The quick start gets an agent running; this page covers doing it properly for production.
Before you deploy
- Pick a narrow use case. One document type, one queue, one workflow. Broad mandates produce noisy escalations.
- Decide the human fallback. Who reviews escalations? Where do they land (email, ticket queue, Slack)?
- Set a success metric. Typical: % of operations completed autonomously without correction.
Deployment checklist
- Connect the source integration and verify the agent can read sample inputs.
- Deploy the agent with a high confidence threshold (0.9+). Expect heavy escalation at first — that is the point.
- Run a one-week shadow period: the agent processes real work, humans verify everything.
- Review the escalation log together with the team that owns the workflow.
- Lower the threshold gradually (0.05 at a time) as autonomous accuracy holds.
Common first-week issues
| Symptom | Likely cause | Fix |
|---|---|---|
| Everything escalates | Threshold too high for input quality | Improve input data or lower threshold slowly |
Agent stuck in deploying |
No operation has been sent yet | Send a first operation via the API or integration |
| Operations rejected with 402 | Monthly quota exhausted | Upgrade plan or wait for the next billing month |