When the Agent Takes the Desk: The Quiet Displacement at the Heart of AI Economics
The first thing a business discovers about an AI agent is that it is not a chatbot with a longer to-do list. It is a different economic animal. IDC projects that more than 1 billion AI agents will be at work worldwide by 2029, roughly 40 times the 2025 total, and that those agents will consume 3.7 trillion tokens and API calls a day, as reported by Forbes. [1] Each of those calls carries a price. That price is what determines whether a company can afford to let software handle millions of actions — or whether a person keeps the job instead.
The distinction matters because the agent does not simply replace a task. It replaces a person’s judgment about when the task is finished. A retail operation that wants an agent to process returns can watch a sales demonstration in which the process looks effortless: the agent locates the order, checks the policy, approves the refund, and sends the customer a shipping label. The investment decision becomes harder when the agent meets a missing receipt, a disputed purchase, or an exception to the policy. Every additional check consumes more computing. Every mistake and every unresolved case pulls an employee back into the process. Before expanding the deployment, the retailer has to weigh those computing costs against the expense of human help — and that comparison is where the human role either survives or quietly disappears.
Two Stages, One Deciding Line
Training and inference are the two stages in the life of the models behind AI agents. During training, a model learns patterns from data and develops capabilities such as understanding language. Inference happens when someone uses that trained model to do work, such as interpreting a customer’s return request or determining whether it meets the retailer’s policy. For a business leader, the financial distinction is the whole ballgame. Training is an upfront investment that gets repeated when models are updated. Inference happens every time the trained model is put to work. Training is the schooling. Inference is the job.
A company buying an AI service may never train its own model, but its usage still generates inference costs that the provider must recover through pricing. That ties a leader’s AI budget to a race among the chipmakers that build AI infrastructure. More efficient chips could make more workflows worth AI enabling, while wasteful ones could eat the savings a business expected. A falling price for AI processing does not guarantee a cheaper AI bill. As agents take on more steps, investigate exceptions, and check their own work, their total usage can grow faster than prices decline. The arithmetic of substitution — machine for human — runs through every line of that budget.
What the Tokens Reveal About the Replaced Worker
Marc Bolitho, chief executive of Tensordyne, estimates that agents consume five to 10 times as many tokens as chatbots, sometimes more, as reported by Forbes. [1] Providers often use tokens — which measure the information a model processes and generates — to calculate usage charges. For leaders budgeting an agent deployment, the implication is that last year’s chatbot bill offers a poor guide to what a more ambitious agent will cost. The chatbot answered a question; the agent finishes a job. The difference in consumption is the price of the judgment the agent is now expected to exercise.
Response time belongs in that budget as well. A customer waiting for a refund authorization needs a prompt answer, while an overnight report can tolerate a longer wait. Leaders should ask providers to price the workload at the speed the business requires, then test whether the economics hold as volume and complexity increase. Bolitho argues that Nvidia’s competitive moat is stronger in training AI models than in running them. [1] His view is that inference creates more room for specialized competitors to win on cost and speed. Nvidia and AMD each pair their chips with software designed to run models efficiently, so competition is already underway. For business leaders, the distinction between training and inference opens another purchasing decision: the hardware used to train a model does not automatically determine the best platform for running it.
The Denominator That Hides the Person
Consider the retailer again. Dividing its model bill by the number of returns the agent attempts would leave out much of the cost of operating the service. The calculation also needs to capture integration, maintenance, monitoring, failed attempts, and the employees who correct mistakes or finish unresolved cases. A more useful measure is the total operating cost divided by the number of returns resolved to the retailer’s required standard. The business should also track how often customers contact support again, because an inexpensive interaction can become an expensive resolution when the customer has to repeat it.
That measure makes infrastructure claims easier to evaluate. More tokens per watt may improve a provider’s economics, but a retailer buying AI through that provider needs to see the benefit in its own pricing or service. A company purchasing equipment directly must also account for how much of its capacity it will use. The same discipline applies to time saved. If an agent frees employees from routine returns, leaders should identify what those employees can now accomplish and measure the change. Saved time supports a business case when it produces additional capacity, better service, or an expense the company can avoid — and when it does not, the saved time is simply the space left behind by a role that no longer exists.
Five Questions for the Person Who Signs Off
Doug Levin, an executive fellow at Harvard Business School, recommends clearer accounting for AI investments. His guidance translates into five questions leaders can use to connect deployment decisions with financial oversight. Where does AI spending appear in the budget? What does each successfully completed task cost? Who is accountable for AI spending? Which agent deployments require approval? How often does the audit committee receive reports on AI costs?
The answers should reveal how spending changes as an agent takes on more work. A deployment that is economical for routine returns may become costly when it expands into disputes that require repeated checks and employee intervention. Leaders need enough visibility to recognize that change before a successful pilot becomes a disappointing investment. Each of those questions, pointed at the right place, asks the same thing in a different register: who is still responsible when the machine is doing the work?
The Evidence the Retailer Should Demand
At the time of the interview, Bolitho said Napier’s projected system performance will be checked against simulations. Tensordyne’s published comparisons with Nvidia Blackwell systems claim up to 17 times as many tokens per watt and 13 times the throughput — company projections of energy efficiency and processing capacity, not independently established results or equivalent reductions in a customer’s bill. [1]. Independent testing on finished systems will need to establish whether those advantages hold under realistic workloads while preserving model accuracy, and whether lower processing costs let an agent perform additional checks or handle more complicated cases within an acceptable budget. A retailer evaluating that technology should bring its difficult returns to the test, including disputed purchases and policy exceptions. Before expanding, leaders should require evidence that improvements in inference translate into a lower total cost per correctly resolved return, delivered within the service standard their customers expect from AI agents.
None of that testing answers the deeper question the economics keep circling. The agent is cheaper per action, faster per step, and available at every hour. What it does not do is decide, on its own authority, that a disputed purchase deserves a refund, that a policy exception is fair, or that a customer’s frustration outweighs the rule. Every one of those decisions still lands on a person — the employee pulled back into the process, the manager who signs off on the expansion, the audit committee that reads the cost report. The system can be measured down to the token, but the responsibility for what it gets wrong cannot be divided that finely. Lower processing costs may let an agent take on more checks and harder cases, yet whether that produces a better result still depends on the agent’s accuracy, the workflow, and the price the customer ultimately pays. That is the line the cost models leave out, and it is the line that decides what the human role is still for.
