“Human in the loop” means little if the human only clicks approve at the end. A reliable workflow gives automation bounded tasks while people retain control over purpose, evidence, judgement and publication. The aim is to remove repetitive work without hiding responsibility.

The seven steps below can be adapted to articles, newsletters, product education and internal documentation.

1. Assign an accountable owner

One person should own the brief and the published result. Record the audience, reader need, desired action, subject boundaries and risk level. Health, finance, legal advice and claims about identifiable people require stronger expertise and review than a general productivity guide.

The owner decides where AI may help and where it may not be used.

2. Build a source pack before drafting

Collect primary sources, interviews, internal data and approved examples. For each material claim, note which source supports it and the date checked. Separate source facts from your team’s interpretation.

Do not ask a model to “research everything” and treat its answer as evidence. Search features can accelerate discovery, but the editor should open and assess the underlying source.

3. Create a constrained brief

A useful brief includes:

  • the reader and question;
  • the proposed conclusion;
  • approved sources;
  • facts that must appear;
  • claims that must not be made;
  • tone and terminology;
  • required sections and length;
  • disclosure or compliance requirements.

Ask the model to flag missing evidence instead of filling gaps confidently.

4. Generate structure before prose

Use AI to challenge the outline: What would a reader need to know next? Which section repeats another? What evidence is missing? Approve the structure before requesting paragraphs.

This makes problems cheaper to fix. It also prevents a fluent draft from locking the team into a weak argument.

5. Draft in reviewable sections

Generate one logical section at a time, with the relevant sources attached. Mark quotations, numbers, dates and product capabilities for verification. Avoid presenting generated customer stories or invented first-hand experience.

The human editor should add the value a generic model cannot supply: a real observation, decision framework, tested example or documented organizational experience.

6. Run separate review passes

Do not combine every check into “make this better.” Use distinct passes:

  1. Evidence: Does every material claim match a source?
  2. Reasoning: Do the conclusions follow from the evidence?
  3. Original value: Does the page offer more than a summary?
  4. Language: Is it specific, natural and free of repetitive filler?
  5. Risk: Are privacy, copyright, attribution and disclosure handled?
  6. SEO and accessibility: Are title, headings, links and alternative text useful and accurate?

Another AI can help surface inconsistencies, but it cannot be the final authority for facts it may reproduce incorrectly.

7. Approve, disclose and preserve the record

The accountable owner reads the final page in its published layout. Save the brief, source list, meaningful revisions, reviewer and approval date. Explain the use of automation where that context would help readers understand how the content was created.

For work where the human contribution matters commercially, keep a compact AI-writing authorship record rather than relying on a generic “human reviewed” label.

Google recommends focusing on accuracy, quality and relevance and suggests providing context about automation when appropriate. The U.S. Copyright Office also distinguishes human creative contribution from the mere provision of prompts. A documented workflow helps demonstrate where human decisions occurred.

A lightweight operating rule

For a small team, use a simple traffic-light system:

  • Green: outlines, headline alternatives, formatting and summaries of approved text;
  • Amber: factual drafts, comparisons and public recommendations—human evidence review required;
  • Red: confidential inputs, fabricated experience, unverified accusations or autonomous publishing.

Review this rule whenever the tools, data or business risk changes.

What the human must still own

Someone must define the purpose, choose the evidence, make the judgement and accept responsibility for publication. AI can accelerate parts of the path, but the editorial record should show who made the decisions and why the final article deserves trust.

Sources