Eighty-seven percent of professional services organizations now plan to manage AI agents as part of their workforce, according to Kantata’s 2026 State of Professional Services report. Not their software stack. Their workforce. And 89% of those leaders say future revenue growth depends more on how effectively they scale AI than on how they scale headcount.
That single data point captures a shift that most companies still haven’t processed. AI agents are not the next SaaS product you buy. They are the next labor category you hire, manage, and budget for. The Brookings Institution put it plainly: this is the first technology capable of independent economic activity. When your “software” can reason through problems, take actions across multiple systems, and produce measurable output without human involvement, calling it software is just a comfortable lie.
The Reclassification That Breaks Your Budget Model
Every company has a clean line between two types of spending. Technology goes in one bucket: licenses, infrastructure, maintenance. People go in another: salaries, benefits, training, management overhead.
AI agents sit in neither.
They cost like technology (compute, API calls, infrastructure) but behave like workers (they make decisions, interact with customers, produce deliverables). Salesforce’s Agentforce platform processed over 3.2 trillion tokens in Q3 FY2026, with 83% of customer service queries resolving entirely without human intervention. That is not a chatbot. That is a department.
The problem is that most finance teams still allocate agent costs as a line item under “AI/ML tools” in the IT budget. When agents are handling 83% of your customer interactions, they should be in the workforce planning discussion. The Kantata report found that 90% of organizations say their systems need to attribute work, costs, and value across both humans and AI agents. That attribution is impossible when agents are buried in the technology budget and humans are tracked in HR.
Why This Is Not Just an Accounting Exercise
Budget categories shape organizational behavior. When agents are classified as IT tools, IT owns them. When agents are classified as workforce, someone has to decide who manages them, who evaluates their output, and who is accountable when they make a $200,000 mistake.
The Carnegie Endowment documented that US employment in the information sector has dropped roughly 10% since early 2020, even as overall employment grew. That is not a recession story. It is a substitution story. The work did not disappear. The category of worker doing it changed. Companies that do not reclassify agents as workforce will keep being surprised by these shifts because they are not measuring them.
What Changes When Labor Has a Marginal Cost of Near Zero
Traditional labor economics has a foundational assumption: adding one more worker costs roughly the same as the last one. Salary, benefits, onboarding, equipment. In the US, a fully loaded knowledge worker costs $80,000 to $200,000 per year.
Spinning up one more AI agent instance costs compute. Depending on complexity, that is somewhere between $0.01 per task (classification, routing) and $10 per task (multi-step research and analysis). There is no salary negotiation, no benefits package, no three-week onboarding program.
This breaks the supply curve in ways that traditional workforce planning cannot account for. When you can scale from 10 to 10,000 workers in an afternoon, headcount planning becomes capacity planning. The constraint is not “can we hire fast enough” but “can we supervise at this scale.”
The Supervision Bottleneck
Here is where the theory crashes into reality. The same Kantata survey found that 89% of leaders spend significant time verifying AI agent outputs, even though 88% trust those outputs enough for operational decisions. Only 12% fully trust their system data, down from 24% a year ago.
Near-zero marginal cost per agent means nothing if each agent needs a human checking its work. The real unit economics of AI agents are: compute cost per task + (human supervisor salary / number of agents one human can oversee).
Nvidia CEO Jensen Huang proposed a concrete model in March 2026: giving engineers AI token budgets of $100,000 to $150,000 and targeting a 100:1 agent-to-human ratio within a decade. At that ratio, the supervision cost per agent drops enough to make the economics work. At a 5:1 ratio, which is where most organizations operate today, agents are expensive augmentation, not cheap labor.
The Job Titles That Did Not Exist 18 Months Ago
When you reclassify agents as workforce, you need people to manage that workforce. This is creating an entirely new job category.
Eightfold AI identified “AI orchestration specialist” as one of the most important emerging roles of 2026. The job: coordinating multiple AI agents, defining their workflows, monitoring their output, and intervening when they fail. It sits at the intersection of project management, systems engineering, and domain expertise.
This is not prompt engineering. Prompt engineering was about getting a single model to produce the right output. Orchestration is about managing a team of agents that work together, hand off tasks, and need different levels of autonomy depending on context. It is closer to people management than programming.
Other roles appearing in job postings for the first time: agent reliability engineer (keeping agents running in production), AI workforce planner (deciding which tasks go to agents vs. humans), and agent compliance officer (ensuring agents meet regulatory requirements).
The Skills Premium
Workers who develop agent management skills are being paid for it. Research shows that workers with AI skills command wage premiums up to 56% higher than their peers. The 5% of workers who are fully AI-fluent earn 4.5x more than those who are not.
This is a classic supply-demand signal. Companies need people who can manage an AI workforce. There are not enough of those people. So the price goes up. The irony: the role that AI agents create the most demand for is the role of managing AI agents.
How to Actually Budget for a Workforce That Does Not Draw a Salary
If agents are workforce, you need workforce-level financial planning. Here is what that looks like in practice.
Step 1: Separate Agent Costs from IT Infrastructure
Create a distinct budget category for AI agent operations. This should include compute costs (API calls, model inference), tooling costs (orchestration platforms, monitoring), and supervision costs (the humans who manage agents). Do not lump it into “cloud infrastructure” or “AI/ML research.” Those categories hide the true cost of your digital workforce.
Step 2: Measure Output Like Headcount, Not Like Uptime
Traditional software metrics (uptime, latency, error rates) tell you whether your agents are running. They do not tell you whether your agents are productive. Salesforce introduced the Agentic Work Unit (AWU) to measure what agents actually accomplish, not just whether they are online. Think of it as the agent equivalent of a job performance review.
The metric you actually want: cost per completed task vs. cost per completed task by a human doing the same work. That is the number that tells you whether your agent workforce is worth scaling.
Step 3: Plan for a Blended Workforce Ratio
Nvidia’s 100:1 target is aspirational. Most organizations should start with realistic ratios: 5:1 for complex knowledge work, 20:1 for structured process work, 100:1 for classification and routing tasks. The ratio determines your supervision budget, which is the largest hidden cost in agent deployment.
Deloitte’s agentic AI framework puts it well: the human role shifts from task execution to orchestration and oversight. Budget accordingly. You will hire fewer task executors and more orchestrators. The total headcount may decrease, but the skill requirements (and salary expectations) for the remaining humans will increase.
Frequently Asked Questions
Why are AI agents considered workforce instead of software?
AI agents make decisions, interact with customers, produce deliverables, and take autonomous actions across multiple systems. Unlike traditional software that executes predefined logic, agents reason through problems and adapt their behavior. The Brookings Institution calls them the first technology capable of independent economic activity. When 87% of professional services organizations plan to manage agents as part of their workforce, the reclassification from software to labor is already happening in practice.
How should companies budget for AI agents?
Companies should create a separate budget category for AI agent operations that includes compute costs, tooling, and human supervision costs. Do not bury agent costs in IT infrastructure or AI research budgets. Measure agent output using productivity metrics like cost per completed task, not just uptime or error rates. The largest hidden cost is supervision: at a 5:1 agent-to-human ratio, human oversight is expensive. At 20:1 or higher, the economics start to work.
What new jobs are AI agents creating?
AI agents are creating demand for AI orchestration specialists, agent reliability engineers, AI workforce planners, and agent compliance officers. Eightfold AI identified AI orchestration specialist as one of the most important emerging roles of 2026. Workers with AI agent management skills command wage premiums up to 56% higher than peers without those skills.
What is the ideal agent-to-human ratio?
It depends on the work. For complex knowledge work, a 5:1 ratio is realistic today. For structured process work, 20:1 is achievable. For classification and routing tasks, 100:1 or higher works. Nvidia CEO Jensen Huang has proposed a long-term target of 100:1 across the entire organization, with AI token budgets of $100,000 to $150,000 per engineer to fund the compute.
How does treating AI agents as labor change workforce planning?
When agents are workforce, headcount planning becomes capacity planning. The constraint shifts from hiring speed to supervision capacity. Organizations need to track work attribution across both humans and agents (90% say their systems need this). Finance teams must model blended workforce costs that combine human salaries with agent compute costs. The total headcount may decrease, but remaining human roles require higher skills and command higher salaries.