AI as a New Kind of Labour

Vendors reframed agentic AI from software into a workforce

Across the 2024–25 enterprise cycle, the leading vendors changed what they were selling. Salesforce announced Agentforce 2.0 as “the digital labor platform for enterprises, enabling a limitless workforce through AI agents for any department,” and Benioff framed the company as “the leader in digital labor solutions,” letting “any company … build a limitless workforce.”¹ A year later, the Agentforce 360 launch generalised the move into a stated vision of the “Agentic Enterprise … where AI doesn’t replace people, it elevates them,” connecting “humans, agents, and data on one trusted platform.”⁷ The vocabulary is hiring vocabulary. A workforce gets built and managed, where software gets deployed and used, and every noun in these releases comes from the first list.

The framing is not confined to one vendor. At Davos in January 2025, Benioff put it in generational terms — “Today’s cohort of CEOs will be the last to lead all-human workforces” — and described leaders as soon “managing not only human workers but also digital workers,” while insisting the technology is “supporting employees, not replacing them.”² Microsoft’s 2025 Work Trend Index introduced the Frontier Firm, an organisation “built around … intelligence on tap and human-agent teams,” and the agent boss, someone who “builds, delegates to and manages agents to amplify their impact.”¹⁰ Microsoft frames digital labour for these firms as being as critical to the business model as human labour.¹⁰ NVIDIA’s Jensen Huang, in his CES 2025 keynote, placed the same role at the infrastructure layer: “In a lot of ways, the IT department of every company is going to be the HR department of AI agents in the future,” with agents that must be recruited, trained on company-specific vocabulary and policy, and supervised.³⁴

The investment case sizes AI against the wage bill, and the sizings disagree threefold

Under the marketing sits a venture thesis that explains why the labour framing is more than branding. Foundation Capital’s Joanne Chen and Jaya Gupta named the shift from software-as-a-service to service-as-software. Where SaaS sells a tool and leaves the customer responsible for the outcome, the new model has the vendor assume the outcome — “Instead of QuickBooks, you offer tax services … conducted by an AI accountant.”⁵ The sizing follows from the wage bill. Foundation Capital frames a $4.6 trillion opportunity (roughly $2.3T of global salaries plus $2.3T of outsourced IT and business-process services) and is explicit that “the real prize isn’t the $200B SaaS market: it’s the $4.6T enterprises spend on salaries and services.”⁵⁶ The unit the essays keep returning to is the gap between a software vendor’s revenue and the wage bill it could displace. Salesforce earns ~$35B a year against the ~$1.1T companies spend annually on sales and marketing salaries.⁵

The mechanism underneath the sizing is a reclassification of what software is. Alex Rampell’s a16z essay compresses it to “software becomes labor.” Traditional enterprise software digitised offline processes (filing cabinets into Workday, tickets into Zendesk, ledgers into QuickBooks), but those systems “required human ‘users.’” The break with that past is that “the ‘users’ of the digitized filing cabinet do not have to be humans.”¹² Once the user can be a machine, the relevant market is the labour budget, which for white-collar work Rampell puts at “many, many trillions of dollars a year.”¹²

The pricing consequence lands on the incumbents first. Workday, Intuit, Zendesk and Salesforce are per-seat businesses, and “a business will need fewer (if any!) seats as the system … takes actions on its own.”¹² Foundation Capital’s outcome-based-pricing argument is the same coin’s other face, since charging for results “creates a much more scalable pricing model” and gives the vendor an incentive to expand usage.⁵ The repricing is already on price lists. Salesforce’s Sales Development and Sales Coaching skills cost “$2 per conversation,” a per-unit-of-work price.¹ The labour-economics pitch completes the picture with scaling that skilled labour cannot match — “AI will always show up to work, can be trained instantly,” where nursing takes years and mortgage brokers could not be conjured when rates dropped in 2021.¹²

The attribution question hides a substantive one. Two sizings of nominally the same opportunity differ by a factor of three — HFS’s ~$1.5 trillion by 2035 against Foundation Capital’s $4.6 trillion²³⁶ — and the gap is itself information. Numbers this soft are theses, and the spread between them is the honest error bar on the category.

The direct counter to the sizing comes from Daron Acemoglu, whose task-based model makes aggregate gains depend on two terms multiplied together: the fraction of tasks AI actually affects, and the average cost saving on each.¹⁴ Counting the whole $4.6T wage-and-services bill as addressable assumes the first term is close to one. Acemoglu’s estimate of what the two terms deliver is “no more than a 0.71% increase in total factor productivity over 10 years,” revised to “less than 0.55%” once harder-to-learn tasks are counted, and he predicts the gains accrue to capital while labour’s share erodes.¹⁴

The fraction-of-tasks term now has an early measurement, from Anthropic’s pairing of Claude usage data with occupational task lists. Claude covered 33% of Computer & Math occupational tasks, the category where its use runs heaviest,¹⁷ despite the tasks’ theoretical suitability; across all observed Claude tasks, 97% fall into categories rated feasible. In the report’s words, “AI is far from reaching its theoretical capability: actual coverage remains a fraction of what’s feasible.”¹⁸ Around 30% of workers showed no coverage at all, and penetration concentrates in specific task clusters within occupations, leaving the rest of each job standing.¹⁸ The same measure locates the addressable slice among the well paid, since workers in the top exposure quartile earn 47% more on average than unexposed workers.¹⁸ The wage bill is the theoretical ceiling. The measured floor, so far, is a minority of tasks in the most exposed occupations.

Foundation Capital’s own year-one retrospective is useful precisely because it is less triumphant. The firms separating from the hype share three traits: forward-deployed engineers became “one of the most strategic assets,” because “integration is not a post-sale activity. It is the product surface”; the customer “expects to experience functionality, integration, and outcome before a contract is signed”; and pricing is moving from seats toward outcomes.⁶ The thesis, in other words, only pays off where the vendor can actually deliver an outcome.

Vendors promise capacity without headcount; deployment remains a minority

The operational consequence vendors emphasise is a break in the link between how much a firm can produce and how many people it employs. Microsoft’s Work Trend Index makes the demand-side case with a picture of a workforce out of capacity. Its telemetry has employees interrupted every two minutes, and the survey layer confirms the pattern: most of the global workforce reports too little time or energy for the job, and a majority of leaders say productivity must rise.⁹¹⁰ The prescribed remedy is digital labour, and 45% of leaders already name “expanding team capacity with digital labor” a top priority.⁹

The supply-side promise is vendor outcome data. Salesforce reports its own support agents resolving 83% of customer queries without human intervention, and customer deployments resolving large shares of work autonomously.¹⁷ Figure 2 collects the claims.

Every number in the two paragraphs above comes from a company selling the remedy. The chart carries the warning in its title because the outcomes are vendor-reported and unaudited, and the same discount applies to the Work Trend Index. It is Microsoft’s own research programme, fielded with Edelman Data x Intelligence and published by the vendor of Copilot and of the Agent 365 control plane for managing agents;⁹¹¹ its statistics describe demand for the thing its publisher sells.

With that discount stated, Microsoft’s survey carries two headcount signals, and they point in both directions at once. AI-native startups grew headcount 20.6% year-over-year against 10.6% for Big Tech, while 33% of leaders say they are considering headcount reductions and 78% are considering hiring for new AI-specific roles (95% at Frontier Firms).⁹¹⁰ If the survey is believed, capacity is uncoupling from bodies at both ends — fewer of some roles, net-new categories of others. The empirical question is whether output-without-headcount shows up in employment records kept by someone with nothing to sell.

Set against the promise, deployment is thin. BCG finds 13% of organisations have deployed agents integrated into workflows, and only a third of workers claim a clear understanding of what an agent even is.²⁰ McKinsey, surveying organisations rather than employees, finds the same shape, with most still experimenting with agents and only a minority scaling them.¹⁹ Figure 3 carries the shares from both surveys. The labour framing is running well ahead of the deployed technology.

In theory, span of control and layer count become design parameters

If output capacity uncouples from headcount, two pillars of organisation design stop being fixed by human cognitive limits: the span of control (how many reports one manager can handle) and the number of layers between the front line and the top. Two working papers model what follows, and both complicate the flat-org story.

Farach models AI as “agent capital,” a distinct production input that “reduces the friction of managing workers, expanding spans of control, compressing hierarchies.”¹⁶ The model’s central finding is that the same technology forks. Under positive “task-creation elasticity,” AI “expand[s] the frontier of feasible work” rather than only cutting jobs, and an “elite-complementarity” parameter determines whether AI acts as general infrastructure, with broad-based gains, or “amplifies top managers selectively,” concentrating them.¹⁶ Structure becomes a chosen regime, and the parameter “becomes a policy lever.”¹⁶ The org chart becomes a decision with distributional consequences.

Xu, Hou, Chen and Xie puncture two simplifications. First, the much-discussed decline in junior employment “reflects deployment choices favoring automation over augmentation, not an inevitable consequence of GenAI itself.”¹⁵ Stanford’s declines sat in the same place, concentrated where AI automates.¹³ Second, span of control moves non-monotonically. “Across all four deployment architectures, the span of control initially contracts before eventually expanding” as the technology improves, because early, fallible AI needs more skilled oversight before it needs less.¹⁵ The design space they model is a 2×2 — automation against augmentation, worker- against expert-level deployment — which is the set of choices a firm now has to make deliberately.

Both models arrive with the same caveat. Farach’s paper is a single-author arXiv preprint, and Xu et al. is an unreviewed working paper; neither has been through peer review, and no deployment data yet tests either. The three claims of this paper therefore stand on different evidence. Deployment outcomes are vendor-reported, the org-design claim rests on theory, and only the substitution question has been run against administrative data — the ADP payroll records behind the Stanford study and Anthropic’s designed exposure measure. The org-design claim should be weighted as what it is, a coherent piece of theory waiting for data.

Running many agents recreates management, and a new role with it

The moment a firm runs more than one agent, it inherits a coordination problem that looks exactly like management. IBM’s reference definition of hierarchical orchestration describes a literal hierarchy: “a manager agent decomposes a complex goal into sub-tasks and delegates each to a specialist agent. The manager monitors progress, handles failures, and reassembles results into a final output.”²² The other patterns, sequential and concurrent, are the same primitives an operations manager already knows — work passed down a line, or split and merged.²² Deloitte describes the enterprise versions as supervisor-agent models, adaptive networks and hybrids, and predicts firms will move along an “autonomy spectrum” from human in-the-loop to on-the-loop to out-of-the-loop, with the shift toward human-on-the-loop orchestration beginning in 2026.²¹

The human role this creates recurs across the vendor literature under different names. Microsoft’s agent boss “builds, delegates to and manages agents,” applies “at any organizational level,” and already comes with new job categories attached (AI Trainer, Agent Specialist and ROI Analyst among them).⁹¹⁰ Huang’s IT-as-HR line is the same role at the infrastructure layer, recruiting, onboarding and supervising a digital workforce.³⁴ Microsoft has already turned the function into a product. Agent 365 is a “control plane for AI agents,” with a registry providing a “comprehensive inventory of all agents,” unique agent IDs operating under “principle of least privilege,” role-based dashboards, and “unified observability across your entire agent fleet.”¹¹ The vocabulary — fleet, registry, least-privilege, observability — is borrowed equally from HR and from IT, which is precisely Huang’s point.

Scale is what turns this from a curiosity into an operating problem. Deloitte sizes the autonomous-agent market at $8.5B in 2026, rising to $35B by 2030 or $45B “with improved orchestration,” and Microsoft cites an IDC projection of 1.3 billion agents by 2028.²¹¹¹ A firm managing thousands of agents needs what a firm managing thousands of people needs (identity, access control, performance monitoring, and a manager), which is why “agent sprawl” and the registry that contains it are now product categories.¹¹

The discount on all of this is that the projects the role would manage are mostly experiments. Gartner predicts “more than 40% of agentic AI projects will be canceled by the end of 2027,” on a poll of 3,400+ organisations, because “most agentic AI projects right now are early-stage experiments or proof of concepts that are mostly driven by hype.”⁸ The firm coined “agent washing” for chatbots and RPA rebranded as agents, and estimates that of the thousands of vendors claiming agentic solutions “only around 130 offer real agentic features.”⁸ Deloitte independently echoes the cancellation figure.²¹ An agent-manager role built on today’s project pipeline inherits today’s cancellation risk.

The reframing is real, the org-design claim is early theory, and substitution at scale is unshown

The three claims come apart cleanly. The reframing is real and near-universal across vendors and investors. Agentic AI is being built, priced and governed as labour, with a coherent economic logic behind it: software as capital that produces work, sized against the wage bill.¹⁵¹² The organisational implication follows from the coordination problem itself, which is why a management role for agents has been independently named across the vendor and analyst literature and modelled by the academics; once managing workers gets cheaper, span of control and layer count become variables a firm sets.^15161022 That claim is the one this paper judges most likely to outlast the hype cycle, and it is also the one still waiting for deployment evidence.

The substitution claim is the one the evidence declines to support. Production deployment is a minority,²⁰¹⁹ a plurality of projects is expected to fail,⁸ measured task coverage runs at a third of what is theoretically feasible in the most exposed occupational category,¹⁸ and the employment effect so far is concentrated in entry-level hiring; broad displacement has not appeared in the data.¹³¹⁸ The single cleanest behavioural signal, the split of Claude.ai usage between augmentation and automation, points the same way. Augmentation has just overtaken automation, 52% to 45%, reversing the August-2025 position; Figure 5 carries the full series.¹⁷ For now, the modal interaction in the measured usage data is a person working with an agent.

References