Why Task Queues Run Dry: The Economics Behind AI Training's Biggest Frustration

2026-09-10 · Expert Match AI team

Read any worker forum for any AI training platform and the top complaint is the same, and it is not the rate. It is the empty queue. You pass the assessment, you get onboarded, you do a week of tasks - and then the dashboard says "no projects available" for eleven days. Outlier workers call it feast-or-famine. DataAnnotation workers report unexplained quiet stretches. Mercor's own communities have a name for it.

We track ten platforms and every one of them gets this complaint, from the best-reviewed to the worst. That tells you it is not a bad-platform problem. It is how the market works. This post explains the mechanism, with the churn numbers from our own board to back it up.

The work is project-shaped, not job-shaped

A lab does not hire AI trainers the way a company hires staff. It commissions a dataset: 40,000 graded responses in a domain, a benchmark of 2,000 hard problems, a batch of transcribed audio in one language. The platform recruits a pool for that batch, the pool works through it, the batch closes, and the queue goes quiet until the next commission lands.

From the worker's side, that looks like being hired and then abandoned. From the lab's side, nothing went wrong - the project finished. Nobody on the platform side is obliged to have your next project ready, and usually nobody does.

Three consequences follow, and they explain almost every "why is my queue empty" thread we have read.

Pools are deliberately over-recruited. A lab that needs 300 active graders will ask the platform for 900 approved ones, because attrition is high and quality filtering removes a share after the first tasks. Being approved means you are in the pool. It does not mean you are in the 300.

Batches are lumpy. Work arrives in floods because that is how commissions arrive. Two projects landing in the same week means everyone is busy; two closing in the same week means everyone is idle. The averages you read about (hours per week, dollars per month) are smoothed over months of this, and no individual week looks like the average.

Quality gates run silently. Most platforms route the next batch to the graders whose earlier work scored best. If your queue went quiet while a colleague's did not, that may be the reason, and almost no platform will tell you.

What the churn looks like from where we sit

We have recomputed our board every six hours since July 2, which gives us a direct view of listings appearing and disappearing - the same dynamic that empties a task queue, one level up.

It is fast. On the September 7 snapshot, 75 of the 236 priced roles on the board had first appeared in the previous seven days - nearly a third of the whole market, replaced in a week. Three days later, on September 10, the live count had dropped from 330 roles to 316. Fourteen listings closed in three days on a board of ten platforms, and the newest arrivals are dominated by Alignerr transcription batches at $10–35/hr and micro1 data-collection listings at $8–17 - exactly the project-shaped, batch-sized work that will close again once the dataset is delivered.

The other direction is just as telling. In our August trend check, the cohort of listings we had first seen before August had shrunk from 134 roles to 58 in sixteen days; by September 7 it was 51. The ones that survived are the evergreen premium tiers - Surge's physician and attorney queues, Mercor's talent networks, Braintrust's senior engineering bands - which are open continuously because they are recruiting pools for whatever comes next, not for one batch. Cheap listings churn; premium standing pools stay open. That is the single most useful pattern in our data for someone deciding where to build.

One more: Mindrift's priced listings went from 9 to 3 between the July and August reports. That is one platform's queue drying up, visible from the outside.

Which queues dry up fastest

Putting the churn data next to the forum record, the pattern is consistent.

Fastest to empty: generalist, low-credential, high-volume work. Annotation, transcription, video and data collection, generic "AI trainer" roles at $15–35/hr. These are the batches a lab commissions most often and closes most abruptly, and the pools are the most over-recruited because anyone can qualify. This is where the Outlier and DataAnnotation complaints come from, and it is where the 75-roles-in-a-week turnover on our board lives.

Slowest to empty: credentialed standing pools. Physician, attorney, PhD and senior-engineer tiers at $100–450/hr are not recruited per batch. They are recruited once and drawn on repeatedly, because labs cannot re-source a board-certified oncologist every time a project starts. The catch is on the way in: these are the queues that take months to accept you, and some (Surge) never say no, they just never say anything.

In between: mid-band specialist work - the $50–140/hr software, product, analyst and clinical-review roles that make up the bulk of micro1's board. Steadier than annotation, lumpier than the credentialed tiers. This is where most applicants with a degree and some experience will actually land, and our micro1 analysis shows why volume there does not translate into ease.

Why stacking is the rational response, not a coping strategy

If you take the mechanism seriously, one conclusion follows. No single platform can promise you continuous work, because no single platform controls when its labs commission the next batch. The only way to smooth a lumpy income is to draw on several uncorrelated lumps at once.

That is what working two or three platforms does, and it is why the workers who report steady income in this market have all converged on it. An anchor platform for baseline volume, a specialist bet where your credential commands a premium standing pool, and a long-shot application at the top. When the anchor's batch closes, the specialist queue may be mid-project; when both are quiet, the assessment you passed at the third one two months ago finally pays off.

Two things stacking does not fix. It does not turn project income into a salary - three lumpy streams are smoother than one, but they are still lumpy, and the first-90-days plan sets honest expectations by month. And it does not fix a silent quality gate: if one platform's queue has been empty for a month while its forum is busy, the most likely explanation is routing, and the fix is to improve on the next batch you do get, not to refresh harder.

What to do when your queue is empty

The jobs board shows what arrived this week, the salary report recomputes every six hours, and uploading a resume on the homepage ranks every live role against your background so you can see which standing pools you already qualify for - free, no account, resume never stored.

Published by the Expert Match AI team. Listing counts and churn figures are from our own tracker, which began July 2, 2026; platform availability observations reflect public worker reviews - see each platform page for sources. Pay figures are listed rates, not accepted offers. Some outbound application links on this site carry disclosed referral codes; rankings and recommendations are never influenced by them.

Live roles right now

AI Facial Data Collection Associate (Twins)NEWmicro1 · Remote worldwide · AI training project
ML, Data & AI
see listing
Associate Forward Deployed EngineerNEWmicro1 · Remote worldwide · AI training project
Software
$40–$60/hr
Bilingual Mental Health ExpertNEWmicro1 · Remote worldwide · AI training project
Science
$100–$200/hr
Equity research ExpertNEWmicro1 · Remote worldwide · AI training project
Finance
$175–$300/hr

See all 316 matching live roles → refreshed every 6 hours

Browse live roles →See the salary report