What you get
By the end of this page you have a Finn that answers questions about your business correctly on a test call. You will not have a live phone number, an audience, or a deployment. Nothing you do here is reachable by a real caller until you create a deployment, so you can be careless with drafts.
The wizard has seven tabs: Identity, Behavior, Knowledge Base, Workflow, Tasks, Call Settings, Post-Call Analysis. This quickstart uses four of them and ignores the rest. Every field is editable later.
1. Open the wizard
Sidebar → Create New Finn. Creating and editing a Finn is a dashboard operation. There is no wizard endpoint.
Either start from scratch or click Browse Templates and pick one. A template pre-fills Identity, Behavior, and voice, which means your first test call has something to say. Starting from scratch is fine but you will write more before the agent is coherent.
There is a Save button in the top right of every tab. Saving does not publish, deploy, or dial anything. Use it between every step below.
2. Set identity and voice
On the Identity tab:
| Field | What to put | What breaks if you get it wrong |
|---|---|---|
| Name | Internal label, e.g. Inbound — Demo Booking | Nothing on calls. You lose track of your own Finns. |
| Voice | Pick from the library, listen to the sample first | Premium voices add latency. The library shows latency tiers. |
| Language | The language your callers actually speak | Drives speech recognition and transcripts, and constrains which voices are available |
| Time Zone | Where your callers live, not where you live | Only matters once you schedule deployments |
| Use Case | Closest category | Analytics grouping and a few defaults |
See agents-identity and agents-voice for the full field list.
3. Write the behavior fields
On the Behavior tab, three fields do most of the work.
Identity Text — who the agent is, in specifics.
You are Maya, a friendly receptionist at Acme Dental Clinic. The clinic is in San Francisco. You help patients book, reschedule, and cancel appointments.
Style and Guardrails — how it speaks and what it must never do. This is where most first attempts are too thin. A model with no stated limits will improvise into pricing quotes, medical advice, and commitments you cannot honor.
Speak in short, conversational sentences. Never quote prices. Defer pricing questions to "let me have someone call you back with that."
Response Guidelines — formatting and confirmation rules.
Always confirm the appointment date and time back to the caller before ending the call. Keep responses under two sentences.
Then set the Welcome Message. Dynamic lets the agent generate an opener. Defined plays an exact script, which is what you want for anything compliance-sensitive such as a recording notice. More in agents-prompting, agents-guardrails, and agents-welcome.
4. Give it something to be correct about
Prompting alone will not make the agent answer questions about your business. It will guess. Go to the Knowledge Base tab and add one source: an uploaded file (PDF, DOCX, plain text), pasted text, or a website URL to crawl.
One focused document beats a large vague one. Indexing takes a short time and shows a progress indicator. The agent searches across all sources during the call and pulls in relevant snippets. See agents-knowledge-base.
5. Test on a call
Click Test on Call and enter your own phone number. The Finn calls you. Test calls are free.
Ask a question that is answered in the document you just uploaded. That is the check that matters. If it answers from the document, your setup works end to end.
Test on Chat is a text simulation in the browser. Use it for fast prompt iteration, then confirm on a real call, because chat does not exercise speech recognition or interruption handling.
| Symptom | Likely cause |
|---|---|
| Call never arrives | Phone number missing country code |
| Agent sounds robotic | Voice choice. Try another from the library. |
| Agent ignores your document | Document too long or off-topic. Use a narrower file. |
| Agent invents a price or a promise | Guardrails too vague. Name the forbidden behavior explicitly. |
| Agent mishears a brand or product name | Add it to Custom Vocabulary on the Behavior tab. See agents-vocabulary. |
Iterate: change one field, save, test again. Changes apply to new calls only. In-flight calls finish under the old configuration.
Next
- agents-testing — test edge cases before real callers hit them
- agents-tasks — booking, webhooks, CRM writes
- post-call-analysis — define the structured fields extracted from every call
- quickstart-inbound or quickstart-outbound — put the agent on real calls