AI agents capable of operating computers have crossed a dual inflection point in both cost and accuracy according to new research published by Andreessen Horowitz partner Fabrizio Serafini. This development marks a turning point for the business process outsourcing industry. The fully loaded cost of running an AI agent now ranges from $6 to $8 per hour and sits below the roughly $10 per hour cost for offshore business process outsourcing labor in India. These agents also outperform humans on standard desktop task benchmarks.

Accuracy Surpassing Human Testers

Leaderboard data from llm-stats.com cited in the a16z research shows the top-performing AI model Claude Fable 5 achieves an 85 percent task completion rate on the OSWorld-Verified benchmark. This benchmark simulates real desktop environments across Ubuntu Windows and macOS. Human testers score approximately 72 percent on the same tasks. A year ago the best model managed just 42 percent completion rate.

The benchmark evaluates whether an agent can independently perform tasks such as browser navigation form filling and button clicking. In production enterprises build on raw APIs rather than consumer-facing products. They wrap model capabilities in sandboxed virtual machines with custom orchestration and retry logic. One founder told the a16z team that models were not good enough to use in production on their own until Opus 4.6 in February 2026. Still an 85 percent completion rate means 15 failures per 100 tasks and back-office processes demand success at every step.

Serafini emphasized that decisive factors lie beyond the model itself. These include validation mechanisms anomaly reporting error handling and adaptation when for example a retailer's portal redesigns its layout overnight.

The Economic Case

The cost comparison draws on multiple sources. AI agent inference runs approximately $6 to $8 per hour in the most expensive mode which involves frame-by-frame screenshot processing with frontier models. A practical range of $3 to $15 applies depending on screenshot frequency context length and how much work is delegated to deterministic code. Offshore BPO labor in India costs roughly $8 to $15 per hour based on 2026 pricing data from Globalify HiveDesk and 1840 & Co. U.S. back-office employees cost $30 to $45 per hour when benefits and overhead are included.

Serafini cautioned that these figures should be treated as order-of-magnitude estimates. Well-built systems reserve expensive model calls for tasks that require judgment and let cheap deterministic code handle repeatable steps. This approach drives blended costs down further over a workflow's lifetime.

Production Deployments at Scale

Production deployments are already running at scale. A consumer goods data platform processes 15 to 20 million automated portal interactions monthly. It uses agents as a self-healing fallback that allowed it to halve its scraper maintenance engineering team. A global systems integrator runs 27 live workflows processing 1,500 to 2,100 IT tickets daily.

The Moat Shifts to Context

As raw UI navigation becomes a commodity at the model layer Serafini argues the durable competitive advantage has moved to the application and context layers. This covers understanding a company's internal terminology preferred formats escalation paths and failure-handling protocols. One operator running millions of automated tasks monthly told the a16z team he does not even know which underlying model powers his system. His vendor swaps models underneath him and his trust lies in the system's consistent ability to complete tasks reliably.