What a Finn is
A Finn is the AI voice agent itself: a named configuration made up of identity, voice, behavior text, knowledge, optional workflow, tasks, call settings, and post-call analysis fields. A Finn does not place or receive calls on its own. It becomes reachable only when you attach it to a deployment. See concepts for how Finns, deployments, and calls relate.
Create one from the dashboard sidebar under Create New Finn. The wizard is organized into tabs: Identity, Behavior, Knowledge Base, Workflow, Tasks, Call Settings, Post-Call Analysis. Nothing is live until you create a deployment, and test calls do not require a deployment.
The configuration blocks
| Block | What it holds | Detail page |
|---|---|---|
| Identity | Name, voice, language, language group, time zone, use case, icon | agents-voice, agents-multilingual |
| Behavior | Identity text, style and guardrails, response guidelines, welcome message, custom vocabulary | agents-prompting |
| Knowledge Base | Uploaded files, pasted text, website imports | agents-knowledge-base |
| Workflow | Optional node graph enforcing conversation order and branching | tools-overview |
| Tasks | Structured side-effects: calendar booking, webhook, CRM write | tools-builtin, tools-custom |
| Call Settings | Idle handling, duration ceiling, voicemail behavior, handoff number | below |
| Post-Call Analysis | Fields extracted from the transcript after each call | post-call-analysis |
| AI services | LLM, TTS, STT selection | below |
Identity fields
The Name is an internal label. Callers never hear it. Use something you can find later, such as Inbound — Demo Booking.
Language drives speech recognition, transcript output, and which voices are selectable. Language Group tunes voice and STT for a region; most accounts default to global. Time Zone governs scheduled deployments and calling-hour windows, so set it to where your callers live, not where you live. Getting this wrong is a common cause of campaigns dialing at the wrong hour.
Use Case affects analytics grouping and a few default behaviors. Finn Icon is cosmetic.
Behavior fields
Identity text says who the agent is. Style and guardrails say how it speaks and what it must never do. Response guidelines cover formatting and confirmation rules. Vague identity text produces vague behavior, and missing guardrails let the model improvise into pricing or advice you did not authorize.
The welcome message has two modes:
| Mode | Behavior | Use for |
|---|---|---|
| Dynamic | The Finn generates an opener from identity plus context | Inbound |
| Defined | Exact script, played verbatim | Outbound, and any compliance-sensitive opening such as a recording notice |
Custom vocabulary is a list of words the recognizer should expect: brand names, SKUs, drug names, proper nouns. Without it, transcription of domain terms degrades and post-call extraction inherits the errors.
Call settings
| Field | What it controls | Sensible default |
|---|---|---|
max_idle_duration | Silence allowed before the Finn re-prompts | 10s |
idle_reminder | What the Finn says on re-prompt | "Are you still there?" |
limit_call_duration | Hard ceiling on call length | 5 mins |
voicemail_detection | What to do when voicemail answers | End call if voicemail detected |
handoff_number | Number to transfer to on escalation | optional |
AI service configuration
LLM, TTS, and STT services are selectable. Defaults work for most use cases, and TTS is tied to your voice selection. Change these only with a specific reason, since a mismatched TTS service can invalidate a chosen voice.
What is editable, and when it takes effect
Every field above is editable after creation. Edit from the dashboard: View All Finn → click the Finn → Edit.
| Change | Applies to |
|---|---|
| Any Finn config field | New calls only |
| Workflow version | New deployments only |
| In-flight call | Finishes under the config it started with |
Active deployments keep running against the workflow version they launched with. Editing a workflow does not retroactively change a running campaign, so a fix you make mid-campaign will not appear until you create a new deployment. Plan changes around deployment boundaries rather than assuming a save propagates.
Versioning and rollback
The platform keeps a history of Finn configurations, reachable at View All Finn → [Finn] → History. You can restore an earlier version if a change degrades performance. Workflows version separately on every save, with their own History → Restore, and a restored version becomes the current version.
Workflows also export and import as JSON from Edit Workflow → ⋯ menu, which is how you copy a flow between Finns or keep it in version control.
Testing before deployment
From the Finn page, Test on Call dials your phone and Test on Chat runs a text simulation in the browser. Chat iterates faster; call is the only test that exercises the real audio path, including voice latency and recognition of your custom vocabulary. Test hostile, confused, and off-topic callers before you deploy, because free-form mode has no structural guarantee that the conversation stays on track.
Next
- agents-prompting for writing identity and guardrail text
- agents-variables for passing per-call data in
- quickstart-inbound or quickstart-outbound to put a Finn on real calls
- api-finns for the Finn resource over HTTP