AI Agents Are Expensive to Run — And That Changes the Replacement Narrative

AI replacing humans as labor has been a hot topic, and a bold one, with two specific people on the hook for it claiming it may also be becoming too hyped. In the past few months, two significant technology investors who actually have dollars behind their decisions are now questioning what happens if there is more expense associated with an AI than there ever would be with the person who was originally to be replaced?

This is not merely a question; it is happening right now.

The $300-a-Day Reality Check

At a recent taping of the All-In podcast, Jason Calacanis — a venture investor and co-host of the podcast — stated that he pays $300 a day to run an Anthropic Claude AI agent to run his businesses. This amounts to approximately $9,000 per month, but here's the kicker: The agent's performance is only between 10% and 20% of maximum capability, according to Calacanis. 

So when you do the math on cost-per-useful-task at $300 per day for partial output, it begins to look less like a disruptive technology and more like an expensive experiment.

Calacanis also described the issue simply: "When do tokens become more expensive than paying an employee?" The tokens are the units of usage that all AI models charge for, and as an AI agent continues to perform tasks during the day, the tokens add up quickly. The more tasks the agent performs, the more expensive your bill will be. Unlike an employee on a salary, there is no limit on how many tokens you are charged for as long as the agent is actively using tokens without you setting limits on token usage.

Chamath's Two-Times Rule

Chamath Palihapitiya, the CEO of Social Capital and co-host of the All-In podcast, shared that he faced a similar issue as Social Capital, which he paraphrased, "AI Agents can only justify their costs to me if they provide me with a minimum of 2x a human employee’s output at the same job."

Using those two times as a basis gives us a way to analyze AI's effectiveness versus a human employee because there are several factors that contribute to their total operating cost. Some of these factors include:

- Cost of running (roughly $70,000 - $150,000/year depending on cloud vs on-premise)

- Cost of managing and prompting (roughly $20,000 - $50,000/year depending on how often they are used)

- Time spent fixing errors due to mistakes or poor quality data

- Productivity loss during time periods when AI is under performing

When you add up the total operating costs for a human company vs AIs, when a company can hire a human to do a job for $70k or more than $150k, a company regardless of AI vs human is going to find the same cost to hire the person. It becomes obvious that the human company will preferably want to hire the person as they can easily hire someone who has worked with them for years, trained and learned about their business to do a job similar to a new AI.

He advised that some companies will probably implement some form of hard budget cap to limit the amount of AI used (this is surprising to hear from someone who has spent years betting on the transformative ability of technology), indicating that even people who are dedicated to AI are eventually going to be treated just like any other line item in a company's budget and thus will need to provide justification for the expenditure.

By Author

The Token Problem Nobody Warned You About

Calacanis points out that there is a more critical problem in AI articles mostly focused on performance metrics and demo video information than token costs. Most of the AI being used by enterprise companies involves a lot of inputs, outputs and transactions that involve an agent interacting with a system. In some cases these will add up at a reasonable level for small jobs but when you start adding an agent doing its job for 8 hours per day across all the departments in a business, the number of tokens being used will reach significant amounts in no time at all.

The business model for an AI agent operating on a token-based model is drastically different from hiring a human resource to perform a job. An employee working for a company on an annual salary will have a set cost, no matter how productive he or she is on any given day. An AI working under a token-based model will have a cost that increases the more productive he or she is — or, in other words, the more value created for a company.

The threshold for productivity required for an AI agent to return a positive ROI for the company that purchased it continually changes, as AI model providers will be forced to compete against each other with price drops over time. However, as of early-2026, there is enough difference, in terms of monetary cost, between what the agents cost and what the companies that own them can expect from them to generate conversation between two individuals that both paid significant amounts of money for them are now sharing publicly.

This Doesn't Mean AI Agents Are a Dead End

Let us be very clear, neither Calacanis nor Palihapitiya have dismissed AI Agents they are both very much committed to this area from the standpoint of investments (both financially and intellectually). They take a more focused and thus pragmatic approach in their argument against the hypothesis of AI agents being substitutes. Examples of this include how on a macroeconomic level that the economics must work for a substitution to happen; however, currently the macroeconomics of AI agents is quite chaotic.

In the future, as the models become more efficient and providers continue to compete for market share, the feat of AI agents will decrease and therefore cost less. Both the effort of OpenAI, Anthropic, Google and the various Open Source alternatives all push for more efficient models. Once the costs per token decrease and reliability increases the economics would start to favour switching to AI Agents. However, using "in the future" to support the current hypothesis of AI Agents being substitutes is very misleading and the users who are actively using these tools daily can provide real world data about how things actually are.

By Author

What to Watch Next

The principal aspect is price. If both OpenAI and Anthropic, along with their competitors, proceed to decrease the cost per token (and the trends show that this is occurring), then the breakeven point outlined will alter. Companies that develop processes concerning agents now (despite negative margins) could be preparing themselves for a different cost structure in the next 12 to 24 months.

However, any individuals using AI to define workforce arrangements at present because of the belief AI is lower cost than human workers will need to reconsider based on the historical experiences of the investors involved in the operator networks.