Content Approval Workflow for AI-Assisted Publishing: Three Gates, Because Fluent and Wrong Is the New Failure Mode

Content approval workflows were designed for human drafts, and human drafts fail in a recognisable way: they are rough, unfinished, sometimes off-brief, and honest about it. A reviewer can see where the writer ran out of time. AI drafts fail differently. They arrive complete, confident and well-formatted, and one paragraph in five is wrong in a way that reads exactly like the four that are right. A fabricated citation has the same shape as a real one. An approval workflow that was tuned to catch rough drafts approves fluent ones, and the error goes live under someone’s byline.

This article is a content approval workflow for teams that publish AI-assisted content. It has three gates, in a deliberate order: a mechanical verification gate that a script can run, a brief-conformance gate that an editor runs against the brief, and a human sign-off that is short because the first two gates did their job. It is the approval step of the content production workflow, designed for the draft that now arrives at it.

Key takeaways

• AI drafts fail as fluent and wrong, not rough and honest, so an approval workflow has to verify before it reviews; Walters and Wilder found 18% of GPT-4 citations entirely fabricated in AI-generated literature reviews.
• Gate one is mechanical: every statistic resolves to a source fetched in the same session, every external link is on a verified list, and no sentence appeals to unnamed authority.
• Gate two is conformance: the draft answers the brief’s question, uses the brief’s sources, links where the brief said, and does not compete with an existing page.
• Gate three is the human signature on the byline, which is a decision about responsibility, not a proofread, and should take minutes.
• Google’s position is that appropriate use of AI is not against its guidelines; the risk the workflow manages is being wrong, not being automated.

What a content approval workflow is for

A content approval workflow is the set of checks and sign-offs a draft has to pass before it is published, with a defined owner for each check and a defined outcome: approved, returned with changes, or rejected. Its purpose is to make sure the organisation stands behind what goes out under its name. That purpose has not changed. What changed is the draft.

With a human writer, approval catches the things a busy writer misses: a claim without a source, a paragraph that drifts from the brief, a tone that does not match the house. The reviewer reads, marks, returns. The workflow assumes the writer knows where every number came from and can be asked.

With a model in the drafting seat, that assumption fails. The draft cannot be asked. It can be prompted for sources, and it will produce them, and some of them will not exist. Content Science’s description of AI-era workflows makes the general point: AI reshapes workflows rather than replacing them, and mature workflows add explicit human-review checkpoints and new roles such as an AI reviewer. The approval workflow is where those checkpoints live.

Why AI drafts fail differently

Three failure modes are specific to model-written drafts, and all three are invisible to a reviewer reading for quality.

Fabricated and misattributed citations. Walters and Wilder analysed 636 bibliographic citations across 84 AI-generated literature reviews and found 18% of GPT-4 citations entirely fabricated, with 24% of the non-fabricated ones containing substantive errors, measured on GPT-4 as of mid-2023, with the 24% covering the non-fabricated citations only. Models have improved since, and the failure mode has not gone away; it has become rarer and therefore harder to spot by reading.

Plausible numbers with no origin. A draft that says “teams save 40% of production time” is making a claim that came from nowhere in particular. A human writer would have a source or would hedge. The model hedges only if asked, and the number reads as a fact.

Unnamed authority. “Research shows”, “experts agree”, “studies have found” are the model’s way of sounding sourced without being sourced. Each one is a claim with no owner, and each one is a liability if a reader, or a competitor, asks which research.

None of these are caught by reading for quality, because the sentences are good. They are caught by checking, which is a different job, and the approval workflow has to separate it from reviewing.

The three gates

The workflow is three gates in a fixed order. Each has an owner, a scope, a pass condition and a time budget. The order matters: the mechanical gate runs first because it is cheap and catches the failures that reading cannot, and the human gate runs last because it is expensive and should see only drafts that are already correct.

Gate	Owner	Checks	Pass condition	Budget
1. Verify	A script, with the editor reading its report	Sources, links, statistics, authority phrases, metadata	Zero failures in the report	Seconds to run, ten minutes to read
2. Conform	Editor	Question answered, angle held, required sources used, internal links present, no cannibalisation	Matches the brief line by line	Fifteen minutes
3. Sign	The person whose name goes on it	Voice, judgement, responsibility	They would defend every claim in it	Fifteen minutes

A draft that fails gate one goes back to drafting with the report attached. A draft that fails gate two goes back with the brief diff. A draft that reaches gate three and is rejected is a signal that gates one and two are letting something through, and the workflow, not the draft, gets fixed.

Gate one: the verification harness

The first gate is a set of checks a script runs on the draft, producing a report the editor reads. The checks below are the ones this site runs on every article before a human sees it; the tooling is described in content automation, but the list is what matters, and any of it can be done by hand on a small team.

• Claims table. Every statistic, date, named study and quoted figure in the draft is listed with the sentence it appears in and the source it came from. A row with no source fails the gate. This is the check that makes fabrication visible: a fabricated citation has a URL that does not resolve, or resolves to a page that does not contain the claim.
• URL whitelist. Every external link in the draft must be on the list of sources that were fetched and quoted during research. A link that was not fetched was not verified, whatever it looks like.
• Quote match. For each cited claim, the source page is fetched and the claim’s evidence is matched against the page text. A claim whose evidence cannot be found on the page is either wrong or paraphrased beyond recognition, and either way it fails.
• Authority phrases. The draft is scanned for the unnamed-authority patterns, the family of phrases that assert an unnamed source, a vague appeal to studies, experts or industry consensus with no name and no link. Each instance without an adjacent link fails.
• Canonical caveats. Statistics the organisation has used before are matched against a claims register that holds the canonical wording and the caveat that must travel with the figure. The Walters and Wilder fabrication figure used above carries one: it was measured on GPT-4 as of mid-2023, with the 24% covering the non-fabricated citations only, and that clause appears because the register requires it.
• Metadata and structure. Title and description within the lengths search engines display, a takeaways block, an FAQ with question-form headings, internal links that resolve, schema dates valid.

The output is a report, not a verdict. The editor reads the report, not the draft, and decides in minutes whether the draft is ready for gate two. On a draft that passes, the editor’s ten minutes are spent on the claims table, confirming that the sources are the right ones to have used, which is the beginning of gate two.

Gate two: conformance to the brief

The second gate compares the draft with the brief, and it only works if the brief exists in a form you can compare against: a named reader, the question they are asking, the angle, the required sources, the internal links, and the result of the cannibalisation check done before anything was written. If your briefs are a title and a word count, gate two has nothing to check, and the fix is upstream in the production workflow.

The checks are five, and each is a yes or no.

1. Does the draft answer the brief’s question in the first screen, not the fourth?
2. Does it hold the angle, or has it regressed to the generic version of the topic that already ranks?
3. Does it use the sources the brief required, and only sources that passed gate one?
4. Does it carry the internal links the brief specified, in context rather than in a list at the end?
5. Does it still not compete with an existing page? The draft may have drifted toward an intent another page already owns, which is a cannibalisation problem gate one cannot see.

A draft that fails goes back with the failing lines marked. Because the brief is the reference, the return is specific, and the rework is bounded. The rework rate at this gate is also the best measure of brief quality you will get: if more than one draft in five fails here, the briefs are the problem.

Gate three: the signature

The third gate is a person putting their name on the piece, and it should be treated as what it is: a decision about responsibility. The question the signer answers is not “is this good” but “would I defend every claim in this to a reader who challenged it”. The first two gates exist so that the honest answer is usually yes after fifteen minutes.

Three decisions belong at this gate and nowhere else.

Voice. Whether the piece sounds like the house. Mechanical checks cannot judge this, and it is the one thing the senior reviewer is uniquely placed to do.

Disclosure. Whether and how the use of AI in drafting is disclosed. Google’s stated position is that appropriate use of AI or automation is not against its guidelines, while its scaled content abuse policy treats many pages generated primarily to manipulate rankings as spam no matter how they are created; the question for the signer is what the house policy is and whether this piece follows it. How one model provider marks its own output is covered in how Claude marks AI-generated content.

Ownership. The US Copyright Office’s report on AI addresses the copyrightability of outputs created using generative AI, and the answer depends on how much human authorship the final piece contains. The signer is the human author of record, and the sign-off is where that authorship becomes real: the edits, the judgement calls and the decision to publish are theirs.

Routing without a queue

Most approval workflows fail not on the checks but on the waiting. The draft sits with the senior person for a week because they are also the strategy lead, the sales support and the person who writes the briefs. Three rules keep the gates from becoming queues.

• Gate one has no queue. It runs when the draft is submitted and returns a report in seconds. Nobody waits on a person for the mechanical checks.
• Gate two has a budget and a deputy. Fifteen minutes per draft, and a named second editor who takes the gate when the first is out. Conformance is a comparison against a document, so it does not need seniority.
• Gate three has a service level. Two working days from a draft passing gate two to a signature or a return. If it is missed twice in a month, the fix is a second signer, not a reminder.

Keep an audit trail for every piece: the gate one report, the gate two checklist, and the signer’s name and date. It costs nothing, it is what you will need if a claim is ever challenged, and it is the record that lets you improve the workflow by looking at where drafts were returned.

Where the gates sit in the wider stack

The approval workflow is one stage in a longer chain. Upstream, the content production workflow decides what gets drafted and how the brief is written. Downstream, automated content creation covers what the pipeline produces and how to keep it honest, and the case for building the gates as software rather than a checklist is made in content automation. For teams comparing tools, Optix against Jasper is a comparison built around exactly this question: whether the tool verifies what it writes before a human has to. The build-it-for-us version is AI tools and software.

The Content Marketing Institute’s B2B research reports that among B2B marketers using AI for content creation, 39% say content performance improved, 34% saw no change, and 12% say content quality decreased. The approval workflow is the difference between the first group and the last.

Frequently asked questions

What is a content approval workflow?

The set of checks and sign-offs a draft has to pass before it is published, each with an owner, a scope and a pass condition. For AI-assisted content it has three gates: a mechanical verification of sources, links and statistics; a conformance check against the brief; and a human signature that takes responsibility for the piece.

Why do AI drafts need a different approval workflow?

Because they fail differently. Human drafts fail as rough and unfinished, which a reviewer can see. AI drafts fail as fluent and wrong: fabricated citations, invented figures and unnamed-authority phrases that read exactly like correct prose. Reading for quality does not catch them; checking each claim against its source does, so verification has to run before review.

Who should approve AI-generated content?

The person whose name goes on it. Verification and brief conformance can be delegated to a script and an editor, but the final sign-off is a decision about responsibility, and it belongs with the human author of record who would defend every claim in the piece.

Does Google penalise AI-generated content?

Google states that appropriate use of AI or automation is not against its guidelines, and that using automation to generate content primarily to manipulate rankings violates its spam policies. The risk an approval workflow manages is publishing content that is wrong or valueless, not the fact that a model helped draft it.

How long should content approval take?

The mechanical gate runs in seconds and takes about ten minutes to read. The conformance check takes about fifteen minutes against a proper brief. The signature should take fifteen minutes on a draft that passed the first two gates, with a two-working-day service level so approval does not become a queue.