McKinsey Warns: AI Agent Costs Can Vary by Up to 30x
nashnova research
McKinsey's latest survey finds that AI agents can cost up to 30 times more or less to complete the same task, as enterprise AI spending accelerates with far greater volatility than traditional text-based AI tools — putting budget planning in uncharted territory.
Why can the same task cost 30x more?
AI agents — programs that autonomously execute multi-step tasks — don't just answer questions like a chatbot. They call models repeatedly, consuming large volumes of tokens (the billing unit for AI models), and different execution paths produce wildly different bills.
McKinsey senior partner Lari Hämäläinen put it bluntly: "Imagine running a business where costs differ by 30x from one day to the next."
This means → Companies can't budget for AI agents the way they budget for cloud servers. The spend looks more like an open-ended tab with no predictable ceiling.
Where is the money going?
Software development is the fastest-growing AI spend category. Using agents to automate coding is "extremely token-intensive," with single-task call volumes far exceeding ordinary conversations.
McKinsey's 2026 survey: roughly one-third of companies now allocate over 10% of their tech and communications budget to AI; 60% plan to increase spending next year.
About one-fifth of respondents say AI spending is already squeezing operating costs. Senior partner Tanguy Catlin was direct: "The spending is now substantial and visible."
Can the efficiency gains cover the costs?
A McKinsey report this month showed AI agents can cut human time on specific office tasks by 35% to 40%, sometimes exceeding 70%.
But McKinsey also stressed: efficiency gains alone don't mean the math works — companies must verify that the value agents generate actually covers what they cost.
In plain terms = the labor hours saved are the "return"; the tokens consumed are the "cost." If the cost side grows faster than the return side, the investment is underwater.
Are companies already hitting the brakes?
Signs of deliberate AI-usage limits are emerging: Coinbase and Salesforce have started capping AI use.
Amazon's case is more vivid — employees created an informal AI-token leaderboard, some gaming it by running tasks just to climb the ranks. Amazon shut the leaderboard down.
This reflects a deeper issue: when AI use has no cost constraint, waste happens automatically. Hämäläinen said: "There's a lot we're only beginning to figure out — the next 12 months will be critical."
市场有风险，内容仅供研究参考，不构成投资建议。