MIT CISR at Johnson & Johnson: What a 130,000-Person Skills Enablement Program Looks Like

See also (wiki): training-architecture

⚠️ TEMPORAL NOTE: Primary evidence in this file predates current model capabilities. The underlying MIT CISR Working Paper No. 461 is Nov 2023 (Tier 4) with an April 2024 Research Briefing (Tier 3) supplying the 20% learning uplift, 90%+ reach, and 41-skill taxonomy figures. The program architecture (taxonomy-first, capped use cases, employee self-assessment before AI-inferred score) remains operational and transferable; quantitative adoption figures should be read against a prior model generation and pre-agentic L&D tooling. Check queue.md for newer MIT CISR workforce evidence; van der Meulen track has continued into 2025-2026.

Executive Summary

• Johnson & Johnson built MySkills, a machine-learning talent platform that infers employee skill proficiency from four existing systems (HRIS, recruiting, LMS, project management) rather than asking employees to fill out another survey. The starting pilot covered 4,000 technologists; the foundation now extends across 130,000+ employees.
• The program’s headline behavioral result is simple: after the first skills-inference round, voluntary learning activity rose 20%, and by March 2024, more than 90% of the J&J Technology group had accessed J&J Learn. No mandate. The architecture did the work.
• The skills taxonomy, not the model, carried the weight: 41 future-ready skills, 11 capability groupings, validated by 100+ senior leaders before a single prediction was run. MIT CISR’s own guidance: establish the foundation before deploying AI systems.
• Trust was engineered into the interface. Employees submitted their own proficiency estimates first; the AI-inferred score appeared after, alongside an “agreement score.” Opt-out was offered. Use cases were deliberately capped at two — employee development and aggregated workforce planning — to prevent scope creep into performance management or surveillance.
• For mid-market CHROs and CLOs: most of this is transferable at a fraction of J&J’s scale. The bottleneck is not the ML model. It is the taxonomy work, the governance guardrails, and the decision to bound use cases before launch.

Why This Case Matters for the Training Corpus

Most enterprise AI upskilling evidence is either aggregate benchmark data (ATD, Deloitte, BCG’s workforce studies) or narrow tool-adoption case studies (Colgate, Citi, IKEA on specific Copilot or ChatGPT rollouts). The J&J case fills a specific gap: a large-enterprise program that is not about training employees on AI tools, but about using AI to run the training function itself — infer skill gaps, personalize development, and feed strategic workforce planning.

That distinction matters. A mid-market CHRO staring at a Copilot rollout has a tool-adoption problem. A mid-market CHRO staring at a 2,000-person workforce with unclear skill inventory, uneven L&D participation, and no way to answer “who can do what right now” has an architecture problem. MySkills is an architecture answer.

The program is authored by Nick van der Meulen and collaborators at MIT CISR — the same research center that produced the “Enterprise AI Maturity” (2025) and “Business Models in the AI Era” (Peter Weill, 2025) work already in the corpus. This case sits on the workforce track of that research program.

The Three-Step Architecture

MIT CISR describes the process as skills inference, and the order is non-negotiable.

Step 1 — Taxonomy. J&J’s Digital Talent team built a taxonomy of 41 future-ready skills grouped into 11 capability clusters, derived from strategic plans and industry benchmarks. More than 100 senior leaders validated it; subject matter experts refined it. No model training began until the taxonomy was signed off. MIT CISR’s explicit recommendation for organizations following this path: “Gather expert input to create a skills taxonomy aligned with organizational purpose and strategy. Establish this foundation before deploying AI systems.”

Step 2 — Evidence gathering. Four data sources fed the model: the HRIS, the recruiting database, the LMS, and the project management platform. No new surveys. No new fields. Employees were encouraged to update existing records to sharpen accuracy. The target coverage was 60-70% of each employee’s skills visible from data.

Step 3 — Assessment. A natural-language-processing model scored proficiency on a 0-5 scale (5 = thought leadership). Validation came from two sources: employee self-assessments and manager evaluations. The engineering target was a deviation of no more than one point between AI-inferred and human-assessed scores.

The interface choice is the detail most worth copying: the AI-inferred score is hidden until the employee submits their own. The system then shows an “agreement score” measuring consistency. This sequence does three things at once — it reduces anchoring bias, it gives the employee authorship of their own profile, and it converts disagreement into a coaching signal rather than an evaluation dispute.

What Changed in the Workforce

The outcomes MIT CISR reports from the first rounds:

• 20% increase in voluntary learning activities after the initial skills inference
• 90%+ of J&J Technology group employees accessed J&J Learn by March 2024
• Enhanced hiring and retention
• Improved internal talent movement and advancement
• Broader adoption of continuous learning platforms

These are behavior metrics, not sentiment metrics. That matters. Most corporate L&D case studies report satisfaction scores. The J&J reporting is about whether employees showed up to learn something they were not required to.

The Governance Decisions That Made It Work

The program is structurally conservative in ways that are easy to miss and impossible to retrofit.

Bounded use cases. Only two were permitted: employee development and strategic workforce planning. Performance management, compensation decisions, layoff prioritization — all excluded by design. The bounds were announced, not discovered.

Opt-out existed. Employees could decline. The platform remained useful to those who participated because the 60-70% data coverage target came from existing systems, not new employee input.

Transparency. The AI use was disclosed. The data sources were disclosed. The intent was disclosed.

Deidentified aggregate reporting. Executive dashboards showed geography and line-of-business heat maps. They did not show named employees.

Data governance partnership. HR data experts and external partners co-owned the data quality and privacy layer. The ML team did not get unmediated access.

This is what MIT CISR means by “prioritize human factors over technology.” The model was the last component built.

What to Copy, What Not to Copy

For a mid-market organization with 500-5,000 employees, the full J&J architecture is overbuilt. The transferable elements:

The wrong lesson is “buy the platform.” The right lesson is “do the taxonomy work and bound the use cases before you buy any platform.”

Key Data Points

Publication dates: Working Paper 461 — November 7, 2023 (Tier 4: predates current model generation; the taxonomy, governance, and UX findings remain operational; the ML specifics have since been commoditized). Companion briefing — April 18, 2024 (Tier 3: published during prior model generation). The structural findings are architectural, not model-dependent, and remain relevant.

What This Means for Your Organization

If you are a CHRO or CLO considering a skills platform, the J&J case points to three decisions that happen before vendor selection. First, the taxonomy. Who owns it, who validates it, and whether it reflects where the business is going — not where it has been. Second, the use-case perimeter. Decide in advance what the data will and will not be used for, and publish it. If performance management creeps in later, the program loses trust and the data degrades. Third, the interface sequence. Employee self-assessment first, AI score second, agreement score as the coaching artifact. This is the single cheapest element of the J&J design and the one most frequently inverted in off-the-shelf implementations.

For a 1,000-person organization, the budget to run an equivalent program is not the $40M-range investment J&J likely made. It is the cost of a skills-cloud license ($15-40 per employee per year for the major platforms as of 2026), two quarters of HR/IT taxonomy work, and the discipline to bound the use cases. The gating factor is organizational, not technical.

If this raised questions specific to your workforce architecture or how to sequence the taxonomy work before a platform decision, I’d welcome the conversation — brandon@brandonsneider.com.

Sources

• van der Meulen, N., Tona, O., Someh, I. A., Wixom, B. H., Leidner, D. E. “Developing a Digital-First Workforce: AI-Driven Skills Enablement at Johnson & Johnson.” MIT CISR Working Paper No. 461, November 7, 2023. https://cisr.mit.edu/publication/MIT_CISRwp461_JohnsonandJohnsonAIDrivenSkills_VanderMeulenTonaSomehWixomLeidner — Credibility: HIGH (peer research center, named authors, named executive sponsor at J&J). Tier 4 by publication date; architectural findings remain operational.
• van der Meulen, N., Tona, O., Leidner, D. E. “Resolving Workforce Skills Gaps with AI-Powered Insights.” MIT CISR Research Briefing, April 18, 2024. https://cisr.mit.edu/publication/2024_0401_DigitalTalentTransformation_VanderMeulenTonaLeidner — Credibility: HIGH. Tier 3. Primary source for the 20% learning uplift, 90%+ reach, 41-skill taxonomy, and three-step architecture.
• van der Meulen, N. “Unleashing Workforce Potential with AI.” MIT CISR Board & C-Suite Online Summit, December 10-11, 2024. https://cisr.mit.edu/publication/2024_1211_WorkforceAI_Meulen — Member-gated; confirms continuing research track through late 2024.
• MIT CISR 2022 survey on workforce retraining (n=342 leaders), cited in the April 2024 briefing for the 38% retraining/replacement estimate within three years.

Brandon Sneider | brandon@brandonsneider.com April 2026