McKinsey Experts Evaluate the AI Cost Paradox
McKinsey partners Tanguy Catlin and Lari Hämäläinen discussed why corporate spending continues to rise despite AI models becoming cheaper.
During a virtual session where they shared findings from the firm's State of AI 2026 survey, McKinsey senior partners Tanguy Catlin and Lari Hämäläinen addressed the paradox that corporate spending continues to increase even as unit costs for artificial intelligence decline.
Decline in AI Costs and Increasing Spending
McKinsey senior partners Tanguy Catlin and Lari Hämäläinen evaluated the cost contradiction in the field of artificial intelligence during a virtual session. It was noted that intelligence at a specific capability level has become much more economical.
For example, compared to when GPT-4 was launched, models with similar performance can now be operated at much lower costs. Despite this, companies continue to allocate significantly more budgets to artificial intelligence than before.
Consumption Explosion of Autonomous Agents
While the cost of producing a unit of intelligence is falling, businesses' consumption volumes are growing rapidly. As the unit capability cost of models becomes cheaper, companies demand much more reasoning and work from them through autonomous agents.
In software development processes, AI agents can repeatedly review and modify entire codebases, producing far more code than a human could ever touch.
Economic Dynamics of Agentic AI
Companies are only just beginning to understand the economics of agentic AI. Unlike traditional software, the cost of running a task is far from predictable because agents can reach the same outcome through different paths.
The same task can generate a cost up to thirty times higher from one run to another. A large portion of this cost stems from the reasoning and continuous improvement processes behind the final output.
System Design and Task-Level Evaluation
System design choices directly impact the cost of task completion. It is emphasized that leaders should evaluate agents at the task level rather than measuring them by cost per token.
The cost of a task is analyzed by taking into account the agent success rate and the human time required to verify its work. Agents begin to create value when verification time is very low compared to human time.
Methods for Managing AI Spending
Tanguy Catlin, director at the McKinsey Global Institute, shared three key areas where companies can manage their AI spending. First, visibility must be established regarding which use cases are driving expenses.
Second, it is recommended to optimize workflows, adapt model complexity to the task, and limit unnecessary tool calls. Finally, procurement discipline and avoiding vendor lock-in are of great importance.