Go-live checklist

Work through this before you send customers to an agent. Each item is one thing to check, with the page that shows how.

Agent

• Pin a published version. Edit the draft, publish, and point the phone number at that version so later edits do not reach callers by surprise. See Phone numbers and Agent versions.
• Set a voice. There is no default voice, so an agent without a voice_id cannot speak. See Voices.
• Set a first message. Leave it empty and the agent waits for the caller to speak first. See Prompt agents.
• Set the call limits you want. Check max_call_duration_ms and end_call_after_silence_ms, so a stuck call ends by itself. See Prompt agents.
• Run your test suite against the draft, then publish. Persona cases vary, so judge a change by several runs. A suite can also block publishing until it passes. See Testing agents.

Telephony

• Route the number. Confirm it has an inbound agent. A call to a number with no agent is refused. See Phone numbers.
• Check your routing webhook, if you use one. It has 1.5 seconds in total; if it is slow or down, the number's own agent answers. See Inbound routing.
• Call your transfer destination with a real number. Confirm it rings, the right person answers, and the call connects. See Voicemail and transfer.
• Decide what happens at a machine. Choose whether the agent hangs up or leaves a message, and test it with real numbers, since detection is a best guess. See Voicemail and transfer.

Webhooks

• Use a public HTTPS endpoint. Private addresses are refused and redirects are not followed. Use Send test to confirm it. See Set up a webhook endpoint.
• Verify the signature on the raw body. Reject anything that fails. See Verify signatures.
• Answer 2xx quickly. Store the event, reply, and do the slow work afterwards. See Delivery and retries.
• Be idempotent on the delivery id. Delivery is at least once, and X-Dialgem-Delivery is the same on every retry. See Delivery and retries.
• Store the signing secret in your secrets manager. It is shown in full only once. See Set up a webhook endpoint.

API keys

• Scope each key. Give it only the resources it needs. A key without scopes has full access. See API key scopes.
• Limit by address and expiry. Use allowed_ips where your servers have fixed addresses, and set an expiry. See API key scopes.
• Keep keys in a secrets store, on your server. Never in a browser or mobile app. See Get your API keys.
• Check the key at startup. GET /v1/api-key-info makes a wrong, expired or under-scoped key fail at deploy time rather than on a customer call. See Get your API keys.
• Write down a rotation plan. Create the new key, deploy it, wait until the old one is idle, then revoke. See Get your API keys.

Billing

• Fund the account for expected traffic. At zero balance new calls are refused and campaigns pause; calls in progress carry on. See Billing and pricing.
• Consider auto-recharge. It tops up from a saved card when the balance falls below a level you set. See Billing and pricing.
• Set a credit-balance alert and a spend alert. See Alerts.

Privacy and data

• Decide on recording and transcripts. opt_out_call_recording and opt_out_transcript_storage turn storage off per agent. See Security.
• Turn on redaction if you need it. pii_redaction_enabled redacts personal data from stored transcripts. See Prompt agents.
• Set a retention window. Choose an organization default, with an optional per-agent override. See Data retention and erasure.

Limits

• Web calls: the default is 10 calls at once and 60 calls created per minute per organization. Throttle your own start-call endpoint. See Tokens and allowed websites.
• Campaigns: max_concurrency (1 to 50) caps calls in progress for a campaign. See Campaigns.
• Everything else: do not assume a number for any limit that is not documented. Ask support before launching something large.

Monitoring

• Watch analytics. Success, failures, cost and latency, filterable by agent. See Analytics.
• Set alerts. Failure rate, latency, spend, credit and stalled campaigns, sent by email, Slack or signed webhook. See Alerts.
• Know the call list. Filter by status, agent, end reason and metadata. See Calls.
• Know how to investigate a failed call. Open the record, read disconnection_reason and the transcript, then check the webhook delivery log. Send the call id to support if it is still unclear. See Troubleshooting.

Next steps