SkillBench sells $15.1M to measure AI's effect on developer skills

Why it matters

AI coding budgets are growing faster than reliable evidence about their effects. SkillBench is betting enterprises will pay to measure output, human contribution and retained expertise together.

Matt Beane (@mattbeane) and Juho Kim (@imjuhokim) have sold $15.07 million in equity for SkillBench, financing an attempt to answer a question hanging over every engineering organization buying AI coding tools: what is the software actually doing to productivity and developer skill?

A new SEC Form D filing shows that the Santa Barbara, California-based startup sold $15,069,899 toward a $17,144,930 offering. The first sale occurred on July 23rd, 2026, and 20 investors participated. SkillBench had $2,075,031 left to sell when the filing was submitted on August 6th.

The filing describes the financing as an equity offering under Rule 506(b). It does not identify the investors, name a lead or disclose SkillBench's valuation. Asia2G Capital lists SkillBench in its portfolio, though the available records do not establish its role in this offering.

Beane and Kim are building SkillBench around a problem they had already spent years studying: intelligent systems can improve short-term output while changing, and sometimes weakening, the way people develop expertise. That makes SkillBench a founder-led research thesis translated into enterprise software, rather than another analytics dashboard assembled around the latest AI adoption wave.

Measuring what happens between the prompt and the pull request

SkillBench says its platform captures turn-by-turn developer workflow signals and separates human and AI contributions. The intended output is anonymized, auditable analysis that engineering leaders can use to assess productivity, quality and skill development as developers work with Claude Code, Cursor, GitHub Copilot and similar systems.

The emphasis on workflow matters. Repository activity can show that code shipped, a pull request moved faster or an AI tool appeared in a commit. Those measures say less about whether a developer understood the resulting system, caught an agent's mistakes, learned a reusable technique or became dependent on generated output that another engineer will later have to maintain.

SkillBench is entering an increasingly crowded developer-intelligence market. Jellyfish, LinearB and DX already sell tools that connect AI adoption with engineering delivery, developer experience and code attribution. SkillBench's proposed distinction is finer-grained telemetry combined with skill-development analysis and a privacy model designed to avoid individual surveillance.

That distinction also creates the central product challenge. SkillBench needs enough detail to explain how developers and agents divide work without producing a monitoring system that employees experience as performance surveillance. Its claim of anonymized, auditable measurement will have to hold up inside security reviews and in engineering cultures where developers are sensitive to context-free productivity scoring.

The funding arrives while SkillBench is still hiring for production infrastructure. Its recruiting page describes work on multi-tenant services, telemetry pipelines, observability, security controls and assessment systems for upcoming launches. The same page retains a product v1 target of the first quarter of 2026, a date that has passed, while continuing to advertise productionization roles. That points to the practical use of the round: turning research and early enterprise work into software that can withstand regulated customers, sensitive source code and large engineering organizations.

Beane has spent his career studying lost expertise

Beane is an associate professor in UC Santa Barbara's Technology Management Program, where his research examines how people build skills while working with AI and robotics. His university biography identifies him as SkillBench's co-founder and CEO and describes the product as a way for chief technology officers to steer AI adoption while helping developers grow through it.

He earned his Ph.D. from MIT Sloan and previously took two years away from his doctoral work to help found and finance Humatics, an MIT-connected industrial sensing startup. Beane later published "The Skill Code," a 2024 book focused on preserving human ability as intelligent machines take over parts of expert work.

His prior field research gives SkillBench's thesis a longer history than the current coding-agent boom. Beane studied settings including robotic surgery, where technology can place junior workers farther from the hands-on work through which expertise traditionally develops. AI coding agents create a similar organizational risk at software scale: companies can collect immediate output while giving newer engineers fewer chances to form mental models, diagnose failures and learn from senior colleagues.

Kim brings the product and human-computer interaction side of that thesis. He is a KAIST computer science professor, directs the university's KIXLAB interaction lab and serves as SkillBench's CTO, according to his academic site. Kim earned his Ph.D. at MIT after studying at Stanford and Seoul National University. His research covers human-AI interaction and systems designed to make AI more transparent, controllable and collaborative.

The founders previously operated under the name Tacitly, which the SEC filing records as SkillBench's former corporate name. SkillBench was incorporated in Delaware in 2024 and filed an earlier Form D on August 30th of that year.

Reid Hoffman takes a board seat

The financing also reveals a deeper role for Reid Hoffman (@reidhoffman). SkillBench's recruiting materials have named the LinkedIn co-founder alongside Wharton professor Ethan Mollick, former Amazon Worldwide Consumer CEO Jeff Wilke and former Atlassian President Anu Bharadwaj as advisors. The SEC filing goes further, identifying Hoffman as a SkillBench director.

Hoffman's board position gives Beane and Kim an experienced operator and investor at the point where SkillBench must establish a new enterprise software category without being absorbed into generic developer-productivity tooling. The founders need to convince buyers that measuring retained skill deserves budget alongside delivery speed, security and cloud cost.

There is evidence that simple productivity claims deserve scrutiny. In a 2025 randomized study, METR found that 16 experienced open-source developers took 19% longer on 246 tasks when permitted to use early-2025 AI tools, even though the developers believed the tools made them faster. METR cautioned that the result covered one population and one generation of tools. By February 2026, the research group said wider agent use had made clean measurement harder because experienced developers increasingly resisted working without AI.

That measurement gap is the opening SkillBench is pursuing. Engineering leaders already know how much they spend on AI licenses. Beane and Kim are betting that the larger software business will form around proving what those licenses change inside the work itself, including the capabilities organizations may be trading away while chasing faster output.