details of the specific document or case being processed,
(2) Use those details to generate a step-by-step workflow (intake checks / docket entry / scheduling / or a case-file summary for handoff),
(3) Cover the full process from receipt through completion,
(4) Flag where a self-represented filer vs. an attorney changes what the clerk should check,
(5) Include a REQUIRED "Clerk Review & Sign-Off" checkpoint before any consequential step is finalized, and list exactly what the human must verify,
(6) Be formatted so a clerk with no prior experience can follow it.

Then test it: use the template as if I am a clerk processing [A REALISTIC SAMPLE SCENARIO, use a fictional case, no real identifiers]. Walk me through what it generates.

Finally, tell me: what did this template miss that a clerk with 10 years of experience would know to cover? What did it assume that isn't true for all case types? What local-rule language needs to be added?
What good output looks like
  • The draft template has clear input fields and a required human sign-off step.
  • The test run produces a workflow a new clerk could actually follow.
  • The gap analysis names real omissions (deadline math, jurisdiction edge cases, confidential-record handling), not generic ones.

Verify before using AI's output

Step 4

Stage 4: Finalize, Add Guardrails, and Make It an Asset

Lock the template, bake in compliance checks, and assign an owner so it stays current.

Confidentiality Guardrail: Read Before Pasting
  • Use role labels, not real names (for example: the client, a partner, the team).
  • Strip identifying details, signed dollar values, account numbers, and anything that re-identifies a person or organization.
  • Confirm the AI tool you are pasting into is approved for this kind of information.
  • You, not the AI, own the final product. AI cannot invent facts.

Use the PII Redactor tool to scrub before pasting.

Produce the final version of our Clerk Workflow AI Template. Format it as a structured document I can save and share with my team. Include:
(1) Instructions for how clerks should use the template,
(2) The input fields the clerk fills in,
(3) The full workflow the AI generates from those inputs,
(4) A built-in compliance section that prompts the clerk to confirm the output complies with current local rules and records-retention policy BEFORE use, and flags areas where rules may have changed,
(5) The required "Clerk Review & Sign-Off" checkpoint with an explicit verification list,
(6) A privacy reminder: never enter party names, case numbers, SSNs, DOBs, or sealed-record contents into the AI. Use placeholders,
(7) A version date and a named template owner responsible for updating it when rules change.
What good output looks like
  • A clean, saveable template with usage instructions, input fields, compliance section, sign-off checkpoint, privacy reminder, and version/owner fields.
  • The human-in-the-loop checkpoint is impossible to skip.

Verify before using AI's output