From Cost-per-Head to Capability Yield

Why the most trusted number in workforce management no longer measures what we assume it does, and what people analytics should measure instead.

The most consequential number in your company may be one nobody questions: total cost of workforce per full-time equivalent (TCOW per FTE). It sits in the budget, the workforce plan, the board deck, and the diligence data room (the file room a buyer combs through when a company is bought or sold), and it has earned its place through decades of dependable service. I have spent much of my career as a labor economist and people analytics leader building analyses, dashboards, and prescriptions on top of it. Which is why it matters to say plainly: TCOW per FTE has quietly stopped meaning what we think it does.

The disconnect has already been noticed, in two places that rarely read each other. In software finance, practitioners have been documenting the symptom for over a year: Subscript (2025) showed that revenue per employee misleads exactly where it is most used, in board decks and diligence rooms, because vendors, contractors, and AI now produce output that never enters the denominator. In people analytics, the issue has been raised for years, most recently and sharply by Colby Nesbitt and Yuyan Sun (Nesbitt & Sun, 2026), two distinctive voices shaping the future of people analytics.

The missing link is the connection between the two. Finance holds the symptom; people analytics holds the reframing. Neither, however, has followed the distortion down to the actual numbers a company is valued on. This article attempts to build that bridge with a nod to an unsung expert who bestrides the worlds of private equity and HR.

One number, two jobs

The mechanism that follows is Garrett Walker's. He is my colleague on The Conference Board's agentic AI expert panel, and I had the privilege of reading his recent paper in draft. He credits me in the acknowledgments for that reading and for the return-on-workforce formulation that appears in his Section 2 (Walker, 2026a). I set out his argument at some length here because it deserves a wider audience, and because what I propose in the second half of this article, a successor metric and a job for people analytics, only holds if the first half is right.

“Cost per head [the everyday name for TCOW per FTE] has always done two jobs at once…It priced the input: what a head costs in cash, fully loaded—the number payroll, budgeting and compensation benchmarking require. And it stood in as a proxy for the output: what head produces” (Walker, 2026a). One person, one stable bundle of work. Because the two moved together, one number could do both jobs, and for a hundred years nobody needed to notice it was doing two.

Every ratio we manage a business by is built on that weld. TCOW per FTE (workforce cost per unit of labor). Output per FTE (productivity per unit of labor). And the one I suggested to Walker as he revised the paper: divide output per FTE over TCOW per FTE and the FTEs cancel out, leaving output per dollar of workforce cost, more familiarly, the return on the workforce. Each ratio works only as long as one unit of labor—an FTE, a headcount, a job—reliably equals one unit of work.

The quiet severing

Agentic AI breaks the weld. When software takes over the mechanical part of a role—the scheduling, the data checks, the first drafts, and the routine requests—the salary line does not move. Yet the bundle of work behind it has shrunk. What stays with the person is the judgment: the calls, the exceptions, the relationships. The routine half now lives in software, and the P&L (the profit-and-loss statement) records it somewhere else entirely. The cash measure is still perfectly right. The output proxy has lost its anchor. As Walker puts it: “cost-per-head now prices the role, not the work.”

Walker names this event an unissued restatement: a change in a measurement definition material enough that, had an accounting standard-setter issued it, every finance function would have been compelled to trace the impact through every downstream calculation and restate what was material. He is precise about the category. It is not an error, not a change in estimate, and not a change in principle, but "a change in the basis on which a reported figure is measured … enacted in the underlying activity rather than announced by a standard-setter" (Walker, 2026a). No one issued it. It arrived without announcement, one automated task at a time, with no formal notice and no start date. The number stayed on every dashboard while the thing it measured changed underneath it.

The distortion does not stay in the workforce line; it flows into numbers finance treats as hard. The distortion is quiet because of how the accounting works. Overhead rates are set once a year, in advance: budgeted pool over budgeted hours. When agents take hours out of the base, the first thing that surfaces is not a higher rate but an unfavorable variance: less overhead was charged out than was actually incurred, and the gap drops to the P&L. At the close, it gets explained the way volume misses are always explained: soft demand, unfavorable mix, a slow quarter. Nobody codes it as automation. The next year the rate is reset on the smaller base, and the higher cost stops being a variance to explain and becomes the standard everything else is measured against.

From there it is no longer a workforce question at all. The rate sets price floors, decides make-versus-buy against an outside quote, flags idle capacity for consolidation, and determines which lines and segments earn their keep. The controller (the senior finance manager who owns the company’s books) running that analysis sees a rate, a variance, and a segment that has stopped covering its costs; nothing in the system tells them the denominator moved because agents took over triage. HR knows exactly what happened and records it as a deployment milestone. Nobody else in the enterprise owns the joint between workforce data and financial data. Here is a case where the joint is load-bearing and unowned, and the output is a board recommendation to divest a healthy segment.

Walker provides an example (Walker, 2026a). A plant spreads a $10 million overhead pool across 200,000 direct-labor hours and charges $50 an hour. Agents absorb some of that labor, but the pool does not shrink, so the same $10 million now spreads across 120,000 hours, and the rate climbs to $83. The work still done by people absorbs what the automated work no longer carries. A segment reads as unprofitable. Nothing about the product changed. As Walker puts it, "the distortion has become a recommendation to cut the wrong thing." The same break runs through transfer pricing (how a multinational prices transactions between its entities in different countries), loan covenants (the financial guardrails imposed by lenders, such as limits on debt or requirements for minimum cash flow), and the diligence metrics a buyer tests at exit, such as revenue growth, profitability, cash generation.

The same reading can now hide opposite realities. A rising revenue-per-employee figure can no longer tell you whether the team got stronger or whether the margin-makers left and software is covering for them for now. Two opposite outcomes, one identical number (Walker, 2026b). A metric that cannot distinguish a strengthening workforce from a stripped one is not merely noisy; it is broken. It is no longer suited for the decision it is being used to make.

The argument so far has been mechanical: this is what must happen when software takes on part of a role. It also yields a testable prediction. Recall the metric’s two jobs. The input job is untouched: the salary does not move when the work moves. The output job is the one that breaks. So across an industry adopting agents, the two should come apart on the page: measures of output should respond while labor cost per head stays comparatively still. Cost should lag output rather than track it. If instead cost per head fell in step with automation intensity, the severing thesis would be wrong, and the weld would be holding.

Walker (2026a) cites a paper that offers a first look at whether this is visible in the data. Yu and Li (2026) read automation intensity out of the language of regulatory filings and track it against operating outcomes at five large U.S. financial institutions: Bank of America, BNY Mellon, JPMorgan Chase, S&P Global, and State Street. They find no statistical relationship between automation intensity and labor expense per employee (coefficient 0.0107 and robust standard error 0.0219). Automation intensity rises; what a firm pays per head does not follow. Their sample is small, and they present the result as descriptive rather than causal, calling it a structured pilot, so this establishes nothing on its own. What it does is fail to contradict the prediction, in the one place someone has looked.

Why finance will care before HR does

If this sounds like an HR debate, look at where the pressure is actually arriving: the transaction market.

Walker’s companion brief on private equity traces the collapse of the model that carried buyout returns for twenty years. More than half of past returns came from multiple expansion: buying at one earnings multiple (the price paid per dollar of annual profit) and selling at a higher one. It is the single largest driver of buyout value creation, about 56% of the average deal from 2016 to 2021, versus 48% for the prior five-year period (Bain & Company, 2022, using CEPRES data). For comparison, the other primary sources of value creation are revenue growth and margin expansion, contributing 38% and 6%, respectively, in the more recent period.

That tailwind is gone, and the operating growth a deal must deliver has roughly doubled. Bain’s shorthand, “twelve is the new five,” refers to the annual EBITDA (earnings before interest, taxes, depreciation, and amortization, the profit measure buyout deals are judged on) growth rate a deal now needs. Buyout holding periods at exit are hovering around seven years, up from an average of five to six years from 2010 to 2021 (Bain & Company, 2026). What must carry returns now is durable operating growth. In human-capital-intensive businesses, the variable that settles the exit is what Walker calls “talent density,” the margin-producing capability held by the people who remain. Buyers, burned by a year of AI disappointment, discount what cannot be proven (EY, 2026). “Deals now close on proof” (Walker, 2026b). The pressure is not only from buyers.

The macroeconomic arithmetic points in the same direction. Acemoglu (2025) estimates that applying AI to the existing task mix alone would yield aggregate productivity gains of well under one percentage point over a decade. The estimate is not a forecast that AI will disappoint; it is a statement about what the gain is conditional on. Larger gains depend on redesigning work and creating new tasks. The field evidence reinforces that argument. As Garrett Walker put it during our Conference Board expert panel, “AI amplifies whatever system it inherits,” an idea he subsequently formalized as the “Systems Law” in Walker (2026c). 

MIT NANDA (2025) reports that 95% of organizations in its study saw no measurable P&L impact from their generative AI initiatives. The figure has been much debated. McKinsey & Company (2025) offers a more careful picture: 39% of respondents report some enterprise-level EBIT (operating profit) impact from AI, but only about 6% qualify as "AI high performers," organizations reporting 5% or more of EBIT attributable to AI and significant value from its use. The distinction matters: measurable impact is not the same as material enterprise value. Our new Conference Board research takes the same view—whatever the exact figure, the pattern holds—and adds the diagnosis: the failures are overwhelmingly organizational, not technical (Schweyer et al., 2026; McKinsey & Company, 2025).

The obvious objection is that this is premature. If most AI initiatives produce no measurable P&L impact, how can the metric behind them already have broken? The two claims measure different things. The value literature asks whether enterprises have captured a gain. The severing asks whether one person still contains one stable bundle of work. A task moves to software the moment it moves, whether or not anyone books the benefit. Cost per head breaks at absorption, not at capture, which makes the record of failure the aggravating condition rather than the counterweight.

Organizations that redesign the work and redeploy the capacity that comes free will eventually see their numbers re-cohere around the new arrangement. Organizations that automate without redesigning hold the cost, shed the work, capture nothing, and get a metric that reports no problem. That is the worse of the two cases, and it is where most companies currently sit. If the work is not redesigned and the measurement is not rebuilt, there is nothing for the value to show up in.

The successor: capability yield

In our Conference Board series on agentic AI and work redesign, we propose the successor metric: capability yield. This is the value people and agents produce together, relative to their full cost (Schweyer et al., 2026). It is the return-on-workforce ratio, rebuilt for a workforce whose unit of work is no longer the person. Nesbitt and Sun (2026) asked what capacity an organization actually needs, how it should be composed, and how to deploy it toward outcomes that matter. Capability yield is an attempt to answer them in a number a CFO will accept.

Three design choices distinguish capability yield from the number it replaces.

The numerator is joint output, attributed, not assumed. It measures what humans and agents produce together, because that is where the value now lives, and it counts only the margin produced over what would have happened anyway, measured against a stated counterfactual (a baseline, written down in advance, of what would have happened without the change), with overlapping levers counted once, and the claim stated as a conservative floor. This is attribution, not correlation. Walker’s useful illustration: a gross sum of workforce levers that reads as 195–310 basis points of margin (roughly 2 to 3 percentage points) resolves, after subtracting the baseline and removing double counting, to roughly 80–120. “The governed figure is what survives an audit” (Walker, 2026a).

The denominator is the full cost of the blended workforce. People, fully loaded, and agents, fully loaded. Agents are not software you buy once; they are a workforce you keep paying for: consumption-based usage pricing (you pay per unit of use, like a utility bill), plus governance, monitoring, orchestration, and updating. In our research panels, one member could not predict whether a single HR agent would cost $200,000 or $1 million a year under usage pricing. Mavvrik (2025) finds only 15% of 372 enterprises in its study forecast their AI costs within 10% of the actual figure. Budget agents the way you budget staff, and charge them to the metric the way you charge staff (Schweyer et al., 2026).

The unit is the capability, not the head. Jesuthasan and Boudreau (2022) argued that the job had stopped being the right container for work. The same logic now runs through measurement: the head has stopped being the right container for cost. Following the workforce-scorecard tradition of pivotal positions (Huselid et al., 2005), capability yield is computed where deconstruction creates pivotal value (the “A-positions” where performance differences swing business outcomes, and the zones where work has been redesigned around AI), not smeared across the enterprise as an average. And attribution carries a governance bonus Walker (2026a) notes in passing: “the line of attribution is the line of accountability: a human owns the number even when the AI agent did the work.”

Two measures sit beneath the ratio, and they are what a people analytics team would actually build. The first is the redeployment rate: of the hours agents free up, what share gets redirected into more valuable work rather than quietly absorbed? Capacity that is freed and goes nowhere produces no numerator. The second is the attribution rate: of the value a business claims. What share survives the checks a careful CFO would apply? In our modeled cases, only one-third to one-half of claimed gains survive them (Schweyer et al., 2026). Walker's illustrative figure, constructed independently, lands in the same range. Two methods, one conclusion: most of what gets claimed will not hold up.

What should people analytics teams do now?

This is people analytics work, for the reason the absorption case made concrete: nobody else owns that load-bearing joint between finance and HR data. And my Second Law of Workforce Analytics, the consumption of analytics takes time and effort (Mohindra, 2015), applies with full force. A successor metric needs to be adopted, not just announced. There are four starting moves.

First, build the divergence signal. Watch how far revenue per employee drifts from output measured with agents included. The drift shows where the old metric fails, and it is detectable in your own books before a buyer detects it for you (Walker, 2026a). Second, pilot capability yield alongside cost per head in one function—one redesigned workflow, a stated counterfactual, agents charged at full cost—and let the two numbers argue in front of your CFO. Third, keep cost per head for what it is still good for: payroll, cash planning, and compensation benchmarking. The discipline is knowing which decisions it can still support. Fourth, apply the who-benefits test before deployment, not after: does the AI dividend go to investors as lowered expense, to customers as better service and speed, or to employees as reduced workload? The technology cannot make that allocation; leaders must, and a successful transformation delivers to all three. Say the choice out loud. Unlike most employer promises, this one is visible and verifiable (Schweyer et al., 2026).

For decades, we priced the role because the role and the work were the same thing. They are not anymore. The organizations that learn to price the work—capability, attributed, at full cost—will carry defensible numbers into every budget debate, board meeting, and data room. Everyone else will find out what their workforce was worth when someone across the table prices it for them.

AI disclosure. I made extensive use of generative AI (Anthropic’s Claude and OpenAI’s ChatGPT) to check quotations and citations against their sources and to draft and refine prose throughout the article. I reviewed all sources and take full responsibility for the final work.

References

Acemoglu, D. (2025). The simple macroeconomics of AI. Economic Policy, 40(121), 13–58. https://doi.org/10.1093/epolic/eiae042

Bain & Company. (2022). Global private equity report 2022. https://www.bain.com/globalassets/noindex/2022/bain_report_global-private-equity-report-2022.pdf.

Bain & Company. (2026). Global private equity report 2026. https://www.bain.com/insights/topics/global-private-equity-report/

EY. (2026). EY global private equity exit readiness study 2026. https://www.ey.com/en_gl/insights/private-equity/private-equity-exit-readiness-study

Huselid, M. A., Becker, B. E., & Beatty, R. W. (2005). The workforce scorecard: Managing human capital to execute strategy. Harvard Business School Press.

Jesuthasan, R., & Boudreau, J. W. (2022). Work without jobs: How to reboot your organization's work operating system. MIT Press.

Mavvrik. (2025, September 10). 2025 state of AI cost management research finds 85% of companies miss AI forecasts by >10% [Press release]. PR Newswire. https://www.prnewswire.com/news-releases/2025-state-of-ai-cost-management-research-finds-85-of-companies-miss-ai-forecasts-by-10-302551947.html

McKinsey & Company. (2025, November). The state of AI in 2025: Agents, innovation, and transformation. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

MIT NANDA. (2025, July). The GenAI divide: State of AI in business 2025. https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf

Mohindra, A. B. (2015, April 19). Three laws of workforce analytics [Post]. LinkedIn. https://www.linkedin.com/pulse/three-laws-workforce-analytics-amit-mohindra/

Nesbitt, C. K., & Sun, Y. (2026, April 28). People analytics is in a midlife crisis (Part 2): A field reborn around capacity and leverage. Variance, Explained. https://varianceexplained.substack.com/p/people-analytics-is-in-a-midlife-b8a

Schweyer, A., Fumento, M., Weathers, A., & Mohindra, A. B. (2026, August 10). A framework for agentic AI and work redesign: Executive summary. The Conference Board.

Subscript. (2025, June 27). Why revenue per employee is misleading in 2025. The Dive. https://www.subscript.com/the-dive/why-revenue-per-employee-is-misleading-in-2025

Walker, G. (2026a). The unissued restatement: How agentic AI severed the "labor proxy" embedded in the P&L. OSF. https://doi.org/10.17605/OSF.IO/VZ7QS

Walker, G. (2026b, July 27). The PE uptick, and the shift beneath [Strategic brief]. Workforce Value Creation & Advisory. https://systemslaw.substack.com/p/the-pe-uptick-and-the-shift-beneath

Walker, G. (2026c). The workforce operating system: A unified architecture for agentic workforce management and enterprise value. SSRN. https://doi.org/10.2139/ssrn.6900960

Yu, L., & Li, X. (2026). From clerks to agentic AI: How will technology transform the labor market in finance? arXiv. https://arxiv.org/abs/2604.19833