--- title: "AI Spend and Token Tracking" canonical: https://harmony.io/platform/ai-governance/ai-spend-management --- # AI Spend and Token Tracking Token usage and estimated cost per person for Claude, usage for ChatGPT, Cursor and Microsoft Copilot, and approved spend-limit increases instead of surprise invoices. ## AI spend management per person, per model See who uses which AI tools, how many tokens they consume and what it costs, and handle requests for more usage with an approval instead of an open budget. ## What is AI spend management? AI spend management is knowing what your company spends on AI tools, who drives that spend and whether it is worth it. Unlike a seat license, usage-based AI pricing changes every month with how much people use it, and each provider's console shows API keys and totals, not your employees. Harmony connects to each provider's admin API and maps usage to employees. For [Claude](https://harmony.io/integrations/claude), that is token usage per person across Claude and Claude Code, broken down by model and API key, with an estimated cost in US dollars. For [ChatGPT](https://harmony.io/integrations/chatgpt), it is daily token counts per person and model from the OpenAI API platform. For [Cursor](https://harmony.io/integrations/cursor), it is activity, request volume and each person's most-used model. For [Microsoft Copilot](https://harmony.io/integrations/microsoft-copilot), it is licensed users and usage events. When someone needs more, the AI Usage Limits agent handles it. The employee asks, an approver confirms, and Harmony raises that person's Claude Enterprise or Cursor Enterprise spend limit, doubling it by default within a cap. The increase is temporary by default and resets when the period ends. ## AI usage you can see and approve - **Tokens per person:** Claude and Claude Code input and output tokens for every employee, by model. - **Estimated cost:** An estimated cost per person for Claude, with seat-based, consumption-based and hybrid pricing supported. - **ChatGPT usage:** Daily token counts per person and model from the OpenAI API platform, refreshed every six hours. - **Cursor activity:** Agent, chat and Composer requests, and each person's most-used model. - **Microsoft Copilot:** Licensed users and usage events, collected daily. - **Usage by department:** Active users of each AI tool by department and job title. - **Spend-limit requests:** The AI Usage Limits agent raises a Claude Enterprise or Cursor Enterprise limit after approval, temporarily by default. - **Ghost and inactive users:** AI seats nobody uses, counted on each tool's page. ## How AI spend tracking works **Trigger:** A team adopts a new AI tool, usage spikes, or someone hits their spend limit. 1. **Connect each provider** - Add an admin or analytics key for Claude, the OpenAI API platform or Cursor, and connect Microsoft Copilot. 2. **Map usage to people** - Usage is matched to employees, so tokens and cost show by person, department and model. 3. **Handle limit requests** - When someone needs more, they ask. An approver confirms and the agent raises their limit. 4. **Reset on schedule** - Temporary increases go back to the default when the period ends. **Outcome:** AI usage you can see by person and approve by policy. ## Fireblocks grants AI model access for 12 hours at a time At Fireblocks, an employee tags Harmony in Slack asking for access to a new AI model. The agent asks for a business justification and uses AI to turn away the empty ones. On a real justification it adds the user to the Okta group, and 12 hours later it removes them. It went live two days after the spec was posted and has run 269 times since. ## Agents for AI tool governance - [Application Access Request](https://harmony.io/agents/application-access-request) - [Discovered Applications Review](https://harmony.io/agents/discovered-applications-review) - [Unauthorized Applications Detection](https://harmony.io/agents/unauthorized-applications-detection) - [Unused Application Discovery](https://harmony.io/agents/unused-application-discovery) ## Usage from the AI tools your teams use - **AI tools: **[Claude](https://harmony.io/integrations/claude), [ChatGPT](https://harmony.io/integrations/chatgpt), [Cursor](https://harmony.io/integrations/cursor), [Microsoft Copilot](https://harmony.io/integrations/microsoft-copilot) - **Identity: **[Okta](https://harmony.io/integrations/okta), [Microsoft Entra ID](https://harmony.io/integrations/microsoft-entra-id), [Google Workspace](https://harmony.io/integrations/google-workspace) ## FAQ ### What is AI spend management? Tracking what a company spends on AI tools, who drives that spend and on which models, so usage-based costs are visible and increases are approved rather than discovered on the invoice. ### Which AI tools does Harmony track? Claude and Claude Code (tokens and estimated cost per person), ChatGPT through the OpenAI API platform (daily tokens per person and model), Cursor (activity, requests and most-used model) and Microsoft Copilot (licensed users and usage events). ### Does Harmony show the cost of AI usage? For Claude, yes: an estimated cost per person, by model and API key. ChatGPT and Cursor report usage through their APIs but not cost, and Microsoft Copilot reports usage events rather than tokens. ### Can Harmony enforce AI spend limits? Harmony raises a person's Claude Enterprise or Cursor Enterprise spend limit after approval, through the AI Usage Limits agent, and the increase resets at the end of the period by default. The baseline limits stay in each provider. ### How is this different from each provider's admin console? A provider console shows its own tool. Harmony maps usage across tools to the same employee record that holds their department, devices and access, and ties requests for more usage to an approval. ### Does tracking Microsoft Copilot need extra setup? Yes. Copilot usage events come from the Microsoft Purview audit log, so audit logging must be turned on in your Microsoft 365 tenant.