a16z's Anish Acharya: personal AI agents cost $1,000 per user per year today — but deflation is coming fast

Personal AI agents: the cost problem and what comes after

Anish Acharya, a partner at Andreessen Horowitz, puts the current cost of running a personal AI agent at $1,000 per user per year — a figure he acknowledges is prohibitive at scale. The business model math is punishing: the average American spends roughly $5,000 to $6,000 a year on online travel and commerce combined, so even a generous take rate on transactions leaves the economics underwater unless compute costs fall sharply.

Acharya thinks they will. He points to recent progress on computer-use models as evidence, singling out a system he describes as "100x cheaper and faster" than earlier computer-use approaches as the kind of deflationary shock that could change the calculus quickly. The pattern has repeated itself across LLMs, reasoning systems, and agentic coding, each of which initially looked expensive before costs collapsed.

“Our best estimates are a thousand dollars per user per year, which is obviously prohibitive if you're running this. The number one thing that needs to happen is cost needs to deflate, and they will. The most exciting thing I saw in the last two weeks was Jeb — computer use is 100x cheaper and faster today.”

Friction with merchants

The near-term friction point is agents running into blocks on retail sites, including a reported case of a personal agent getting stopped while trying to make a purchase on Adidas.com. Acharya expects most transaction-focused retailers to adapt fast, because intermediary friction costs them revenue. The harder case is platforms whose business model depends on attention and advertising rather than fulfillment — Amazon being the clearest example — where agent-driven purchasing bypasses the ad stack entirely. Retailers that think of themselves as destination brands but are really just fulfillment operations will face pressure to decide which side they are actually on.

The consumer empowerment angle

Acharya frames the agent-driven demand for refunds and payment disputes not as a merchant threat but as a consumer access story. Bureaucratic processes — changing schools after a move, navigating insurance after illness, fighting an incorrect charge — have always been tilted toward people who could afford a lawyer or had time to pursue small claims. Agents effectively democratize access to those processes, and Acharya treats that as an investment opportunity rather than a problem to hedge against. Brazil, where anyone can file a lawsuit digitally with minimal friction, gives a preview of how adversarial this can get: merchants there already buy specialized software to manage the volume of digital claims.

Agent wallets and the economy beyond

The more speculative thread is what happens when agents become independent economic actors with their own payment credentials. A dedicated agent credit card — attached to something like an Instinct or Muse — would let a personal agent service monetize transaction volume without needing opt-in agreements from every long-tail retailer it touches. Acharya acknowledges the take rates on card transactions alone are thin, but sees the larger prize as agents that can seek out work, get paid, and deploy specialized skills learned from their users. That agent-to-agent economy is further out, but people are already scaling outbound sales operations that effectively function as human proxies for tasks models can already do — the transition is starting at the edges.

The immediate constraint remains compute cost. Until that $1,000 per user per year figure drops by an order of magnitude, personal agent services are fighting a two-front battle: customer acquisition and infrastructure spend at the same time.