Stories that sit between review states and nobody can say whose they are

Measure before asking anyone. For every transition pair in the workflow — draft to sub, sub to check, check to ready — record the count, the median age and the p90 age over the last 14 days, bucketed by the hour and weekday the story entered the state. The four shapes that come out of that table each mean something different, and between them they cover almost every stalled-story complaint a newsroom produces. Asking the desk where the bottleneck is produces the state everyone finds most annoying, which is rarely the state with the longest tail.

Measure the transitions before asking the desk

This needs a transitions log, not a status field. If the platform records only the current state, there is no age to measure and the investigation cannot start. Legal review is deliberately out of scope here: it has a legitimately long and irregular tail, and it belongs in its own analysis in the pre-publication legal read as a workflow state.

Four distributions, four different problems

Bucketing by entry hour is what separates the second and third rows, and it is the step most teams skip. A tail made entirely of stories that entered a state after 18:00 or on a Saturday is not a workflow defect at all — it is a rota, and no amount of state redesign will change it.

The four causes, ranked by how often they are the real one

1. The state has no owner. Assignment is to a group — 'with subs', 'production' — which means no individual is accountable and no individual has a queue. This is the most common cause by a distance, and its signature is the long tail: the items that stall are the ones nobody happened to pick up. Assign at entry, to a person, with a visible personal queue and a way to hand it on.
2. The state models the org chart rather than an action. 'With the Deputy Editor' describes a location, not a task, and nobody can tell whether it has been done. Apply the verb test: name every state as the action it demands — sub it, check the figures, clear the picture, approve the splash. A state you cannot name as a verb is usually two states wearing one label, or a state that should not exist.
3. Parallel work is modelled as a queue. Picture desk, standards, subbing and legal do not need each other's output, and modelling them as a chain makes total time the sum instead of the maximum. Four checks taking 20, 30, 45 and 60 minutes cost 155 minutes in sequence and 60 in parallel — and the sequential version also blocks all four on whichever person is unavailable. Model these as independent gates on one state, with the story advancing when all gates are satisfied.
4. The desk routes around the workflow entirely. The states exist, and the real coordination happens in a chat channel. The confirming query is precise: for each published story, measure the interval between its last recorded transition and its publication. A large cluster under 5 minutes means the states are being clicked through as paperwork immediately before publishing, which tells you the workflow is documentation of a process that happens somewhere else.

When the chat channel is the real workflow

The fourth cause is the uncomfortable one, because the fix is not technical. If the desk has built a parallel system in chat, the platform's states are describing a process nobody follows, and adding fields will make the gap wider rather than narrower. Two questions settle what to do. Does the chat version work? If it does, the honest move is to model what the desk actually does and delete the states that describe the process somebody once designed. If it does not work — things are lost, ownership is unclear, nobody can reconstruct who approved what — then the states are right and the tool is losing to chat on convenience, which is a latency and notification problem rather than a modelling one.

One category is worth exempting before you delete anything. Gates that exist to catch a specific failure are not slow because they are pointless; they are slow because nobody owns them. The check that generated text does not reach readers unreviewed is exactly that kind of gate — it must exist, it must fail closed, and it needs a named owner rather than good intentions, as set out in machine-drafted copy reaching the live site.

A state with no owner is not a step in a process. It is a place a story waits for somebody to feel responsible.

Remove the state, or give it an owner and a deadline

1. Does the state ever refuse anything? Count backward transitions over 90 days. Zero means delete it and record whatever it was capturing as a field on the story instead.
2. If it does refuse things, can you name it as a verb? If not, split it until you can. Two people waiting on one ambiguous state is the commonest way a story stops.
3. Is it genuinely dependent on the state before it? If not, take it out of the chain and make it a parallel gate. Total time drops to the slowest gate rather than the sum of all of them.
4. Who owns it at entry? Name a person, not a desk, and give them a queue they can see without being told to look at it.
5. What happens when the deadline passes? Escalate to a named second, or advance with a recorded exception. A deadline with no defined action is a reminder, and reminders are ignored within about a month.
6. Agree the deadline with the people who work the state, and publish the age-in-state numbers to them weekly. An SLA imposed on a desk gets gamed; one the desk set gets defended.

Deadlines and defaults are the part most easily built and most often skipped, because they need a small piece of always-running machinery rather than a screen. That is the shape of work we take on as AI agents and automation, and the queue-age technique itself is not specific to newsrooms — it is the same measurement used on any operational queue with humans in it, applied here the way we describe applying it in AI in logistics operations.

Two adjacent problems are worth naming so they are not misdiagnosed as this one. If the hold-up is verification rather than approval, the evidence belongs in the claim-level record described in a fact-check record that survives a complaint. And if stories stall after publication rather than before it — a change agreed and not propagated to every version that ran — that is a distribution problem, traced in a correction on the original leaves five versions wrong. The rest of this silo sits under editorial workflow and newsroom AI, inside our media and publishing practice.

• editorial workflow
• review states
• newsroom tools
• process design