Human review in an AI content workflow: where the person goes

Human review in an AI content workflow belongs before the draft and at the exceptions, not in one read of the finished draft. Put a person on topic approval, on the voice and rules, on any new internal number and on source checks. Let scripts run the mechanical checks. The last read of a clean draft is the weakest place to spend a person.

That is a design choice, and it has a cost. Below is the map: six touchpoints, what the person decides at each, and the evidence for putting them there.

Where should a person sit in an AI content workflow?

A person should sit wherever judgment is cheap and decisive: approving a topic before any words exist, owning the rules the machine follows, signing off new numbers, and checking sources. Each of those is a short decision with a clear outcome. Reading a finished draft is a long, vague task.

Touchpoint	When	What the person decides	What they do not do
Topic queue approval	Before any draft	Approve or reject each topic row	Rewrite drafts to fit a topic
Voice and rulebook	Before rules change	Wording of the voice document, lint rules	Let the machine edit its own rules
New internal number	At the pull request	Sign off a figure not yet public	Approve a number they cannot trace
Source and fact check	Before trust	Whether each cited figure matches its publisher	Rely on lint to catch a false fact
Failed gate or new number	At the pull request	Fix the draft or the rule, or sign off the number	Re-read a gate-clean draft line by line
After publish	When something was wrong	Fix pull request, repair pass	Patch live pages by hand

Why is one review at the end the weakest place for a person?

A reviewer at the end sees a finished, fluent draft with every mechanical rule already satisfied. That is the moment attention is lowest and persuasion is highest. The research on how people evaluate machine suggestions points the same way, though it comes from a different task and should be read as an analogy.

Beck, Eckman, Kern and Kreuter ran a randomised experiment with 2,784 paid crowdworkers (arXiv, 10 September 2025). Their Bias in the Loop paper had people check AI-suggested values in tables. Requiring corrections for flagged errors reduced engagement and increased the tendency to accept incorrect suggestions. Participants sceptical of AI caught errors more reliably.

Both halves matter. Overall accuracy was slightly higher, 68% against 66%, when corrections were required (p = 0.028). A 2-point gap is small, and the people involved were checking data tables, not editing blog drafts. The lesson we take is narrow: a forced pass over machine output does not guarantee attention, so design where attention is spent.

Readers have noticed. The Reuters Institute’s Generative AI and News Report 2025 (7 October 2025, YouGov, about 2,000 people in each of six countries, fieldwork 5 June to 15 July 2025) found 12% comfortable with news made entirely by AI and 21% when it is mostly AI with a human in the loop. Only 33% thought journalists always or often check AI outputs before publishing. It is self-reported and about news, not business blogs.

What does the person decide at each touchpoint?

The decisions are small and specific. A person approves topics, writes rules, signs numbers, checks sources and handles failures. Nothing in that list asks them to judge whether prose feels right on a fourth read. Here is each point in order, with the reason it earns a person.

Topic queue approval. A person approves each topic row before anything is written. If the queue has no approved topic, the run stops and reports. It does not invent one. Choosing what to publish is the highest-leverage judgment in the whole workflow, and it takes seconds per row.

Voice and rulebook. People write and change the voice document and the lint rules. The machine never edits its own rules. If it could, a bad draft would simply move the rule until the draft passed.

New internal numbers. A new performance figure blocks automatic merge until a named person signs off in the pull request. Figures already public on our proof pages can be reused. A new one needs a human who can trace it.

Source and fact check. This is the step no lint rule covers. The check is opening each cited source at the publisher and comparing the figure and the date. Our lint tests structure and banned patterns. It cannot tell a true sentence from a false one. On our own blog the drafting run opens the sources before a figure is used, and nothing mechanical proves it did. On a client install, the same limit applies, so the supervised first week is where the source rules get written down and tightened.

After publish. When something turns out wrong, the fix goes through a fix pull request, and periodic repair passes sweep for older drift. The person decides what counts as wrong. The machine does the edits.

Does a person click publish on every post?

No, on our own blog or on a client install. A post that passes every gate and carries no new internal number merges and deploys automatically. If a gate fails, or a new number needs sign-off, the pull request stays open for a person.

We made that choice deliberately. On a gate-clean draft with no new number, the publish click was the lowest-value review point we had. The topic was already approved, the rules were already enforced, and the person had nothing left to decide except whether to trust the machine again. We would rather spend that person on the queue and the rulebook.

Clients start with a new voice document and a new rulebook, so the first week is supervised. We watch every run, see what passes and what stops, and tighten the rules in writing until the drafts pass. After that the gates decide, as they do for us.

The gap that matters is facts. Automatic merge means a plausible but wrong figure could pass the lint. Rules and source checks narrow that gap and do not close it, which is why the source check sits in the table as its own touchpoint.

What do Google and the EU AI Act say about human review?

Google says to check AI content before publishing, and the EU AI Act treats human review as meaningful only when someone holds editorial responsibility. Neither source gives a procedure. Both point at the same thing: review needs a named owner who is answerable for what goes out.

Google Search Central’s page on using generative AI content (last updated 1 October 2026) says it is critical to manually factcheck and review AI-generated content for accuracy and trustworthiness before publishing. The same page warns that using AI to generate many pages without adding value for users may violate its scaled content abuse spam policy.

The EU AI Act (Regulation (EU) 2024/1689), in the consolidated text on the European Commission AI Act Service Desk, makes two relevant points. Article 14 tells overseers of high-risk systems to stay aware of automation bias, the tendency to over-rely on output. It applies to high-risk systems only. Article 50 requires disclosure of AI-generated text published to inform the public on matters of public interest, unless it has undergone human review or editorial control and a natural or legal person holds editorial responsibility.

Most small-business blog posts are probably outside Article 50(4), and this is not legal advice. The useful point is the wording: review counts when a person is answerable for the publication. A queue approval, a signed number and a named owner of the rulebook all create that answerability. A click on a finished draft does not create it by itself.

How do you place the people in your own workflow?

Start by listing the decisions that only a person can make: what to publish, what the voice is, which figures are true, which sources are real. Assign each to a named owner. Then check that every mechanical rule runs without a person in the loop.

Our content pipeline example walks one post through these touchpoints, and the quality checks post lists what the gate can and cannot test. The Content Engine installs this on your accounts: $4,900 setup, 2 to 3 weeks, optional $990 per month run, as of July 2026, with a supervised first week. The inbound machine proof shows it on a 363-product B2B catalogue, as of July 2026.

Start async

If you want your own gates designed around your voice and your topics, Start async and send the brief. You get a written reply within one business day.