Design a Human-Checked Workflow

A workflow is not a prompt. A prompt is one instruction inside a larger system of inputs, permissions, tools, reviews, decisions, exceptions, and records. The workflow is what lets another person understand where work came from, who checked it, what changed, and who accepted the result.

AI can accelerate a draft, a summary, a transformation, or a first-pass comparison. It cannot absorb professional accountability, know the buyer's private context, or certify that an output is correct merely because it sounds confident. Design the human review before you celebrate the speed.

Begin With the Promised Result

Start from the accepted proposal and write the output in one sentence:

> Twelve follow-up message drafts, each tied to one approved customer moment, with source notes, a privacy check, and final approval by the buyer's coordinator.

Then identify what must be true for that sentence to count as delivered:

• The right inputs were used.
• The quantity and formats match the scope.
• Material claims can be traced to an approved source.
• Unnecessary personal data is absent.
• A competent person reviewed the specified risks.
• The buyer approved the final version.

Faster output that fails one of these conditions is not a complete delivery. The workflow should make failures visible before they reach the buyer.

Map Five Stages

A reliable small-service workflow can usually be described in five stages.

1. Intake and approval

Record what arrived, who supplied it, when it arrived, and whether it is the approved version. Check that the input belongs within the agreed scope and permission boundary. If the input contains unexpected sensitive or regulated information, stop and ask before processing.

2. Preparation and machine-assisted draft

Use only the approved inputs and approved tools. Prepare a small work packet with the objective, exclusions, source labels, required format, and uncertainty rules. The output should identify its sources rather than presenting every sentence as equally certain.

3. Mechanical checks

Before human review, check objective properties:

• Required elements are present.
• Counts and formats match the brief.
• Names, dates, links, and numbers appear in the expected fields.
• Prohibited fields or obvious sensitive identifiers are absent.
• Output is saved in the approved working location.

Rules and scripts can catch some of this, but they do not replace judgment. A present source note may still cite the wrong paragraph.

4. Human review

A named reviewer checks the work against a risk-specific checklist. The review must be independent enough to catch a plausible error. If the same person writes the prompt, accepts the first output, and signs the final version without a meaningful check, the step is documentation rather than quality control.

5. Approval, release, and record

Only a person with the agreed authority approves the final result. Record the approved version, date, reviewer, unresolved limitations, and handoff location. Preserve the evidence required by the agreement and delete or return material according to the data plan.

These stages can run in one afternoon for a small pilot. Their purpose is not bureaucracy. They make the invisible handoffs visible.

Assign a Risk Level to Each Step

Not every task needs the same review. Use the consequence of an error to set the control:

Low consequence

Examples include internal idea lists, first-pass formatting, or nonpublished brainstorming. A quick human scan may be sufficient if no private data or material decision is involved.

Moderate consequence

Examples include customer-facing drafts, summaries with factual claims, and reports used for an operational decision. Require source checks, a named reviewer, and a correction record.

High consequence

Examples include legal, medical, financial, safety, employment, benefits, identity, or regulated decisions. Do not use a general workflow to make the decision. Escalate to the appropriate qualified process and limit your role to the part you are competent and authorized to perform.

The review level should match the risk even when the tool is fast enough to produce more work.

Make the Prompt a Controlled Artifact

A reusable prompt should be stored with the workflow and include:

• Purpose and user.
• Approved inputs.
• Required output format.
• Facts that must be cited.
• Facts that must not be inferred.
• Prohibited content or personal data.
• Uncertainty instructions.
• Human-review reminder.
• Version and owner.

A prompt might say:

> Draft exactly four sets of three messages from the approved customer moments. Use only the supplied transcript and notes. Attach a source note to every claim about an event, person, date, or prior result. Mark any missing fact as "needs confirmation." Do not invent attendance, revenue, testimonials, or quotations. Do not include attendee names or contact details. The draft is not approved for release until the human checklist is complete.

This prompt controls the shape of the output. It does not prove that the output follows the instructions.

Separate Generation From Verification

The strongest simple pattern is: draft with one step, verify with another. The verification step should begin from the source, not only from the polished output.

Ask the reviewer to:

• Locate each cited claim in the source.
• Compare names, numbers, dates, and quotations.
• Check whether the output omitted an important condition.
• Look for invented relationships or unsupported certainty.
• Confirm that no unnecessary personal data appears.
• Confirm that the result still serves the buyer's approved purpose.

A second AI pass can help flag differences, but it cannot be the final authority. If a generated check says "all claims verified," a human still needs a sampling or risk-based review appropriate to the work.

Design the Exception Path

Normal paths matter less than what happens when something is wrong.

Write instructions for:

• Missing or unreadable input.
• Unexpected personal or confidential data.
• A source conflict.
• A failed quality check.
• A tool outage or terms change.
• A material scope change.
• A privacy incident.
• A buyer request outside your competence.

Each path needs an owner and an action. For example: "If an attendee name appears in the draft, stop release, quarantine the file, notify the coordinator through the approved channel, remove the field at the source, and rerun the check."

"Use judgment" is not an exception path. A calm, testable response is.

Keep a Fallback

If the AI tool is unavailable, slow, unapproved, or unacceptable to the buyer, the service still needs a truthful fallback:

• Manual drafting from the approved source.
• A smaller deliverable.
• A revised delivery date.
• A paused engagement.

Do not promise uninterrupted AI speed if your entire workflow depends on one account or vendor. A fallback can be more expensive. State that cost before the buyer depends on it.

Record the Work Without Recording Everything

Keep enough evidence to answer:

• What input version was used?
• Which workflow and prompt versions ran?
• Who reviewed and approved the result?
• What defects were found and corrected?
• What limitations were handed to the buyer?
• When should the data be deleted?

Do not keep every prompt, response, and intermediate file by default. More records can mean more privacy and access risk. Retain the minimum required to demonstrate delivery, handle a dispute, and improve the workflow.

A Worked Human-Checked Workflow

The following is an illustrative composite, not a claim that any particular tool is accurate or approved.

Nina's workshop pilot uses one redacted transcript and one program notes file. The buyer's coordinator approves the four customer moments. Nina creates a work packet with those inputs and the approved offer.

Her workflow has six controls:

1. The coordinator confirms the input version and removes attendee contact details.
2. Nina drafts twelve messages using only the approved source.
3. A mechanical check confirms that there are four groups of three and that each group names its source moment.
4. Nina creates a source-note sheet with the exact support for every material claim.
5. The coordinator reviews the source sheet and drafts, marking each item approved or returned.
6. Nina corrects the agreed defects, updates the final record, and sends the handoff note.

During the first review, one draft contains a quotation that sounds plausible but does not appear in the transcript. The coordinator flags it. Nina removes the quotation, records the exception, and adds a rule to require exact quotation location. She does not simply ask AI whether the quote is real.

The revised workflow costs more time than the first version, but it produces a result the buyer can trust. On the next engagement, the source-note sheet becomes a standard checkpoint.

Test the Workflow Before a Client Depends on It

Run a dry test with synthetic or approved nonconfidential material:

1. Follow the written steps exactly.
2. Record start and finish time for each stage.
3. Introduce one missing fact and one source conflict.
4. Check whether the reviewer notices both.
5. Ask a peer to follow the workflow without asking you for unwritten instructions.
6. Revise the step that caused the most confusion.

The goal is not to prove perfection. It is to find the point where the process becomes unclear, too slow, or dependent on one person's memory.

Your Chapter Exercise

Using your paid-pilot scope:

1. Write the promised result and five acceptance conditions.
2. Map the five stages from intake to release.
3. List every input, tool, reviewer, and output.
4. Assign low, moderate, or high consequence to each material step.
5. Write a review checklist for the highest-risk step.
6. Draft a controlled prompt or work instruction.
7. Define permissions and prohibited data.
8. Write exception paths for missing input, source conflict, privacy issue, and tool failure.
9. Run one dry test with nonconfidential material and time it.
10. Record one failure and one concrete workflow correction.

Chapter Deliverable

Create a workflow pack containing:

• Workflow map with five stages.
• Input and output register.
• Tool and permission list.
• Review matrix by consequence.
• Versioned work instruction or prompt.
• Human review checklist.
• Exception and escalation guide.
• Fallback plan.
• Dry-run record with timings and corrections.
• Retention and deletion instructions.

Source and Decision Caution

AI outputs can omit context, invent details, alter quotations, and reproduce biases or confidential information. Tool behavior and vendor terms can change by plan, account, region, and date. Check the current product documentation and the buyer's policy rather than relying on a remembered feature or a marketing claim.

Privacy, employment, consumer, professional-service, and regulated-decision requirements vary by jurisdiction. This chapter does not establish that a particular workflow is lawful, secure, or suitable for a regulated task. Use current primary guidance, the agreed contract, and qualified review when the consequence is significant.

Keep personal and confidential inputs out of unapproved tools. Limit records to what the workflow needs, control access, and verify that human reviewers have the competence and authority to perform the named check.

The next chapter turns these review points into a concrete privacy and client-data control plan.