Can AI not do this now?

It does a large part of it, and that has changed the shape of the work rather than removed it. Automated classifiers clear what is obviously acceptable and stop what is obviously not. What reaches a person is the middle — sarcasm, context, local idiom, a listing that is technically compliant and plainly fraudulent, two clauses of your own policy pointing in opposite directions. That work is harder than what moderators did five years ago, which is why the skilled end of this service costs more than it used to.

Who writes the policy?

You do. We apply it, and we are strict about that boundary because a moderator's personal opinion has no value in a queue and your policy is the thing you can defend. Where we add value is by finding the places your policy contradicts itself or fails to cover something real, and raising those in writing with examples rather than quietly deciding on your behalf.

How do you measure whether moderation is any good?

With two numbers together, never one. Precision is what proportion of the items we removed genuinely deserved removal. Recall is what proportion of the items that deserved removal we actually caught. Someone who removes everything scores perfectly on recall and destroys your platform, and someone who removes nothing scores perfectly on precision and lets it rot. Both figures appear monthly, alongside a sample of the disagreements.

What kind of material will you not take on?

The most severe categories are not accepted without professional counselling support arranged and funded as part of the engagement, and we will say no to that work rather than staff it badly. For everything we do accept, three things are non-negotiable — reviewers are told on day one exactly what they will be looking at, shift length is capped with scheduled breaks, and nobody is left on the hardest category continuously.

How fast will our queue be cleared?

To an agreed target we set after measuring, not before. Review speed depends almost entirely on how hard your decisions are — a spam and duplicate-listing queue moves at a completely different pace from one where each item needs context and cross-checking. We measure throughput and decision quality during a calibration period and then commit to numbers we can hold without pushing accuracy down to hit them.

What happens to the record of decisions?

It stays in your system, and it is easy to underrate until the day it is needed. Every action carries a decision, a policy reference and a reason. When a user appeals, when a regulator asks, or when a removal is challenged publicly, the difference between a defensible position and an embarrassing one is whether that record exists and is consistent.

QA, Support & Ongoing Operations

Content moderation against your policy, at queue speed

Posts, listings, images, comments and reviews reviewed to your written policy, with every decision recorded, calibrated reviewers, and precision and recall reported monthly.

From: $700–$3,000/agent/mo
Typical time: ongoing
Usually bought by: Platforms, marketplaces

What this actually solves

The moment you let people post, you take on a responsibility you did not previously have. Three consequences follow, and buyers usually arrive holding one of them. The legal one — in a growing number of jurisdictions a platform is answerable for what its users publish, and the penalties are not symbolic. The reputational one — a single offensive comment screenshotted and shared costs months of goodwill, and it costs it in an afternoon. And the commercial one — a marketplace running fraudulent listings loses buyers first and sellers shortly after, because trust is the product.

The difficulty is not volume, it is consistency, and that is the part that gets underestimated. Ten reviewers looking at the same borderline listing will produce a spread of answers unless they have been deliberately brought into alignment, and an inconsistent queue is worse than a slow one. Users notice immediately when the same behaviour is punished on Tuesday and ignored on Thursday, and once that is visible your policy stops meaning anything.

Automation has changed where the difficulty sits. Classifiers now handle the clear cases at both ends, so the human queue is concentrated on ambiguity by design. That produces one good effect and one that has to be managed honestly. The good one is that the work is more skilled and the people doing it are worth more. The one to be managed is that what reaches a person is now, on average, the most distressing material on the platform, and the operation has to be built around that fact rather than around pretending otherwise.

How the work runs

Read the policy properly and stress-test it. Line by line, against real examples, looking specifically for the places it contradicts itself, uses a word that means two things, or does not cover something that arrives daily. This step gets skipped more than any other and it is the one that determines whether decisions will be consistent. Questions come back to you in writing. Three to five days.
Calibration until decisions converge. Around fifty real items, every reviewer deciding independently with no discussion, then the disagreements worked through together until the group lands in the same place for the same reasons. This is repeated on a cycle, not done once, because drift is gradual and invisible from inside. Two to three days initially, then recurring.
The queue itself. Each item gets a decision, a policy reference, a tag and a reason. Anything genuinely ambiguous, or involving a repeat actor or a potential legal issue, goes to a second tier rather than being guessed at. Nothing is actioned without a record. Ongoing.
Daily quality sampling. Someone other than the original reviewer re-checks a sample of decisions every day — around twenty per reviewer — and disagreements are logged rather than quietly corrected, because the pattern in the disagreements is what improves the operation. Daily, roughly half an hour per reviewer.
Welfare and rotation, run as an operational control. Capped shift length, scheduled breaks away from the queue, and rotation off the hardest categories. This is scheduled and audited in the same way throughput is, because when it slips it slips silently and the first visible symptom is people leaving. Ongoing, reviewed weekly.
Monthly reporting. Volume and throughput, precision and recall, the categories generating the most decisions, appeal and reversal rates, and the policy gaps found that month with proposed wording. Monthly.

What moves the price

The range is $700–$3,000/agent/mo and it is ongoing. Three things decide the position.

How hard the decision is. Spam, duplicate listings and obvious fakes sit at the bottom of the range because the rule is nearly mechanical. A queue requiring context, cultural knowledge or judgement about intent sits at the top, because it needs experienced people and they do fewer items an hour.

Language and locale coverage. Every additional language narrows the available pool sharply, and idiomatic or coded language is exactly where automated filters fail and human review earns its cost. Coverage across several locales is a materially different service from coverage in one.

How expensive an error is. A wrongly removed comment on a hobby forum and a wrongly approved listing on a marketplace handling payments are not comparable risks. The higher the cost of a mistake, the more review, sampling and second-tier escalation the operation needs, and that shows in the rate.

What usually goes wrong

A reviewer substitutes their own judgement for the policy. It always sounds reasonable — this one seems fine to me, this one feels wrong — and it is the fastest way to destroy consistency, because now the outcome depends on who picked the item up. Personal opinion has no standing in a queue. If the policy produces a bad result, that gets raised in writing with the example attached, and the queue keeps running on the policy until you change it.

Only one quality number is measured. Reporting removals alone rewards over-removal, and reporting accuracy of removals alone rewards leaving things up. Either can be made to look excellent while the platform gets worse. Precision and recall have to be read together, and an operation that reports only one of them is either not measuring or not telling you.

One person is left on the worst category indefinitely. It is the most efficient-looking roster and the most damaging one. Somebody who spends every shift for months on the hardest material carries real harm from it — sleep, mood, relationships — and pay does not undo that. Rotation, capped exposure and honest disclosure at hiring are conditions of running this work, not benefits attached to it.

See where the site stands before you brief anyone.

If the site is already live, start from what it scores rather than from a guess. The audit reports what it checked, what it found, and — the part most tools leave out — what it could not check, and why.

See the website audit Talk about this work

Questions people ask before they commit.

Can AI not do this now?: It does a large part of it, and that has changed the shape of the work rather than removed it. Automated classifiers clear what is obviously acceptable and stop what is obviously not. What reaches a person is the middle — sarcasm, context, local idiom, a listing that is technically compliant and plainly fraudulent, two clauses of your own policy pointing in opposite directions. That work is harder than what moderators did five years ago, which is why the skilled end of this service costs more than it used to.
Who writes the policy?: You do. We apply it, and we are strict about that boundary because a moderator's personal opinion has no value in a queue and your policy is the thing you can defend. Where we add value is by finding the places your policy contradicts itself or fails to cover something real, and raising those in writing with examples rather than quietly deciding on your behalf.
How do you measure whether moderation is any good?: With two numbers together, never one. Precision is what proportion of the items we removed genuinely deserved removal. Recall is what proportion of the items that deserved removal we actually caught. Someone who removes everything scores perfectly on recall and destroys your platform, and someone who removes nothing scores perfectly on precision and lets it rot. Both figures appear monthly, alongside a sample of the disagreements.
What kind of material will you not take on?: The most severe categories are not accepted without professional counselling support arranged and funded as part of the engagement, and we will say no to that work rather than staff it badly. For everything we do accept, three things are non-negotiable — reviewers are told on day one exactly what they will be looking at, shift length is capped with scheduled breaks, and nobody is left on the hardest category continuously.
How fast will our queue be cleared?: To an agreed target we set after measuring, not before. Review speed depends almost entirely on how hard your decisions are — a spam and duplicate-listing queue moves at a completely different pace from one where each item needs context and cross-checking. We measure throughput and decision quality during a calibration period and then commit to numbers we can hold without pushing accuracy down to hit them.
What happens to the record of decisions?: It stays in your system, and it is easy to underrate until the day it is needed. Every action carries a decision, a policy reference and a reason. When a user appeals, when a regulator asks, or when a removal is challenged publicly, the difference between a defensible position and an embarrassing one is whether that record exists and is consistent.

Often taken together.

Start with the brief, not the invoice.

Send what you are trying to achieve. You get a written scope, a price and a date — or an honest “this is not the right service for that”.

Talk about Content moderation All QA, Support & Ongoing Operations services