How Getir ran 200K product onboardings a month with 5 people, not 25
Getir runs a marketplace with roughly 90,000 restaurants on it. Every one of those restaurants wants to sell, and none of them can sell a single item until the item is live — photographed, described, categorized, checked for copyright, and approved. That approval step sounds trivial. At Getir's scale it was a wall.
Around 6,000 product images landed in the review queue every day. Behind that queue sat a 25-person QA team, opening images one at a time, checking each against the rules — is this the restaurant's own photo or a stock image lifted off Google, is the food actually in frame, is the category right, is the description usable — and either approving it or kicking it back. It was careful work. It was also the single thing standing between a restaurant signing up and a restaurant making money.
The math that made the queue impossible
Do the arithmetic and the problem explains itself. At roughly 90 seconds of human attention per image and 6,000 images a day, you are spending about 150 hours of focused review time every single day. That is what 25 people on a QA rota actually buys you: enough throughput to mostly keep up, with a backlog that swells the moment volume spikes or someone is out sick.
And volume only ever went one direction. More restaurants meant more products meant more images meant either a bigger team or a longer queue. Neither was acceptable. A bigger team is a linear cost that scales with the catalog forever. A longer queue is lost GMV dressed up as a process. Getir needed to break the link between catalog growth and headcount growth entirely.
The tell, when we mapped the flow with Getir's team, was that almost none of the 90 seconds was judgment. It was assembly and triage: figuring out what kind of request this even was, running the copyright check, spotting the quality problems, grouping the obvious approvals. A reviewer's actual expertise — the call on a genuinely ambiguous image — took a couple of seconds once everything else was laid out. The rest was the machine work a human shouldn't be doing by hand.
The flow we deployed
We built a four-stage flow that runs on every image the moment it hits the queue. The agent moves through all four stages optimistically — it doesn't stop and wait for a human at each gate — and a reviewer confirms the batch at the end in one pass.
Click a step. The agent runs all of them; a human confirms the last call.
The agent reads each incoming submission and sorts it by type and category before a human touches it — new product, image swap, description edit — so nothing lands in the wrong lane and the easy lanes stay separate from the hard ones.
The point of the design is what it doesn't do. It doesn't quietly auto-approve everything and hope. A human is still on the approvals — but that human is now confirming a pre-sorted, pre-checked, pre-scored batch instead of grinding through 6,000 raw images cold. The ambiguous ones still get a real person's eyes. The obvious ones stop stealing that person's day.
What changed
Compressing per-image handling from a minute and a half to under two seconds doesn't shave the cost of the queue. It changes what the queue is.
| Metric | Before | After |
|---|---|---|
| Time per task | 1.5 seconds | |
| QA team size | 5 people | |
| Photo coverage | +30% |
~€150K/yr in direct QA savings, plus €1M+ in GMV upside from restaurants going live the same day instead of three days later.
Per-task time dropped from 90 seconds to 1.5 — a 60× compression on the step that had been the constraint. The team that ran the queue went from 25 people to 5, and the five who remained moved from data-entry mode to genuine review: exceptions, edge cases, the calls that actually need a human. Photo coverage went up 30%, because the quality-enhancement stage caught weak or missing images that used to slip through or bounce around in email.
What the five people actually do now
It's worth being precise about what "25 people to 5" means, because it's easy to hear it as "we fired 20 reviewers and the robot does their jobs." That's not what happened, and if it were, the quality numbers would have gone down, not up.
The 25 were spending their days as a human sorting machine — opening images, running the same copyright check by eye, spotting the same framing problems, hunting for the obvious approvals. That work didn't require their judgment; it required their patience. The five who run the queue today do the opposite. They handle the submissions the agent flags as genuinely ambiguous, they set and adjust the rules the agent enforces, and they own the calls that carry real risk — a borderline copyright case, a category dispute, a restaurant pushing content that games the rules. That's skilled work, and it's the work the queue was drowning it in.
The other thing the five do is teach. Every time an operator approves a flagged image or overrides one the agent got wrong, that decision feeds back into the model, so the agent handles more of the same cases correctly next time and asks for help less often. The queue doesn't just clear faster than it did with 25 people — it gets a little sharper every week the five keep working it. Headcount stopped scaling with the catalog, and quality started compounding instead.
Feel the volume yourself
The reason this works isn't that the agent is perfect. It's that most of a review queue is easy, and easy work should clear itself. Drag the slider to your own daily volume and watch how the split falls out — how much clears automatically versus how much a human still needs to see.
Drag to your daily case volume. Qrambo clears the routine ones; your team stays on the 30% that need judgment.
Illustrative, based on a 70% auto-resolution rate and 1.5 min per manual case. Your numbers are set in the pilot.
Set it to 6,000 and the shape of Getir's problem appears: the overwhelming majority of images are routine and clear on their own, and the human team collapses onto the fraction that genuinely needs judgment. That is the entire mechanism. You are not replacing reviewers — you are aiming them at the 20% of work that was worth their expertise and taking the other 80% off their plate.
Where this fits beyond food delivery
Getir is a food-and-grocery marketplace, but nothing in this flow is about food. Any marketplace or e-commerce operation with a catalog-approval gate has the identical shape: a flood of seller-submitted content, a rulebook that has to be enforced, a human who has to stay accountable for what goes live, and a queue that turns straight into lost sales the moment it backs up. Product image review, listing moderation, seller onboarding, content compliance — same problem, same fix.
The pattern is the one that shows up across every vertical we work in: the calendar is eaten by assembly and triage, not by judgment. Automate the sorting, the checking, and the scoring; keep a human on the last call and log every decision; and let the routine volume clear itself. It's the same playbook that took Cenoa's onboarding from two weeks to two days — a different industry, an identical constraint.
If your catalog queue grows every time your seller base does, and your only two levers today are "hire more reviewers" or "let the backlog grow," there's a third one. Put one real review flow in front of us and we'll map it — the honest version, with a go / no-go at the end.