How Handshake built a $100M AI data business on top of its existing network

Executive overview

AI labs exhausted publicly available internet data for pre-training roughly two years ago. Gains now come almost entirely from post-training — collecting high-quality, expert-produced data to improve model capabilities in specific domains.

Handshake, a decade-old career platform with 18 million students and alumni including 500,000 PhDs, realised its audience was exactly what labs needed. It launched a human data business in January and hit $50M ARR in four months.

The only moat in human data is access to an audience — and Handshake already owned the largest expert network in the world.

What post-training data actually is

• Pre-training fed models the entire internet; gains from that are now largely exhausted.
• Post-training improves specific capability areas: coding, math, law, medicine, biology, finance.
• Labs run experiments like a scientific process — form a hypothesis, collect a small data batch, expand what works.
• Data types include: prompt-response pairs (SFT), preference rankings (RLHF), step-by-step reasoning traces, multimodal (audio, video), and tool-use trajectories.
• A trajectory captures everything a human does to solve a problem — screen, mouse, voice narration — used to teach models how professionals think and act.
• Rubrics let models act as judges in domains without a single correct answer (e.g. educational design, MRI interpretation).

Why experts replaced generalists

• Early labeling used low-cost generalist labor for simple tasks (bounding boxes, basic classification).
• Models improved to the point where generalists can no longer break them — only domain experts can.
• A physics PhD can identify where a model's reasoning fails across sub-domains; an average user cannot.
• Labs now prioritise advanced STEM, then professional domains: accounting, law, medicine, finance.
• Data quality, volume, and speed are the three things model builders care about most.

Handshake's structural advantage

• 18 million professionals, 500,000 PhDs, 3 million master's students — the largest expert network in the world.
• University partnerships with 1,600 institutions mean zero customer acquisition cost for supply.
• Decade of brand trust drives high conversion and retention rates that competitors cannot match.
• Competitors spend tens of millions monthly on performance advertising and 200-person recruiting teams just to find experts.
• Those same competitors then drop experts into a transactional experience with no training — leading to high drop-off.
• Handshake built community-first onboarding: cohort learning, instructional design team, assessments team before any project work.

How the business was discovered and launched

• Middleman companies were already paying Handshake to recruit its PhDs — the company was sending revenue to intermediaries.
• Frontier labs began reaching out directly, trying to cut out the middlemen.
• Garrett spent December doing expert calls with researchers via GLG, Alpha Insights, and similar networks.
• January: launched Handshake AI as a separate internal company with its own team, office space, engineering, design, finance, and recruiting.
• Four months in: $50M ARR. Eight months in: on pace to exceed $100M in year one.
• Working with seven of the top frontier labs.

Running a zero-to-one inside a mature company

• Garrett kept 80%+ of his own time focused on the new business, not delegated to a lieutenant.
• Every person on the new team had zero responsibilities in the core business — one job only.
• Separate all-hands, separate onboarding, separate recruiting, separate engineering.
• DRI (directly responsible individual) assigned per initiative based on capability, not seniority or function.
• Metrics-driven operating cadence from day one — weekly, monthly, quarterly rigor absent from Handshake's early years.
• Hired people comfortable with ambiguity: founders, early-stage operators, people who had never worked at a large company.
• Compensation tied to hurdles in the new business — employees felt like equity holders in a new co.
• Culture built on "leave nothing to chance": late nights, weekend work, public recognition of individual contributions.
• Took top engineers from the core business when scale demanded it — senior staff and principal engineers parachuted in.
• Owned the chaos narrative publicly at all-hands: upfront about expectations, hours, and pace.

The broader opportunity and where this goes

• Human data demand will not disappear until full ASI — lead researchers at labs confirm humans will be needed for at least a decade.
• Data types will evolve: CAD files, scientific tool use, drug discovery workflows, esoteric operating systems, audio, video.
• Synthetic data has a role in verifiable domains but will not dominate — labs consistently confirm this.
• PhDs on the Handshake network earn $100–200/hour; vs $25/hour as teaching assistants.
• AI is not eliminating early-career jobs — employers report one AI-native person can now do the work of a multi-person team.
• Long-term vision: Handshake AI feeds insights back into the core job-matching platform, enabling AI interviewers, skills-based matching, and eliminating the resume-review bottleneck.