Route, approve and account for every AI request.
Know who uses which AI, for what, and at what cost. Marshal is a private control plane for enterprise AI: approved access, scoped keys, quotas, and reports your CIO, FinOps and auditors can rely on, deployed in your data center or VPC.
AI use grows from the bottom up.
Personal accounts, shared keys, tools nobody registered. When finance asks what it costs, or audit asks who approved it, there's no single answer.
Access with approval chains
- Staff request a model and a quota.
- The request goes through your approval chain: as many levels as your organization needs, in your order.
- Once approved, a scoped key is issued with quota and expiry.
- Every step is recorded.
Usage intelligence
- Usage by project, department, provider and model: tokens, cost, latency.
- Work versus personal or unclassified use.
- Reports for management, FinOps, security and audit.
Policy that fails closed
- Requests outside the allowed provider, model, region or quota are denied, with a clear reason.
- No silent fallback to something unapproved.
Private by default
- Runs in your environment.
- Provider keys stay in your secret store; applications never see them.
- Prompts and responses are not stored unless your policy says so.
From request to key, on the record.
The same object that was requested and approved is the one that gets metered and reported. So "who approved this" and "what was used under it" have one answer.
- Access requested: model + monthly quotasubmitted
- Team managerapproved
- IT securityapproved
- FinOpsapproved
- Scoped key issued, with quota and expiryrecorded
- Later request uses a model outside the approvaldenied · reason logged
Example. Levels and order are set by your organization.
Answers for each person who asks.
Reports use metadata and aggregates, so they don't need copies of sensitive content.
Bottom-up use, top-down view.
Employees use Legwork to get more from the AI they already have. Marshal shows IT what it's used for and what it saves.
Fits how your organization approves things
Approval chains with as many levels as your process needs, in your order. Reports shaped to your departments and projects. Our implementation team configures it with you, without custom code forks.
Designed for standard interfaces
Built around OpenAI- and Anthropic-compatible APIs, MCP and OpenTelemetry, so it governs the tools you already run instead of replacing them.
Start with a pilot in your environment.
A time-boxed pilot with acceptance criteria agreed up front. We run a small number at a time.