Every company is becoming a data company, and foundation model labs are the most aggressive buyers. An entirely new, hybrid role for knowledge workers has emerged: the special project lead (SPL). That title is a catch-all for a person who drives and orchestrates experts at scale, ensuring they produce the right kinds of data assets that models train on. Ironically, the process of corralling experts remains bottlenecked by human effort.
What Is a Special Project Lead?
A special project lead (SPL) is a hybrid operator at a data-labeling vendor who drives and orchestrates domain experts at scale. They own the end-to-end pipeline required to produce specialized, high-fidelity data assets and reasoning traces that AI foundation model labs need to train their models.
A Job Description That Resists Easy Definition
“Even people who accept the offer don't fully know what the role is when they show up on day one,” an ex-SPL source tells me. The role absorbs whatever the project requires.
Here’s how it works. A lab hands the vendor a mandate. For example, they may want 10,000 units of a novel math task in three weeks. The mandate lands on an SPL’s desk. From there, the SPL owns everything between the task and the deadline. Raise a hiring requisition for however many experts the volume implies. Write the task instructions. Stand up the annotation platform.
Are people converting from application to first task? Is each contributor producing two or three tasks a day? The SPL tracks today’s forecast with tomorrow’s and the whole run rate against the promised deadline, all while feeding updates back to the customer. The tight-knit customer relationship is of particular importance as the scope of the work, otherwise known as the spec, can change on a whim.
Expert Data and the Importance of Spec
Expert data is markedly different from its generalist equivalent because many domains have specific language and requirements for a correct answer to train the model with. Additionally, labs want “reasoning traces,” where an expert elucidates sequenced steps in thinking while creating the output. In the math example above, a triple integral might be a good type of problem to solve, where the expert must set up the components, work each stage and only then arrive at an answer versus a simpler problem. The SPL has to think carefully how to set up a process, like choosing the questions, for example, to best get this type of fidelity of output.
This role also requires flexibility, as specs change at any given moment. One SPL described a math project where the customer decided, mid-flight, that valid problems now required three or four reasoning steps instead of two. Every unit produced under the old spec was instantly worthless. Labs differ in temperament. The disciplined ones arrive with detailed specifications and change little. Others hand over a loose first spec and mold the project as it goes, so the work is guaranteed to take on many forms over its lifetime.
Building the Expert Layer
Sourcing experts is the first task of an SPL, particularly at early-stage data labeling vendors with no mature growth function. In some cases, this is as easy as filtering through a list of pre-screened candidates. But the moment a project needs investment bankers or criminology majors, the pipeline runs dry, and SPLs have to get creative. One former banker simply worked their own network, circulating sign-up forms to friends of friends. For the most niche roles, SPLs dig into Reddit threads and the online communities where those experts already gather and recruit them there.
Data labeling is new enough that quality is difficult to measure programmatically, and the infrastructure perpetually lags the demand. One SPL, needing to measure inter-rater agreement, built the entire mechanism by hand on spreadsheets, manually routing each submitted task to three independent reviewers.
As data vendors mature, they build dedicated pods for areas like coding and RL environments. Sourcing niche experts in areas like protein folding can still take up to six months, creating lengthy lead times on new projects. Meanwhile, low-skill tasks are increasingly handled by LLMs themselves, and some vendors have cut their generalist teams by as much as 90 percent. “The models need to hillclimb on better data,” an SPL told me, meaning that lower-quality generalist data is no longer as valuable.
How to Get Into the SPL World
Who might be a good fit for the SPL role? Former founders, early engineers, consultants and extremely detail-oriented operators. Several described the job as the closest thing to founding a company without actually doing it. The operational scale is vast with high ownership. They own a P and L and manage large enterprise contracts.
I kept hearing about a consistent set of desired skills from SPLs and ex-SPLs. First is external communication: the lab-facing half of the job is essentially customer success with the most demanding client in technology. Second is people management, where the contributor-facing half is closer to customer support, fielding thousands of moonlighting experts through onboarding, feedback and attrition. Third, owning technical evaluations, running evals and writing small scripts. Consultants tend to spike on the client-facing work. Engineers automate their own pipelines and excel at evaluation.
Data labeling companies screen for external communication and comfort with hard client deadlines and genuine empathy for the people producing the data day in and day out. For generalists who want into get into the AI field, it may be the single fastest training ground, offering direct exposure to lab relationships and the jagged frontier of model development.
A Boom Market With an Unpredictable Future
How long any of this will last is harder to assess. The data market rests on lab training budgets, which rest on a specific bet about where model improvement comes from. If synthetic data closes the gap or if research taste changes, the demand will shift fast.
The deeper structural problem is that human data revenue isn’t recurring. Contracts are project-based: cash flow can grow month over month while nothing is locked in year over year. This is why many vendors race toward enterprises, building repeatable, contractual revenue outside the labs. The counterargument, made to me by more than one source, is that a close enough relationship with the labs is its own hedge. You can simply pivot toward whichever data sets matter most in any given moment.
For now, the SPLs are quietly building the plumbing of a new gig economy. The models need to hillclimb, and the SPLs ensure that the data steers the model toward climbing up the right hills.
