Based on research from McKinsey, Deloitte, IDC, and earnings call disclosures from major tech companies, the typical enterprise AI budget breaks down roughly as follows.

Model API costs and experimentation consume 30-40% of the total AI budget. This includes direct API spend with OpenAI, Anthropic, Google, and others, plus the engineering time spent evaluating models, running proofs of concept, and building demo applications. The problem isn't that experimentation is bad. It's that experimentation without infrastructure produces experiments that never scale. Dozens of teams running independent POCs with independent model contracts and independent integration work is the definition of duplicated effort.

Custom development takes another 25-35%. This is the engineering time spent building bespoke AI solutions: custom RAG pipelines, custom agent frameworks, custom integrations, custom governance layers. Each team builds its own version because no shared infrastructure exists. The custom development line item is where the build-vs-buy decision plays out most visibly. Learn more

Point tool licenses account for 15-20%. Enterprise search tools, AI copilots, chatbot platforms, specific-purpose AI applications. Each solves one problem. Each has its own pricing model, its own integration requirements, and its own security review. A mid-size enterprise running five AI point tools is spending $500K-$2M annually on licenses alone, before considering the integration costs to make them work together.

Infrastructure gets 10-15%. This is the investment in the foundation: context layers, AI gateways, governance frameworks, orchestration platforms, memory systems. The category that determines whether all the other spending compounds or fragments.

Consulting and professional services take 5-10%. Implementation support, AI strategy consulting, change management. This category has grown significantly with OpenAI's Frontier launch and similar services-heavy enterprise AI approaches.