Everyone seems to think they know the answer to how AI is going to impact the workforce.
Every researcher, consultancy, and vendor has "cracked" AI's impact on the workforce. Some with open source data; others with a “proprietary framework” for AI workforce transformation, and some with a “five-step plan” for upskilling humanity for the AI age. Hint: none of them have it figured out. Neither do we, yet. The difference is that we are honest about the size of the question.
Which leads to a question I cannot stop pondering lately: What are the actual roots of AI workforce transformation? Every organization on earth is wrestling with this. It comes up in literally every conversation I have now. So where did this problem come from, and how would you even know if you were solving it correctly?
That is a multi-part story. This piece is written about one specific part, and it happens to be the part almost everyone skips: the data you would need to diagnose the problem in the first place. Before the framework, before the roadmap, before the change-management workshop, you have to be able to see what is happening. Right now, everyone is flying blind.
The way I see it, you need four distinct data sources to truly understand AI's impact on the workforce.
The four data sources
1. External labor market data. Job postings and profiles tell you how tasks and skills are shifting across the open market, independent of any single company's spin. This is the macro signal: which work is being redefined, which skills are appearing and disappearing, and how fast. Providers like Lightcast, Revelio Labs, TechWolf, and Reejig live in this layer, each turning billions of postings and profiles into task and skill data.
2. AI usage data from the model providers themselves. This is the clearest read we have on what AI is actually doing to work, as opposed to what people assume it might do someday. So far it skews heavily toward coding, which is exactly where adoption is deepest. The Anthropic Economic Index maps real Claude usage against tasks in the economy, and controlled studies like the GitHub Copilot productivity research put hard numbers on developer speed. These are studies of usage, not surveys of opinion, which is what makes them valuable.
3. Internal work data. How do your own employees use AI tools, how is the company deploying AI at scale to lift productivity, and what is that productivity actually worth? You cannot answer the third question without measuring it, which means you need productivity measurement to sit right next to adoption data. What are the internal tools and use cases for AI adoption in your organization and what signals are they bringing into view? Most organizations still cannot say, with a straight face, how AI changed output per unit of work or worker last quarter.
4. Third-party, non-employee data. Here is the source everyone forgets: Not all work is done by employees anymore. A meaningful and growing share of work runs through contractors, freelancers, and gig platforms, and those workers are adopting AI faster than your payroll ever will. Upwork's In-Demand Skills research, along with signals from platforms like Fiverr, shows AI-related skill demand climbing at triple-digit growth rates. If your picture of AI's impact stops at the edge of your headcount, your picture is wrong.
Looking at the market, no one is publishing comprehensive guidance that requires all four data points - not to mention AI agents' impacts to your org chart and workforce. Should you treat them as a human or a tool? Everyone talks their own book on what data sources they happen to sell. The labor market data firms tell you it is all about external signals. The model providers tell you it is all about usage. The internal tools tell you it is all about your own people. Each is correct and each is incomplete.
Synthesis is the game
We at HRBench see the answer as the synthesis of these data sources, not the dominance of any one of them. We are on the journey to answer this question. No one is at the destination, and as far as we can tell, no one outside of academia is even interested in starting the race, either. Most players just want to take the data they already own and milk what they can out of it, rather than going looking for the insight that only appears when you combine sources.
That instinct, to combine data for insights, is the root of people intelligence. It is the difference between looking in the rearview mirror and looking at the road ahead.
At the core is workforce planning
Underneath all of this talk of AI workforce transformation sits something much older and far less glamorous than AI. It is traditional workforce planning.
At the highest level of abstraction, workforce planning answers one question: How will the work get done? Everything else is in the details that support answering that question.
Traditional workforce planning also combines four things; internal and external supply with internal and external demand. You have internal supply, the workforce you already have, and internal demand, the work that needs doing. You have external supply, the talent available in the market, and external demand, the competition for that same talent in the market. Wrapped around all of it is cost, and not just employee cost, but total workforce management cost across employees, contractors, and contingent workers. Classically, three functions sit on that stool. HR owns the humans, Operations owns getting the work done, and Finance owns the costs. Three legs, one stool.
AI adds a fourth leg. IT owns AI for the firm at large, but HR also feels it needs to own AI for how HR's own work is done, which means HR cannot outsource its understanding of AI to anyone else. And "AI" is not one thing. There are at least three flavors worth tracking separately: personal usage by individuals, end-to-end AI woven throughout a process, and AI-native products that did not exist before. Each carries its own token economics and its own costs, and if you are not tracking the spend, you are not really planning. There is even a live question about whether AI agents belong on the org chart at all. My answer is a toggle view: show me the humans, show me the agents, or show me both, depending on the question I am asking.
This is where the vendor landscape falls down. Most vendors chase a single branch of this tree, skills, or tasks, or external data, and call the branch the whole tree. The result is that organizations today have pretty decent facts about their humans and almost no facts about their AI. It reminds me of the pre-internet era, when companies genuinely struggled to track their own workforce across locations using paper files and filing cabinets and fax machines. We are back in the same “information fog", except this time the thing we cannot see is the fastest-moving part of the operation: AI’s impact on the work and workforce.
Where HRBench sits
We are not trying to be every vendor combined. That is a losing game. We are trying to be the singular point of intelligence that sits on top of all of it.
Concretely, HRBench is the data layer and the intelligence layer that combine information across systems: internal and external supply and demand, the AI components of work, the org charts, the total workforce management across every worker type, and the token economics that come with all that AI. Our focus is deliberately narrow in one respect. We care about AI that helps HR execute its role in AI workforce transformation, not just for AI but for the entire firm’s workforce.
What good workforce planning actually requires
Skills and tasks are necessary, but they are not sufficient. Good workforce planning has always run on units of productivity achieved and standard ratios (such as revenue per employee, etc.). How many units of these does it take to produce one of those? AI does not exempt us from that discipline. It demands a new ratioed methodology, an AI-aware version of the same logic that has guided workforce planning for decades.
And the workforce plan is not the point. Execution of the workforce plan so the business can function is where value is created or lost. That means upskilling for workers, deliberate deployment of AI agents into real workflows, and the deeply unsexy (or sexy?) work of change management. A perfect plan with no execution is just a prettier rearview mirror.
The organizational network analysis opportunity
Org charts tell you how work is supposed to flow. Organizational network analysis (ONA) tells you how work actually gets done, through the real social networks inside a company. ONA shows you who collaborates, who quietly blocks productivity, and who enables high performance for everyone around them. That is gold for workforce planning, and it becomes even more interesting when you ask where AI agents sit in that network and how they reshape it.
The existing ONA vendors do good work but tend to struggle with the same thing: they lack organizational context. A network without the org chart, the roles, the costs, and the AI layer around it is a picture missing its frame. Combine ONA with that organizational context, and add AI agents into the network as first-class participants, and you have something genuinely new. As far as we can tell, no vendors have seized this opportunity and it is there for the taking.
The honest version of the answer
So, what data is needed to understand AI's impact on the workforce? External labor market data, AI usage data from the providers, internal work and productivity data, third-party non-employee data, and tokenomics, with organizational network analysis layered on top of org charts to show how the work really moves. Many data sources, plus the network that connects them, are synthesized and grounded in the workforce planning discipline rather than floating free of it.
No one has assembled all of that yet. The race starts the moment you stop pretending the one dataset you already own is the whole story. Most of the field has not started running. We are.
