Give every AI agent a token budget before it becomes a surprise invoice

An agentic task can consume roughly 1,000 times the tokens of a standard query. Finance asks why the invoice jumped. The answer is often an agent that looped, retried, and burned tokens while nobody watched.

Agentic work calls tools and summarizes long documents, so treat each agent like a cost center with a budget and an alarm.

In the FinOps Foundation's 2026 State of FinOps survey, 98% of practitioners now manage AI spend and consider FinOps for AI their top forward-looking priority. In a May 2026 McKinsey Enterprise AI FinOps Survey, 93% of enterprises had gone over their AI budgets. The control is a pre-flight token budget for every agent, set before it scales.

The budget is the control, not the dashboard

A dashboard tells you what happened. A budget tells you what is allowed. Assign each agent a token tier, a cap, and an alarm before it gets broad access. The record should name the agent, owner, model, expected monthly tokens, cap, alert threshold, data access, and review date.

Cohere's North platform had been generally available since August 2025 before the North 2 upgrade. North 2's management console, North Admin, reports token consumption for each user and each agent. North 2 lets administrators create usage tiers based on request and token rates and apply an organization-wide cap. North Admin gives administrators live analytics for AI usage, showing token use and limits. The control you want is per-agent visibility, tiered limits, and a company ceiling.

The pre-flight token budget checklist makes agent spend enforceable

• Inventory every agent. List each agent, its owner, the model it calls, the tools it can use, and the data it can reach. The result is a table where no agent is anonymous. If an agent can call itself or loop, mark it for tighter limits.
• Rank by spend and risk. Put high-loop agents, long-document agents, and agents touching customer data at the top. The result is a ranked list with the highest-spend, highest-risk agents first. A chatbot and a research agent are not the same cost class.
• Set the tier, cap, and alert threshold. Give each agent a token tier, a monthly or per-session cap, and an alert threshold below the cap. The rule should say how many tokens the agent may use, when it warns, and where it stops. Set it before the agent gets more users, not after usage has already run high.
• Add privacy and injection guardrails. North 2's agent guardrails check prompts and responses for personally identifiable information and flag attempted prompt injection. Test the guardrail: paste a prompt that asks the agent to reveal PII, and it should block or redact the answer. North 2 provides access-control lists that administrators can attach to libraries and skills. Cohere is planning North 2 connectors for PitchBook Data, S&P Global, and FactSet Research Systems. If an agent can reach those sources, the access-control list should say exactly what it can read.
• Review on a cadence. Revisit the budget monthly, after model changes, and after any incident. The record should show spend versus cap, alerts fired, guardrail blocks, and whether the tier still fits the work. Tetrate Agent Router Enterprise enforces token budgets by tripping a circuit breaker when spend crosses policy and then falling back to an approved alternative. Use that pattern when a budget is a routing decision: stop the expensive path, keep the work moving, and log the exception.

The tooling can live inside the agent platform or across providers. 1Password's AI Spend and Consumption Management consolidates AI usage data from multiple providers and lets organizations configure vendor spend caps, threshold notifications through Slack or email, and usage slices by team, user, vendor, and model. With Ramp's AI Token Spend Management, finance teams can cap spending by team, project, or API key and send anomaly warnings to relevant stakeholders before usage goes over plan. Pick the tool that gives you the limit, the alert, and the owner.

The budget answers finance

Finance does not care about model names. It cares about owner, limit, alert, and exception. 1Password's AI Spend and Consumption Management records token usage at the API level whether the request originates from a human or an AI agent. Put the agent's budget in the same place as the team's budget, and the conversation changes from why usage is high to which agent crossed its limit.