Healthcare · Financial services · Public sector
Understand AI agent costs without sending sensitive content.
Track the economics of agent workflows with model names, token counts, timing, and internal IDs. Your model and tool requests stay on their existing path. Ganivra receives usage metadata separately.
A clear data boundary
Send the facts needed to calculate cost.
Operational metadata
- Provider and model name
- Input, output, and cached token counts
- Latency, status, and error category
- Execution, step, and workflow IDs
- Pseudonymous customer or tenant ID
- MCP server and tool identifiers, when used
For margin, add customer pricing or revenue amounts and transaction references.
Sensitive content
- Prompts and AI responses
- PHI and patient records
- Customer names, emails, and contact details
- Account credentials, card numbers, and bank details
- Source documents and application transcripts
- MCP arguments, results, and authentication tokens
Use neutral labels and internal IDs. Keep sensitive content out of custom attributes too.
How the architecture works
Your agent keeps its existing connections.
Ganivra does not sit between the request and response.
Send operational facts after the call to calculate and attribute cost.
No proxy inspection
Keep using your existing provider clients and tool connections. Instrumentation reports usage metadata without requiring prompt or response capture.
Fail-open SDK delivery
The Python SDK sends telemetry through a bounded background queue. Delivery failures do not intentionally block the model response; telemetry may be dropped rather than hold up your app.
For a direct REST integration, send asynchronously from your backend and handle telemetry failures separately from the user request.
Commercial value
Explain spend by workflow, not by sensitive payload.
Healthcare AI workflows
Attribute model and tool costs to workflows such as scheduling or document processing using neutral workflow IDs. Patient records and conversation transcripts are not needed to see which workflows consume the budget.
Financial research and advisory
Link model calls and MCP tool steps under one report execution. See tracked cost-to-serve using an internal account ID, without sending account credentials or the report content.
Public-sector applications
Compare tracked usage across applications and services with internal identifiers. Give budget owners a cost breakdown without collecting citizen conversations in telemetry.
Customer and contract economics
Connect supplied revenue or configured customer pricing to tracked AI costs. Identify expensive workflows and review pricing, limits, or model choices using the recorded evidence.
Keep your data review focused
Prompts and responses do not need to leave your existing model and tool path for Ganivra. Usage metadata is still sent to and stored by Ganivra: review its fields and hosting requirements against your organization's residency policies before production use.
Start small
Connect one workflow with metadata only.
Send a call, verify its cost, and add customer pricing when you are ready to see margin.