Rising consumption costs could slow the adoption of generative and agent-based AI in the enterprise sector: companies are struggling to correlate growing token usage with measurable returns.
"Token shock" is the budgetary pressure that occurs when AI consumption grows faster than the price per token declines. Software providers are increasingly moving away from fixed, unlimited subscriptions in favor of pay-as-you-go pricing to cover inference costs and protect margins.
Agent-based AI exacerbates the situation. A multi-step task can require multiple model calls, wider context windows, and more tokens for reasoning than a simple query. In one project studied by EY, switching from direct queries to large language models to agent-based workflows increased the cost per interaction by approximately 30 times. Enterprise users are already responding by imposing spending caps. Uber Technologies (NY:UBER) reportedly exhausted its 2026 AI programming budget in four months and set a $1,500 per employee cap for each agent development tool.
Furthermore, a major French insurer scaled back its use of Anthropic's Claude two months after launch when costs exceeded expectations. Walmart (NY:WMT) also limited employee access to its internal AI agent, despite previously providing unlimited tokens.
Discussions with tech and consulting companies revealed that many clients have postponed, scaled back, or canceled at least one AI project due to concerns about costs and the uncertain economic environment. One vendor estimated that less than a quarter of its pilot projects had yielded positive returns.
Inference costs can account for approximately 80% of the total costs over the lifecycle of an AI model. This shifts the focus from the initial costs of training models to the long-term costs of running them in production.
Companies manage these costs by reserving advanced models for complex tasks and routing simpler queries to smaller, cheaper alternatives. Open-source models and proprietary infrastructure are also gaining increasing attention for predictable, high-volume workloads.
The outlook remains cautiously optimistic, although enterprise adoption may take longer than expected. Software development companies can respond with specialized models, improved routing, and performance-based pricing. HubSpot (NY:HUBS), for example, has already implemented performance-based pricing rather than token consumption.