LLM Cost Management for AI Agencies

Why agency LLM costs are difficult to control

A single provider invoice can contain usage from multiple clients, agents and workflows. That makes it difficult to answer basic commercial questions: Which engagement generated the cost? Which agent used a premium model? Is usage still covered by the client’s plan? Did efficient routing improve the margin? The risk increases when an agency sells a fixed retainer or usage allowance while paying variable model costs. More requests, longer prompts or unnecessary premium-model calls can reduce gross margin without changing revenue. Reporting alone does not solve this problem if the report arrives after the budget has already been consumed. Agencies need cost attribution, routing rules and enforceable limits at the point where requests are made.

Create a controlled cost boundary for every engagement

Give each client or agent a dedicated AgentCost API key instead of sending unrelated workloads through one shared credential. AgentCost records request volume, spend and estimated savings, providing a clearer cost trail for each operating unit. A practical structure is to separate keys by client first and by agent when the engagement contains materially different workloads. For example, a support agent and a document-analysis agent may have different request volumes, model requirements and budget expectations. Separate tracking makes those differences visible instead of averaging them into one client total. Keep every key in a server-side environment variable or secrets manager. Do not expose it in browser code, mobile applications, repositories or support messages. Use separate keys for production and other environments when you need their usage and permissions isolated.

Route routine work without making premium models the default

Model selection should reflect the task rather than a single default for every request. Agencies can configure routing logic so suitable work is directed toward a more cost-effective configured model while reserving higher-cost models for tasks that genuinely require them. Start by grouping workloads according to complexity and risk. Classification, extraction, short summaries and templated responses may be candidates for lower-cost routing. Complex reasoning, high-value generation or tasks with strict quality requirements may justify a more capable model. Test output quality against your own acceptance criteria before changing production routing. The decision rule should be operationally clear: use the least expensive configured model that consistently meets the engagement’s quality, latency and reliability requirements. AgentCost records estimated savings per request, but those estimates should support—not replace—quality review.

Set budgets before production traffic reaches them

Translate each client agreement into daily and monthly operating limits. AgentCost can enforce configured ceilings by rejecting requests once a limit is reached, rather than only showing an overage after it occurs. Use a monthly limit to reflect the engagement’s total LLM allowance. Add a daily limit when traffic spikes could consume that allowance too early. The daily ceiling should not simply be the monthly amount divided by the number of days: account for expected weekday patterns, launches, batch jobs and seasonal peaks. Decide what happens when a request receives a non-2xx budget response. Customer-facing applications may need a clear temporary message, while internal workflows may pause for review. The application should handle this response deliberately; repeatedly retrying a budget rejection will not create more available budget.

Calculate margin using attributable model spend

For a fixed-fee engagement, a useful starting formula is: Gross margin = client revenue − delivery labour − attributable LLM spend − other direct delivery costs. Gross margin percentage = gross margin ÷ client revenue × 100. For usage-based work, compare invoiced usage revenue with the corresponding attributable LLM spend and direct service costs. Review both the total client margin and the contribution of individual agents. A profitable engagement can still contain one inefficient workflow that is being subsidised by the others. Illustrative example: if an agent’s monthly revenue allocation is €1,000, delivery labour is €500, attributable model spend is €180 and other direct costs are €70, gross margin is €250, or 25%. This is only an example; real results depend on your contract structure and accounting policy. Use the agency margin calculator to test your own revenue, workload and cost assumptions.

Use a weekly and monthly margin review cadence

Weekly review is for intervention. Check request volume, spend against daily and monthly limits, unusual changes by agent, premium-model usage and failed requests caused by exhausted budgets. Investigate changes against deployments, prompt revisions, traffic growth or new automation steps. Monthly review is for commercial decisions. Reconcile attributable LLM spend with the client agreement, compare actual usage with the included allowance, review estimated routing savings, and calculate the engagement’s gross margin. Then decide whether to adjust routing, change a budget, revise an allowance or discuss pricing with the client. Do not present estimated savings as guaranteed cash savings. The estimate depends on available usage and pricing data, model choice, workload and routing configuration. Record the comparison basis used so the figure remains explainable during a client review.

Client cost-review checklist

Use this copy-ready checklist before a monthly client meeting: • Confirm the reporting period and the client keys included. • Record total requests and attributable LLM spend. • Break out material agents or workflows rather than reporting only one blended total. • Compare actual spend with the agreed monthly allowance or internal ceiling. • Identify any daily or monthly limit events and explain their operational impact. • Review premium-model usage and confirm that routing still matches task requirements. • Record estimated savings and state the model or routing comparison behind the estimate. • Compare revenue, labour, LLM spend and other direct costs to calculate gross margin. • Note workload changes that may affect the next period. • Agree any routing, budget, allowance or pricing changes and assign an owner. The checklist is most useful when the same definitions and reporting period are used every month. That turns cost management into a repeatable client operating process rather than an invoice investigation.

Connect AgentCost without rebuilding your request format

AgentCost exposes an OpenAI-compatible chat completions endpoint at https://agentcostai.eu/v1/chat/completions. Existing OpenAI SDK integrations can use the AgentCost base URL and an AgentCost client key while keeping the familiar chat-completions request shape. Start with a minimal request containing the model and messages fields. Add optional parameters such as temperature or max_tokens only where supported and needed. Before sending production traffic, create the appropriate client or agent key, store it securely on the server, configure daily and monthly budgets, and ensure the application handles authentication, provider and budget errors. This controlled path lets the agency apply routing and budget rules before forwarding requests, then review usage, cost, limits and estimated savings in the portal.

Frequently asked questions

How does AgentCost help an AI agency protect margins?

AgentCost separates client or agent usage with dedicated API keys, routes requests toward configured cost-effective models, enforces daily and monthly budgets, and records spend and estimated savings. Agencies can use that data to calculate engagement margin and identify inefficient workflows.

Can AgentCost stop a client from exceeding a configured budget?

Configured daily and monthly controls can reject requests once a limit is reached. Your application should handle the resulting non-2xx budget response and decide whether to pause work, notify a user or request a budget review.

Should every AI agent have its own key?

Use one key per client at minimum when client-level attribution is required. Separate agents when their workloads, permissions, budgets or margin impact need to be reviewed independently. Avoid unnecessary fragmentation if several agents belong to the same cost centre and do not need separate controls.

Are estimated savings guaranteed?

No. Savings vary with workload, provider pricing, model selection and routing configuration. Treat portal savings as an estimate based on available usage and pricing data, and document the comparison basis used in client reporting.

What should an agency review each month?

Review requests, attributable spend, budget usage, premium-model calls, estimated savings and cost by material agent. Reconcile those figures with client revenue, labour and other direct costs, then decide whether routing, limits, allowances or pricing should change.

Will an existing OpenAI SDK integration need to be rewritten?

AgentCost provides an OpenAI-compatible chat completions endpoint. You can configure the SDK with your AgentCost client key and the https://agentcostai.eu/v1 base URL while retaining the familiar request format. Test the integration and error handling before moving production traffic.