How do I set spend and cost limits to stop an AI agent from racking up runaway bills?

To prevent an AI agent from incurring excessive costs, implement explicit monetary ceilings and integrate cost accounting mechanisms that track and gate spending.

• Implement explicit cost ceilings and autonomy levels: Actions with significant monetary impact require explicit ceilings. Autonomy settings for an agent's capabilities, such as spending money, should range from fully autonomous to never-permitted, with a default of "not permitted unless explicitly authorized". For consequential actions, target autonomy levels 1 or 2, where agents perform narrow, low-risk tasks with human review or act autonomously within a defined scope and escalate at boundaries. This aligns with the NIST AI RMF function of Govern by establishing clear policies and procedures for agent operation.
• Utilize cost accounting and threshold gates: Implement a system that normalizes provider-specific usage into a canonical schema, multiplies tokens by pricing tables, attributes spend, accumulates session totals, and warns or gates the user when spending crosses a threshold. For interactive human-in-the-loop workflows, an interactive gate can interrupt the user at a fixed threshold and require acknowledgment before continuing. This addresses the OWASP LLM Top 10 risk of LLM04: Denial of Service by preventing runaway resource consumption.
• Track costs across sessions: Ensure that costs persist across sessions, so that cumulative spend over multi-day operations is accurately reflected. This allows for a comprehensive view of an agent's financial impact over time.
• Employ precise cost calculation and provenance: Use Decimal-based math for per-million pricing to avoid accumulated float drift, especially when summing many small per-call costs. For multi-provider environments, use accounting systems that provide provenance on every cost number, including status (actual/estimated) and source, to satisfy finance, FinOps, or compliance requirements. This supports the ISO/IEC 42001 control of A.7.2.1 AI system logging by ensuring auditable cost records.
• Budget memory and background compute: Recognize that agent "dreams" (memory consolidation) cost money at standard API token rates and that background compute can create operational load. Implement budgeted injection with retrieval policies and token budgets for each memory layer to manage these costs.