AI Agents May Be More Affordable Than We Can Handle

Zhang mentions that his company earns profits from each conversation after accounting for certain overhead expenses, yet he refrained from discussing the overall profitability of the startup. With $100 million raised from investors like Andreessen Horowitz and Accel, Decagon can focus on growth rather than immediate profits. “Considering whether we could be charging more is always a ‘what if?’” he states. “But we’re generally quite satisfied at the moment.”

“So Affordable”

Erica Brescia, managing director at Redpoint Ventures, realized the potential of AI agent pricing last month. She was taken aback by the $250 cost of Google’s new AI Ultra plan. “Everything feels so inexpensive,” she remembers thinking. “It’s out of sync with the value being delivered to users.” She believes a price at least twice that amount would be more appropriate. (In the same week, Nvidia CEO Jensen Huang expressed to Stratechery that he would employ an AI agent for $100,000 annually “without hesitation.”)

Previously, Brescia served as GitHub’s chief operating officer, which set the standard for AI pricing. GitHub’s Copilot coding assistant launched at $10 per month in 2022, ahead of ChatGPT’s release. Brescia notes that GitHub aimed for a price that would attract a significant user base, gathering data to enhance the service, even if it meant incurring losses for Microsoft, its parent company. In fact, she estimates that a price 100 times higher would now more accurately reflect the value Copilot offers developers. (Kyle Daigle, GitHub’s COO, states that the company’s intent is to assist rather than replace developers, with “pricing reflecting a commitment to democratizing access to powerful tools.”)

Currently, Copilot is priced at a maximum of $21 per month. Other tools in the market have adopted similar pricing, including Zed, which has secured $12.5 million in funding from Redpoint and others. This May, Zed began charging a minimum of $20 monthly for their AI-assisted code editor, developed from scratch.

Zed CEO Nathan Sobo anticipates that AI companies will eventually increase their prices since existing pricing structures may not be viable. However, he emphasizes the importance of keeping AI agents affordable compared to human labor, enabling everyone to enhance their work, create better software, and foster new jobs. “I want as much intelligence at my disposal at the lowest possible cost,” he states. “This includes the possibility of a junior engineer using this technology, ideally at an affordable rate.”

Zhang from Decagon shares a similar perspective on AI coding tools. “Would we consider paying more? Marginally? Yes,” he responds. “But $2,000? Probably not.” He adds that “the demand for skilled engineers is limitless.”

AI entrepreneurs believe that higher prices could be justified if agents were simpler to set up and more reliable. For example, Nandita Giri, a senior software engineer with experience at Amazon, Meta, and Microsoft, says she would be willing to invest thousands annually for a personal AI assistant. “However, it comes with strict stipulations—you can’t become frustrated while using it,” she remarks.

Unfortunately, that optimal scenario seems distant. Giri attempted to create an AI agent to mitigate psychological burnout as a personal project. “It just canceled all my meetings,” she recalls. While a solution, it wasn’t the ideal one.

Currently, some companies are hiring “AI architects” to supervise agent systems and minimize errors. The concern is who will fill these roles in the future if entry-level workers are denied opportunities today. Simon Johnson, an economist at MIT, doubts that companies will factor in the societal costs of career disruptions in their pricing strategies. He advocates for governments to reduce payroll taxes on entry-level jobs to spur hiring. “The appropriate measure to take is one that lowers costs for employers,” Johnson advises.

Arrigoni is opting for a different approach. At Loti AI, he is focusing on gradually hiring junior engineers without relying on AI coding tools. If a job crisis arises, “I don’t want to be responsible,” he asserts.