As enterprises race to deploy agentic AI systems, one stubborn challenge keeps surfacing: runaway token spend. A new report from Business Insider highlights how organizations are wrestling with the cost of large language model tokens while trying to keep autonomous agents focused on their goals. The tension is real—token budgets can balloon quickly, but cutting corners on context can derail an agent's performance.

Why Token Costs Are Spiking in Agentic Workflows

Agentic AI systems rely on iterative reasoning, tool calls, and long context windows to complete multi-step tasks. Unlike simple chatbot queries, these agents often consume thousands of tokens per task—and when scaled across an enterprise, the bills add up fast. The Business Insider report notes that teams are increasingly surprised by how quickly costs escalate when agents need to re-read documents, retry actions, or maintain conversational memory.

The root cause is often over-provisioning: giving the agent more context than necessary, or using expensive high-capability models for routine subtasks. Without careful design, token consumption can spiral, making otherwise promising pilots financially unsustainable.

The Context Window Dilemma

Larger context windows are a double-edged sword. While they allow agents to handle complex workflows, they also encourage developers to stuff in irrelevant data—driving up costs per request. The report suggests that organizations must strike a balance between sufficient context and minimal token usage.

Practical Strategies to Trim Token Usage

Experts cited in the article recommend a multi-pronged approach to keep agentic goals on track without breaking the bank. The key is to treat tokens as a finite resource that must be allocated intentionally.

  • Right-size your model selection: Use smaller, cheaper models for simple tasks and reserve frontier models for complex reasoning.
  • Condense context aggressively: Summarize or filter documents before feeding them to the agent, removing redundant or low-value text.
  • Implement caching: Reuse common prefixes or system prompts to avoid re-billing the same tokens repeatedly.
  • Set hard token budgets: Enforce per-task limits and add escalation paths when agents exceed them.

Another tactic gaining traction is structured output prompting, which forces the model to return only essential data, cutting down on verbose responses. Additionally, periodic "token audits" can reveal which workflows are cost outliers, enabling targeted optimization.

Monitoring and Cost Attribution

The report emphasizes that you cannot manage what you do not measure. Teams should instrument their agent pipelines to track token consumption by task, model, and user. This visibility allows finance and engineering to collaborate on cost controls without stifling innovation.

Keeping Agentic Goals in Focus

Cost management should not come at the expense of agent effectiveness. The article warns against over-optimizing to the point where agents lose critical context and start making errors—which can be far more expensive than the tokens saved. Instead, the goal is to optimize for outcome per token, not just minimize raw usage.

One recommended approach is to use a tiered reasoning pattern: let a cheap model handle initial steps, and only escalate to a premium model when the task's complexity demands it. This keeps average costs low while preserving the highest quality for the most challenging subtasks.

Also, consider asynchronous workflows where agents can pause and resume instead of keeping long conversations alive. This reduces the need to re-send entire histories, lowering token overhead significantly.

Long-Term Cost-Model Design

Looking ahead, the report suggests that organizations should negotiate pricing models with providers that reward efficiency, such as batch processing or reserved capacity. As the market matures, new pricing structures may emerge that better align with agentic workloads.

Key Takeaways

Managing token costs in agentic AI is not about squeezing every penny—it is about building sustainable systems that can scale. The Business Insider report offers a clear message: with deliberate design, monitoring, and a willingness to adjust models and context, companies can keep their agents both effective and affordable.

For teams just starting their agentic journey, the first step is simple: measure your current token usage, identify the biggest culprits, and apply one or two of the strategies above. Small changes can yield significant savings without sacrificing the intelligence that makes agents valuable.