An Editorial Workflow With an AI Step and a Human Gate

Distributed vigilance does not work. Put one gate in the pipeline, give it to a named person, define exactly what it checks, and record what it rejects. Everything before the gate can be as automated as you like; nothing passes it unchecked.

One gate, not oversight everywhere

The instinct is to add review at several points: a check after research, a check after drafting, a check before publishing. It reads as thorough and it produces the opposite, because responsibility spread across three checkpoints belongs to nobody. Each reviewer assumes the others caught the obvious things, and the obvious things are what get through.

A single gate has one property the distributed version cannot have: somebody is answerable. Ask who let a specific error through and the answer is a name, which changes how the check is performed long before anybody has to use the name for anything.

It also has to be the last thing before publication. A gate in the middle guards a draft, and the draft changes afterwards.

Where it goes and who holds it

  brief ──▶ research ──▶ outline ──▶ draft ──▶ style pass ──▶ [GATE] ──▶ publish
             │           │           │          │                │
             sources   claims     prose     checker           named human
             gathered  written    written   script            5 checks
             by human  by human   assisted  automatic         signs off

  Everything left of the gate may be assisted, automated or generated.
  Nothing crosses the gate without a name attached to it.

• The gate holder is not the drafter. A writer checking their own facts is checking their own memory of having checked them.
• The gate holder is named on the piece, in the disclosure note or in internal metadata. Anonymous sign-off is not sign-off.
• The gate has authority to reject. A gate that can only comment is a comment.
• The gate is not a style review. Style is the checker’s job, upstream and automatic — the constraints and the script exist precisely so that the gate does not spend its attention on spelling.

The five checks

Every piece. No exceptions for short pieces, updates, or things somebody is confident about.

1. Every factual claim traces to a source. Open the source. Not the summary of it, not the search result snippet, not a second article citing it. The claim and the source must be in front of you at the same time. Where the piece used the span-verified extraction workflow, this is fast, because the spans are already attached.
2. Every quotation is verbatim against the original. Against the recording, the document or the transcript — not against the notes. Quotations are the single highest-consequence error class, because a misquotation is attributed to a person who will read it.
3. Every name, number, date and title is checked individually. These are the categories a model gets wrong most often and the ones a human reader is most likely to notice, which is a bad combination. Read them aloud; the eye slides over a wrong digit and the mouth does not.
4. The disclosure decision is made and recorded. Above or below the threshold, and why — the threshold is on the disclosure page. Recording the negative decision matters as much as the positive one, because it is the record that the question was asked.
5. The sensitivity read, where it applies. Anything naming a living person, alleging wrongdoing, describing a legal matter, or touching a protected characteristic goes to whoever holds that responsibility. The gate’s job here is routing, not judging.

What is deliberately absent: any check that asks whether the piece is good. That decision was made when it was commissioned, and mixing it into the gate is how gates become slow and then become optional.

What the gate costs, and when it pays

The gate is the reason people abandon these workflows, so it is worth pricing honestly rather than asserting that quality is priceless.

One 1,200-word piece with 8 external claims and 2 quotations.
Assumptions, all substitutable:

  time saved by the assisted draft        40 min
  verification, 8 claims × 3 min          24 min
  quotation check, 2 × 4 min               8 min
  names/numbers/dates pass                 6 min
  disclosure decision, recorded            2 min
  ------------------------------------------------
  gate cost                               40 min
  net saving                               0 min

So on this piece the gate consumes exactly what the assistance saved.

That result is the honest one, and it is why so many teams quietly drop the gate. The case for keeping it is not that it is free; it is what happens on the pieces where it would have caught something.

Break-even on skipping the gate.

  gate cost                                40 min per piece
  cost of a published error, C             correction, republication,
                                           notification, and the time
                                           spent discussing it — call it
                                           180 min

  Skipping pays only if:  p × C  <  40
  where p is the probability a piece contains a publishable error.

  p* = 40 / 180 = 0.22

So the gate is a losing trade only if fewer than about 22% of ungated
pieces would carry an error worth correcting. And that ignores the cost
that cannot be put in minutes: a reader who stops believing the next
piece, whether or not that one was checked.

Two conclusions follow. The first is that the honest way to speed this up is to reduce claims per piece, not to check them faster: a piece with three claims costs fifteen minutes to gate and a piece with twenty costs an hour, and the twenty-claim piece is usually twenty claims because nobody decided which three mattered.

The second is that the gate cost is where the assistance should be aimed. Automating verification — span checks, identifier checks, quotation matching against a transcript — reduces the gate. Making the draft faster does not, and drafting is where every tool budget goes.

The rejection log is the point

Every rejection gets one line: date, piece, what was wrong, which category, where it came from. This costs the gate holder thirty seconds and it is the only mechanism by which the organisation learns anything.

date        piece            category        detail
2026-07-02  q2-retro         number          MRR figure was quarterly, prose said monthly
2026-07-09  vendor-compare   fabricated      cited a case study that does not exist
2026-07-09  vendor-compare   quote           paraphrase presented in quotation marks
2026-07-14  postmortem       name            engineer's surname misspelled
2026-07-21  pricing-explain  stale           quoted a rate changed in May
2026-07-28  security-notes   overclaim       "industry standard" with no source

After thirty entries the log answers questions nothing else can. Which error categories are frequent for your subject matter, which are concentrated in particular pieces or particular writers, and which are mechanical enough to move upstream into the checker script. A category appearing weekly is not a training problem; it is a missing automated check.

The log also produces the number in the break-even calculation above. After a quarter you know your own p instead of assuming one, and the argument for the gate stops being a matter of taste.

The rubber stamp, and auditing the gate

Every gate degrades. The mechanism is not laziness; it is that the checks pass overwhelmingly often, and a check that always passes stops being performed and starts being assumed. This happens within about two months and it happens silently.

• Audit the gate itself. Once a month, someone else re-checks one published piece completely. Fifteen minutes. It is the only way to find out whether the gate is running.
• Watch for zero rejections. A gate with nothing in its log for a month is either receiving flawless drafts or has stopped working, and the first has never happened.
• Rotate the holder. Attention to a repetitive check decays with familiarity, and a new holder finds things in the first fortnight that the old one had stopped seeing.
• Never let volume set the gate’s pace. If the queue exceeds what the gate can check, the correct response is to publish less. Deciding to check less is deciding to publish unverified work, phrased so that nobody has to say so.

What the gate does not do

Being clear about this is what keeps it fast enough to survive.

• It does not decide whether the piece is any good. That is commissioning and editing, both of which happen earlier.
• It does not check style. Automated, upstream.
• It does not make the piece original. A fully verified piece that contains nothing anybody did not already know is accurate and pointless, and the gate will pass it every time — which is the failure mode the slop page is about.
• It does not transfer accountability from the byline. The author remains answerable for the piece. The gate is a second pair of eyes, not an indemnity.