The week in AI training work: Innodata's record quarter, TELUS's AI data dip, agentic roles keep growing

Law, Medicine & Finance remained the largest kind of AI training work, with 1,622 open roles on the board at the end of 8 Nov. On the comparable platforms it grew from 1,579 to 1,610 (+2.0%), with 161 new listings against 130 removals. Math and STEM fell furthest among the larger kinds, from 257 to 238 (-7.4%), with only 3 new listings against 22 removals. Multilingual work went from 862 to 822 (-4.6%) and transcription, audio and speech from 635 to 613 (-3.5%), while coding and software evaluation held at 327 and AI evaluation and rating at 143.

Agentic tasks and RL environments grew for a second week, from 129 to 146 (+13.2%), with 53 new listings against 36 removals. Robotics and physical-AI data went from 98 to 107 (+9.2%), and red-teaming and AI safety from 79 to 82, including AfterQuery's first 5 red-teaming roles, opened on 4 Nov.

Among the comparable platforms, micro1 moved most: 91 new listings and 22 removals took it from 298 to 367 open roles. DataAnnotation went from 112 to 133 with 21 new listings and none removed. Alignerr went from 396 to 334 (52 new, 114 removed), RWS TrainAI Community from 148 to 118, Turing from 227 to 211 and OpenTrain AI from 51 to 30. No country went from no open roles to some, or from some to none, on these platforms.

The median advertised hourly rate of the open roles on the comparable platforms held at $40. Within the kinds, Law, Medicine & Finance rose from $70 to $75, red-teaming and AI safety from $60 to $70, coding from $60 to $62.50 and data annotation from $22.50 to $25, while writing eased from $52.50 to $50. Multilingual work held at $32.50, transcription and audio at $30, math and STEM at $40, agentic work at $50 and robotics data at $10. A median moves when the mix of open roles moves, so these say which roles were open, not that anyone got a raise.

Some of the week's new programs advertised more than the same platform's earlier roles. micro1's 7 new coding roles of 3 Nov advertise a typical $100/hr, against $65/hr on its coding roles already open, and AfterQuery's first red-teaming roles a typical $150/hr. A posted rate is what a role advertises, not what anyone earns.

On the platforms we already read, the hourly detection caught new professions and new kinds of work: DataAnnotation opened 11 roles in medicine and nursing on 7 Nov, from sonographers to surgical clinicians, at a typical $80/hr; Mercor 9 roles in engineering outside software on 5 Nov, at a typical $80/hr; AfterQuery its first 5 red-teaming roles and 2 roles in nuclear, chemical and biological defense on 4 Nov; and Alignerr its first 1 role in earth and environmental science.

Some of the week's changes were in how we read the market, not in the market, and those roles count on neither side of the comparison. Scale AI's Outlier postings carried on the Joveo job-ad network, which we read from 30 Oct, left our feeds from 4 Nov, 52 roles. Weekday's board was read more narrowly from 2 Nov, 48 roles fewer, and Welo Data's from 4 Nov, 21 fewer. On 8 Nov we began reading every post in micro1's sitemap instead of only its first 400, 11 roles more, and the Post-Training department of xAI's board, where it files its design, humanities and writing tutor posts, 3 more.

On 7 Nov 2026, 15 of the 25 AI labs whose public job boards we read had 87 roles open on the teams that buy, run and evaluate human data, 2.4% of their 3,572 open roles, against 102 on 2 Nov. xAI still had the most, 14 of its 303, down from 26 a week earlier.

Evaluations are now the most common of these jobs, 23 roles at 10 labs; a week earlier AI tutors and raters led with 28. 6 roles buy or license data, at Mistral AI, OpenAI, Reflection AI and Thinking Machines Lab. 39 of the staff roles state a yearly salary range, the middle of the median range being $284K, and the 20 contract roles that state an hourly rate pay $18 to $100 an hour. This week's podcast episode is about how reviewers grade AI training work, and what gets a task rejected.