Upcraft Buy Brings Agentic Sales Automation
The bottleneck in frontier AI has shifted from compute to people. Labs racing toward superintelligence spend billions on post-training and reinforcement learning, but the experts who evaluate model outputs, write preference data, and stress-test reasoning capabilities remain stubbornly hard to reach at scale. Labelbox's answer arrived February 10, 2026: the company acquired Upcraft, a Chicago startup that builds AI agents to automate sales outreach, qualification, and engagement. Terms were not disclosed.
Upcraft brings different DNA. Founded in 2021 by Greg Caplan, the startup spent nearly five years teaching AI agents to run complex sales workflows (prospecting, qualifying, nurturing) without human operators in the loop. Caplan described the move as bringing "our AI sales agent expertise to Labelbox" to "contribute to a platform with unmatched resources and reach." Manu Sharma, Labelbox's CEO, framed the logic more bluntly: "Building frontier AI requires connecting elite domain experts with development teams at scale. Upcraft's AI agent expertise will transform how we grow and operate the Alignerr network."
The integration target is specific. Alignerr, Labelbox's expert network, now spans 2.6 million domain specialists across 30-plus languages, up from "over 1 million" cited in the February acquisition announcement. Its growth has relied on manual recruiting, screening, and onboarding of specialists (physicians, physicists, coders, linguists) who then produce the high-quality training data that defines the cutting edge of AI. Upcraft's agents are being repurposed to automate that pipeline: identifying candidates, initiating outreach, assessing fit, and managing engagement cycles.
Sharma's quote hints at the stakes: "AI labs invest billions in post training and reinforcement learning workflows." The labs he references — OpenAI, Anthropic, and their peers — are the same ones posting hundreds of six-figure research roles on specialized job boards. They need data, not demos. Whether agentic automation can deliver the right experts fast enough without diluting quality is the question the next quarter will answer.
Why the Finance Hub Landed in India
Labelbox's finance and operations function now accounts for 39.4 percent of its 386-person workforce, the largest single group inside the company, and it is the only function still expanding, up 1.5 percentage points in the latest Revelio Labs workforce snapshot dated March 2026. That growth is not happening in San Francisco. South Asia represents 12.3 percent of total headcount, a share that has risen alongside the finance build-out, while North America's portion has held at 51.3 percent. The median salary in South Asia sits at $11,000 versus $150,000 in North America, a gap that makes a dedicated India hub economical for the volume of transaction processing frontier AI labs now require.
| Role | Salary Band |
|---|---|
| VP of Research | $453,600–$567,000 |
| Director of Engineering, Physical AI | $302,400–$378,000 |
| Principal Architect, public sector | $298,400–$373,000 |
According to Revelio Labs, active job postings fell 58.8 percent year over year to just 12 listings. Revelio Labs estimates Labelbox's overall hiring velocity has slowed to five new roles per month in 2026, down from 12 in 2025 and 21 in 2024. Yet the finance share keeps climbing. Employee sentiment, tracked by the same workforce intelligence platform, is negative and declining. Engineering remains the second-largest group at 41 percent, but its growth has flattened. Sales and marketing sits at 19.6 percent.
From Email Threads to Programmable Workflow
Labelbox's acquisition of Upcraft and its parallel build-out of a finance operations hub in India converge on a single bottleneck: the time between an AI lab requesting specialized expert data and that data arriving in a training run. The Upcraft agents automate the outreach, qualification, and scheduling that previously required human recruiters to match a lab's niche requirement — say, board-certified radiologists fluent in Mandarin for a multimodal evaluation — to the right experts in the Alignerr network. Meanwhile, the India finance hub handles the compliance, invoicing, and cross-border payment rails that used to stall onboarding for weeks.
The throughput gains show up in delivered projects. For Meta's GIM benchmark, Labelbox produced 820 expert-authored problems across seven cognitive categories, including 229 multimodal items and 528 rubric-graded prompts, with original prompt creation, structured scoring criteria, review, annotation, and quality assurance, all feeding a 2PL IRT model calibrated over more than 200,000 prompt-response pairs. That scale of structured, expert-graded data is the direct output of a pipeline that can recruit, onboard, task, and pay thousands of specialists in parallel. The same infrastructure powers the labeling services that frontier labs buy daily: RLHF preference ranking, supervised fine-tuning data, multimodal LLM evaluation, red-teaming, coding and agent-task annotation, and text-to-image/video/audio generation tasks. Each workflow demands different expert profiles, different rubric designs, and different quality gates.
OpenAI and Anthropic, which together posted 140 new research and engineering roles on Zero G Talent in the past week alone, are the most visible consumers of this capacity. Their hiring velocity signals model development cycles that demand fresh expert data weekly, not quarterly. Labelbox's platform, with Foundry integrating foundational models directly into the labeling workflow, lets those labs iterate preference data against live model outputs, close the RLHF loop, and ship the next checkpoint without waiting for a vendor to spin up a custom annotation campaign.
Horizon, Labelbox's RL environment product, ships preference signals tuned for reasoning, tool use, and computer use across autonomous research, scientific knowledge work, agent coding, and cybersecurity. Terra delivers robotics foundation-model data: video, trajectories, multimodal annotations collected with purpose-built hardware. Recursion closes the loop for enterprise specialist models with RL training on evaluation reward signals. All three product lines draw on the Alignerr network.
Revelio Labs' figures put Labelbox's headcount growth at 93 percent from 200 to 386 between 2023 and March 2026, with finance and operations now comprising 39.4 percent of the workforce, nearly equal to engineering at 41 percent.
Three Ways to Scale Expert Data
This acquisition and its India finance hub build-out arrive as the two incumbents in AI data services, Appen and Scale AI, are making leadership and capital-allocation moves of their own.
Scale AI, which Revelio Labs' data shows employed 6,015 people (roughly 15 times Labelbox's headcount), added 13 roles in the past seven days alone, including a VP of Research banded at $453,600–$567,000 per Zero G Talent's board data, a Director of Engineering for Physical AI at $302,400–$378,000, and a Principal Architect for public-sector work at $298,400–$373,000. The company's overall salary band ($80,000–$331,000 with a $249,000 median) reflects a workforce still heavily weighted toward engineering and operations.
Appen, the Australian-listed veteran with 30 years in human data, emphasizes "expert human data for frontier AI" and its ability to help teams "build, train, and deploy", language that mirrors Labelbox's Alignerr narrative.
The contrast in strategies is instructive. Labelbox is betting that agentic sales automation (via Upcraft) paired with a dedicated finance-operations hub can compress the cycle from expert identification to paid contribution. Scale AI's response leans into leadership credibility and high-value research hires, while Appen leans into incumbent volume. Neither rival has publicly matched the finance-hub model.
What emerges is a three-way race where the differentiator may not be annotation quality — all three serve over 80 percent of leading U.S. AI labs — but the unit economics of expert engagement. It already constitutes that share, a ratio that dwarfs typical SaaS benchmarks and hints at the operational intensity of managing a million-expert network.
Kicker
Labelbox's bet is that the next frontier of AI won't be won by the lab with the most GPUs, but by the one that can summon the right human judgment, on demand, at the speed of a training run. They and the India finance hub are the plumbing. The experts are the fuel.
Working in frontier tech? Zero G Talent tracks the openings: see every open OpenAI role, browse frontier tech jobs, openings at Anthropic and Scale AI, and the people building the field.