Staffing the review side, not the make side

Superside's Breakpoint research found that 80% of creative teams are at or beyond capacity and 70% of creative leaders report burnout — and that is after widespread AI adoption. Output went up. The pressure did not come down.

That combination only makes sense one way. Generation stopped being the thing that limits how much work leaves the building. Filtering, brand governance, rights checks and someone's actual taste became it. Which means the most common capacity decision in creative teams right now — add another person who makes things — is being made against a constraint that moved.

Adding a maker to a team whose approval stage is already full does not increase throughput. It increases work-in-progress, ages every job in the queue, and degrades the review that was already the bottleneck. This is the test that tells you which side of the line your next hire belongs on.

The tell that your constraint moved downstream

Before measuring anything, check for the four signs. If three of them are true, you almost certainly do not have a generation problem.

• The queue of finished-but-unreviewed work is never empty. Not "sometimes backed up." Never empty, including Monday morning.
• Your best reviewer is also your best maker. The person with the taste to approve is the person you keep pulling onto production, so review happens in the evening gaps.
• Rework is concentrated, not random. The same three notes come back across different projects, which means the review pass is catching things a definition should have caught.
• Deadlines slip in the last mile. Assets exist days before the deadline and still ship late.

The fourth is the diagnostic one. If the material is ready and the date still moves, the delay is entirely downstream of making, and no amount of extra making touches it.

The staffing test: two weeks of timestamps

You need four timestamps per job and one number per stage. Nothing else. Two weeks is enough if you ship more than about fifteen jobs in that window; if you ship fewer, run it for a month.

For every job, log:

Timestamp	Recorded when
T0	Intake is complete and the job is startable
T1	Generation output exists and is handed to review
T2	Review verdict is final (approved, or rejected with notes)
T3	Delivered to the client or published

Then, separately, log touch time: the minutes a human was actually working on that job in each stage. Not the calendar span. The hands-on minutes.

Now compute two things per stage.

Elapsed share. Stage elapsed divided by total elapsed (T3 minus T0). This tells you where the calendar goes.

Flow efficiency. Touch time divided by elapsed time for that stage. This tells you whether the stage is busy or waiting. A stage at 60% flow efficiency is genuinely saturated with work. A stage at 8% is a queue.

The decision rule falls out of the arithmetic. If generation is 20% of elapsed time, then making generation infinitely fast improves your delivery date by at most 20%. Doubling generation capacity improves it by considerably less than that. You cannot hire past a stage that is not the majority of the clock.

What two weeks shows	Where the next hire goes
Generation is over half of elapsed time, and its flow efficiency is high	Upstream. The maker is genuinely saturated.
Generation is under a quarter of elapsed time, review queue never empties	Downstream. Hire review, not making.
Review flow efficiency is low but review touch time is also low	Nobody. You have a scheduling problem, not a staffing one.
Both stages fine, but rework rate is high	Nobody. Fix intake.
The largest elapsed block is "waiting on client"	Nobody. Fix the review loop.

That fourth row is the one people misread most often. A high rework rate looks like a capacity shortage because everyone is busy, but the busyness is second passes on jobs that should have been right the first time. Hiring into it buys you more of the same rework at a higher payroll number.

Why more generation makes a saturated stage slower

This is not a soft claim about morale. It is a queue property.

When arrival rate at a stage exceeds its service rate, the queue does not stabilise at a longer length. It grows without bound until something else gives. In a creative team the thing that gives is review quality, because the reviewer is a human who can see the backlog. Under a visible backlog, review gets faster per asset, which means it gets shallower, which means more defects reach delivery, which generates rework, which arrives back at the same saturated stage as new work.

So the second-order effect of adding generation capacity to a full review stage is a rise in the defect rate. You feel it as "quality slipped when we scaled," and it gets blamed on the models.

The related trap is measuring the wrong unit on the generation side. The number that matters is time per usable asset, not per generation. A generator producing more rejects is not producing capacity; it is producing review load. Reroll rates and the shots you throw away covers how to instrument your own version of that number.

What a downstream hire actually does

"Hire a reviewer" sounds like hiring a bottleneck-shaped person, which nobody wants to write a job description for. The role that works is narrower and more mechanical than that.

They own the definition, not just the verdict. A reviewer who only says yes or no is a queue. A reviewer who converts each rejection into a named defect class, a prompt fix, or a checklist line is a rate improvement, because the same defect stops arriving.

They own the checks that do not require taste. Duration, aspect ratio, safe-area, caption presence, loudness, disclosure label, provenance metadata, claim substantiation. None of that needs a creative director. All of it currently interrupts one.

They own rights and provenance. With EU AI Act Article 50 binding since 2 August 2026, and C2PA Content Credentials now shipping by default in a growing number of tools, the delivery spec includes machine-readable disclosure. Someone has to check it survived the export. Content credentials through a real pipeline is the practical version of that check.

The industry framing for this shift is Creator to Curator: senior people who design inputs, curate outputs and enforce brand rather than producing every asset by hand. One vendor source puts the ratio at roughly one AI creative specialist per five to seven traditional creatives; treat that as a starting hypothesis, not a benchmark, because it has not been independently tested. The content team roles that matter in 2026 has the fuller role map, including which roles to stop backfilling.

Three things to try before you hire anyone

A hire is the most expensive way to buy review capacity and the slowest to arrive. Each of these buys some of it in a week.

1. Move checks off the reviewer. Anything a script or a model can verify should never reach a human as a question. A pre-publish machine pass catches the mechanical failures and hands the human a severity rating instead of a raw asset — automated QA before a paid placement goes live walks through where that pass earns its cost and where it does not.

2. Pre-approve formats, not assets. For a repeating series, approve the template once and review only the variable parts. This is the single largest reduction in review load available to most teams, and it costs a meeting. Approval workflows that don't stall covers the failure modes.

3. Make iteration stop consuming review slots. A lot of review time is spent on questions the reviewer should never have been asked — timing, wording, crop. The editor is EDL-based, so the cut is one re-renderable timeline, preview: true returns a 480p pass at no credit cost with a short per-user cooldown between passes, and the final export is charged once regardless of clip count. Settle the mechanical questions at preview resolution before anything enters the review queue. Previews and the final export has the billing shape.

If you run all three and the review queue still never empties, you have a real staffing case, and now you have two weeks of timestamps to make it with. Measuring content team productivity is where those numbers belong in the recurring report.

FAQ

How many jobs do I need before the measurement means anything?

Fifteen completed jobs is enough, because you are looking for a stage at 50% of elapsed time versus one at 15%, not a 5-point difference. If your stages come out within ten points of each other, you do not have one bottleneck, you have a general capacity shortage, and the answer is fewer concurrent jobs rather than a targeted hire.

Isn't a reviewer just a cost centre?

Only if you define the role as approving. Defined as owning the defect definition, the checklist and the rights pass, the role reduces rework at the source, which is a rate improvement on the whole pipeline. The measurable version is rework rate before and after — if named defect classes and a checklist do not move it within a quarter, the role was scoped as a queue.

We are two people. Does any of this apply?

Yes, but it becomes sequencing rather than hiring. At two people, stop generating on days you have a review backlog: you are both stages, and interleaving them makes both worse. Batch making, then batch reviewing, with a hard cap on open jobs. One marketer with the output of a full team is the small-team version.

What if the client is the bottleneck?

Then no hire on your side changes the delivery date, and the fix is contractual and structural: a named single approver, consolidated feedback in one document, a stated feedback window, and share links rather than file transfers so reviewing takes a tap. A client review loop built on previews and share links is the operational build.