THE PROBLEM 

Degrees are losing their signal. Job titles are poor proxies for capability. And the half-life of any given skill is now measured in years, not decades. Enterprises increasingly need to understand not just who their people are, but what they can actually do—and how to deploy, develop, and retain them based on that reality. Yet most large organizations have no coherent, up-to-date map of the skills that live inside them. They are operating in the dark on their most important strategic resource. 

THE OPPORTUNITY 

The opportunity is in building the skills infrastructure layer for the enterprise: AI-powered platforms that continuously infer, validate, and map employee skills from real work artifacts—code commits, documents, project outcomes, communication patterns—rather than self-reported profiles that go stale the moment they're filed. The startups that build a dynamic, defensible skills graph for large organizations will become the operating system for talent mobility, succession planning, and learning investment. For the savvy investor, this is foundational HR infrastructure with the kind of data moat that compounds with every customer. 

Analysis & Implications 

IBM has 280,000 employees. JPMorgan Chase has 300,000. Walmart has 1.6 million. Every one of these organizations has, somewhere inside it, the specific talent it needs for its next strategic initiative. Most of them cannot find it—not because the information doesn't exist, but because they have no system that synthesizes it into an actionable picture. What they have instead is a skills profile that an employee filled out in Workday at onboarding and never updated. It says they're proficient in Java. They've spent the last three years building Kubernetes infrastructure. The system doesn't know that. 

The cost of this blindness is enormous and largely invisible because nobody measures it. A global consulting firm hires an external contractor for a machine learning project while an employee three floors away has spent two years doing exactly that work. A bank creates a transformation program around Python analytics while laying off employees who have built Python analytics systems for years. These mismatches happen constantly, compounding in wasted spend and in the disengagement of capable people who cannot find work that uses their actual capabilities. 

The solution is not better self-reporting. Humans are reliably poor at representing their own skills accurately—because of self-assessment bias, because the taxonomy of skills changes faster than any survey cycle can capture, and because nobody wants to spend twenty minutes updating a profile that nobody reads. The solution is inference at the system level: models that read the artifacts of real work and surface skill signals automatically. 

The signals are there. A software engineer's GitHub commit history reveals programming languages, frameworks, code quality patterns, and project complexity over time. A consultant's document library reveals the industries worked on, the analytical frameworks applied, and the kinds of problems structured and solved. A marketer's campaign history reveals channels, audience types, testing methodology, and performance outcomes. None of this requires asking anyone anything. It exists in systems the enterprise already operates. 

The early players—Degreed, EdCast, Cornerstone—have approached this primarily from the learning side: showing employees what skills to develop rather than cataloging the skills they already have. That's valuable, but it's the less urgent half. The urgent half is talent deployment: helping the enterprise find the capabilities it already owns when it needs them. That's the system that gets budgeted on the spot when you show a CHRO a single example of a capability miss that costs them a significant contract. 

If you're building here, start with software engineering, where the skill taxonomy is well-defined, and work artifacts are digital and accessible. Build a tool that a technology VP can use to understand their team's actual capabilities in thirty seconds—materially better than anything they can do today. That use case sells itself. The rest of the platform follows.