Control · C-091 · Monitoring, Assurance & Incidents
AI cost and capacity monitoring
Usage, latency and spend for each AI system are tracked against budgets and limits, with alerts and automatic caps before a limit is breached.
- Control type
- Detective
- Risks addressed
- 3
- Audience
- Heads-of · Practitioner
- Frameworks
- MAP 3 · Environmental
- Responsible AI domain
- Lifecycle stage
- Reviewed
- 2026-11-11
What it does
It keeps consumption visible and bounded. For each system, the volume of requests, the size of prompts and outputs, response times and cost are measured in near real time and compared with what was planned. Alerts fire as limits are approached, and hard caps stop runaway use. Spend is attributed to the system and team that incurred it. The system owner is accountable for staying within budget; engineering and finance operate the monitoring. It runs continuously.
How to implement it
- Set a budget and a capacity plan for each system at approval.
- Meter usage by system, feature and team, and attribute cost accordingly.
- Alert at set fractions of budget and quota, and cap at the limit.
- Limit steps, retries and context size in agentic and looping workflows.
- Track cost per transaction over time, and investigate drift.
- Review forecasts against actuals, and feed the result into the next business case.
Evidence it operates
Design: budgets, quotas and alert thresholds for each system. Operation: usage and spend reports; alert history and responses; caps that triggered; variance reviews; revised forecasts.
Where it fits
In operation, alongside technical monitoring. It is the primary control for AI capacity and cost overrun and supports unbounded resource consumption. It complements inference usage quotas and rate limiting, which enforces limits at the interface.
Limits
Provider billing data can lag by hours or days, so a fast runaway may outrun the alert. A hard cap protects the budget by stopping the service, which may be the worse outcome for a critical process. Pricing changes are outside the organisation's control. Cost pressure can push teams to cheaper models with unassessed quality.
Risks this mitigates
Primary mappings
- AI capacity and cost overrun Medium
- Tracks usage, latency and spend against limits with alerts. Usage, latency or spend exceeds what was planned, degrading the service or the budget, without any attacker being involved.
- Environmental footprint of AI Medium
- Measures compute use, the main driver of footprint. The energy, water and hardware that training and running AI require conflict with the organisation's environmental commitments, or go unmeasured and undisclosed.
Supporting mappings
- Unbounded resource consumption Medium
- Alerts when spend or capacity crosses a limit. With no limit on the volume, size or cost of requests, a user or attacker drives inference far beyond what the system was sized or budgeted for, degrading service or running up charges.
Framework and clause references
| Framework | Clause | Title |
|---|---|---|
| NIST AI Risk Management Framework (AI RMF 1.0) | MAP 3 | Capabilities, intended use, benefits and costs are understood Costs understood. |
| NIST Generative AI Profile (NIST AI 600-1) | Environmental | Environmental Impacts Compute as the driver of footprint. |
Related terms
How to cite
Axiom Verity (2026). AI cost and capacity monitoring. Reference Library, C-091. https://axiom-verity.com/controls/ai-cost-and-capacity-monitoring. Reviewed 2026-11-11.
Found an error or a better source? Report a correction.