Enterprise generative AI spend increases fivefold to $10m as coding agents drive adoption

Enterprise expenditure on generative artificial intelligence grew fivefold over the past year, according to survey data published by Andreessen Horowitz.

Average annual commitments rose from approximately two million dollars to over ten million dollars per organisation. This expansion marks a decisive transition from exploratory sandbox pilots to production deployments embedded directly within core business units.

Software development represents the primary catalyst for this capital influx. Coding assistants and autonomous engineering agents have demonstrated immediate velocity gains, establishing technical departments as the commercial beachhead for generative adoption. Rather than purchasing broad productivity tools, buyers increasingly allocate capital to specialised programming platforms capable of generating, testing, and reviewing pull requests across large codebases.

Procurement patterns reveal growing demand for architectures that operate with reduced human supervision. Where initial deployments favoured simple internal chatbots and basic retrieval-augmented generation systems, current outlays target multi-step autonomous workflows. These agents execute sequential operations across enterprise resource databases, customer management platforms, and ticketing infrastructure.

Budget mechanisms supporting this procurement have matured in tandem with model capabilities. Early enterprise trials drew predominantly from discretionary research reserves or dedicated innovation pools. Current allocations pull directly from core software budgets and departmental operational expenditure. This reallocation exerts immediate pressure on traditional software vendors, whose seat-based subscription models face displacement by autonomous workflow tools.

Infrastructure diversification accompanies the increase in enterprise commitments. While commercial foundation models still capture the majority of application spend, enterprises increasingly run heterogeneous environments. Engineering teams deploy proprietary frontier models for complex reasoning tasks while directing routine tasks to smaller, fine-tuned open-weight alternatives to manage inference expenditure.

With corporate buyers demanding verifiable efficiency returns on their generative outlays, will autonomous agent platforms sustain this capital intake once early developer gains plateau?