Audon vs. building it yourself.
Both halves of Audon can be built in-house, and some teams should. This page lists what the work involves, from what we had to build, so the estimate is honest on both sides.
Side by side. Facts about Audon from the product and its docs.
| Aspect | Audon | Built in-house |
|---|---|---|
| Provider APIs | Connectors for OpenAI, Anthropic, Claude Enterprise, xAI, Cursor, Devin and OpenTelemetry from Claude Code and Codex, each with its own key type, units and revision rules. | One integration per provider and per API change: amounts arrive as dollars, cents in decimal strings or ACUs. |
| Savings levers | A catalog of levers with provider facts that carry a source and a check date, sized in stage order so they add up. | Pricing facts to research and keep current per provider and model. |
| Proof | Savings measured from billing per unit of work, and verified by an approved experiment with at least 30 successes per window. | A measurement method to design and defend. |
| Agent actions | Agent Action Control: checks under a lock on the idempotency key and the resource, fresh provider context, duplicates within 72 hours, limits and approvals. | Rules inside each agent or tool, kept in step across agents. |
| Unclear outcomes | A timeout after sending is Unknown and is looked up at the provider before any retry. | Retry logic that must never send a refund twice. |
| Evidence | Append-only, hash-chained evidence per action that the database refuses to change. | An audit trail to design and protect. |
When building it yourself makes sense. Said plainly.
- You have a platform team with time to own provider integrations and keep pricing facts current.
- Your agents take one kind of action in one system, and that system already enforces idempotency and limits you trust.
Questions teams ask when they compare.
Can we start with Audon and build later?
Yes. Every capability is available over REST (OpenAPI) and MCP, the audit reports export as CSV or JSON, and the evidence for each action is readable through the API.
Does Audon sit in our request path?
Not for Agent Spend Optimization: it reads billing data only. For Agent Action Control, the actions you protect run through Audon (the SDK or the Action Gateway) so they can be checked before they run.
The products
- Agent Spend Optimization: Finds and proves savings on what AI agents, model APIs and coding tools cost, from read-only billing data.
- Agent Action Control: Checks every refund, credit, discount or subscription change an AI agent makes before it runs, verifies the outcome and keeps the evidence.
Last updated