The cost reality

Traditional software had fixed licenses; cloud had predictable consumption. Agents introduce a third model: variable, usage-driven costs that multiply non-linearly. The baseline numbers are not projections; they are current production numbers.

The three forces at working resolution: the new economic model itself, the pricing reset shifting risk onto buyers, and the infrastructure question underneath both.

A structural shift

The economics of agentic AI are fundamentally different from every previous enterprise technology cycle. Traditional software had fixed licensing costs. Cloud shifted to predictable consumption. Agents introduce a third model. Variable, usage-driven costs that multiply non-linearly with complexity, concurrency and failure. According to McKinsey this is a structural shift from labor to technology as the dominant cost driver. In knowledge-intensive industries technology spending could ultimately even exceed labor costs. Gartner projects that by 2028, AI coding costs driven by ungoverned consumption will be as much per developer as the salary companies pay that person. This is not an incremental change. It is a different economic model.

"Cognitive tax" is an emergent new strategic dependency and it describes value accruing to intelligence infrastructure providers in the same way cloud providers captured value in the previous era. This isn't a line-item cost problem. It is a structural shift in where enterprise value concentrates.

The baseline numbers

The baseline numbers are staggering. Agentic systems require 5-30x more tokens per task than standard conversational tools. A single complex orchestrated interaction in 2026 costs 30x more than a simple workflow interaction in 2023. Agents make several times more LLM calls than chatbots per user request, since each one triggers planning, tool selection, execution, verification and response generation. When multiple agents run concurrently, costs multiply non-linearly. These are not theoretical projections. They are current production numbers.

The pricing reset

The cost structure is also shifting underneath buyers. SaaS vendors are abandoning per-seat pricing while adopting new models that transfer forecasting risk directly onto buyers. Outcome-based pricing sounds appealing but remains difficult to implement, because measuring attribution in complex enterprise environments is genuinely hard. Some enterprises are already spending over $1200 per employee annually across overlapping AI tools. Gartner projects 40% of enterprise SaaS spending will shift to usage, agent or outcome-based models by 2030. The underlying token price itself is volatile too. The Silicon Data LLM Token Expenditure Index, a blended rate in US dollars per million tokens across providers, stood at 1.62 in July 2026, down 20% from its May 2026 peak though still above its December 2025 inception level. Why is unclear: vendor price cuts, a backlash against AI spending, or enterprises switching to less token-heavy models are all plausible. Without cost governance and usage discipline, organizations will face invoice surprises that undermine financial planning.

Infrastructure economics

On the infrastructure side, the global AI inference market is projected to reach $48.8 billion by 2030. Unlike training, which is periodic and latency-tolerant, inference is real-time, latency-sensitive and unrelenting. Organizations running inference on general-purpose infrastructure can pay 2x per million tokens compared to inference-optimized environments. Right-sizing inference infrastructure is now a strategic business decision and not merely an engineering one.

The total cost of agentic AI includes infrastructure, governance, organizational change, failure recovery and regulatory risk. Not just tokens.