Architecture
Five admission checks decide whether a call is answered, and the answer waits at most a second for warm-up. PSTN and WebRTC build the same pipeline.
Four things matter before you build on Vertical AI. An admission sequence decides whether a call is answered. Both transports share one pipeline. Some surfaces trigger the post-call chain and some do not. And the version boundary splits settings a version owns from settings that edit live. For which region each part of that pipeline runs in, read where a Vertical AI call runs.
What happens when a call arrives
The caller's network delivers the call to our carrier, which posts a signature-verified webhook to the voice runtime. Everything below happens before the caller's phone stops ringing.
Number resolution
The runtime resolves the dialled number to an agent. If the lookup fails or the number is not assigned, the call is refused cleanly and the caller can redial.
Admission checks
Five checks decide whether the call is answered. They run concurrently with warm-up, so on the healthy path they cost no perceptible time.
| Check | On failure | On error |
|---|---|---|
| Blocked caller | The call is refused, or the caller hears your message and the call ends | Fails open |
| Billing | Call refused before answer, no charge | Fails open |
| Answering hours | Outside your hours, the call gets your closed handling: a message or a forward | Fails open |
| Duplicate call | Second arrival for the same call refused | Fails open |
| Platform capacity | Caller forwarded to your overflow number | Fails open |
Blocked callers and answering hours apply only when the agent sets them. The duplicate guard is off by default. Billing and capacity are always active.
Order matters. The results are read in the order above, so a workspace that can't be billed is refused rather than forwarded. A pipeline slot is claimed while the checks run, and a call that any check turns away gives its slot back.
Warm-up during ringback
While the checks run, the runtime loads the agent's settings, warms the prompt cache and pre-renders the greeting audio. The answer waits at most a second for the warm-up; after that the call is answered anyway.
The caller hears ringback rather than dead air. A greeting rendered in advance is the difference between an agent that starts talking and one that starts thinking.
Answer and media
The runtime answers, the media stream opens, the greeting plays, and the conversation begins. Latency targets are on Reliability.
Two transports, one pipeline
| Transport | Used by |
|---|---|
| PSTN | Inbound calls, outbound campaigns, onboarding, call-me test |
| WebRTC | In-app test call, share-link voice call, website voice widget |
Both build the same pipeline. A browser test call exercises the same agent and conversational machinery as a phone call. What it does not exercise is the carrier leg: line quality, codec compression and speakerphone echo only show up on a real phone call.
After the call
Three passes run in order:
- Deterministic metrics. Model-free measurements from the call record.
- Judge metrics. A language model grades only the metrics the agent has enabled. None enabled means no model calls and no charge.
- Analysis. Sentiment, outcome, caller type, urgency, summary and extracted facts. Rolls the call into the caller profile when memory is on.
Analysis reads metric results, so it runs after the judge pass. Each pass is isolated, so one failing does not stop the next.
On the first analysis only, the platform fans out to
alerts, post-call follow-up, the agent's post-call
functions and the call_log.completed webhook. Re-analysis does not fire them
again. Turning call analysis off also stops all of them.
Which surfaces run the post-call chain
| Surface | Transport | Post-call chain |
|---|---|---|
| Real inbound or outbound call | PSTN | Yes |
| In-app test call, share link, website widget | WebRTC | Yes |
| Voice eval, phone mode | PSTN | Yes, simulated |
| Voice eval, bridge mode | Direct | Yes, simulated |
| Text eval | None | Yes, simulated |
| Chat: test panel, share link, website widget | None | No |
A phone-mode eval dials the agent's own number, so its agent leg is an ordinary inbound call. Every eval runs the chain marked simulated: metrics and analysis are written, and each post-call function records what it would have sent instead of sending it. So a test can't send a real message to a stranger.
Evals send nothing after the call
An eval sends no webhook, raises no alert and runs no follow-up. If your integration depends on those, check it on a real call.
Related
Core concepts
An agent is one language model, one system prompt and a flat list of tools. The workspace owns the phone numbers, and the published version answers calls.
The version boundary
Which settings a published version owns and restores on a rollback, and which settings edit live on the next call with no publish step.