What you get
By the end of this page you have a Finn that answers questions about your business on a test call. It won't be reachable by real callers yet: a Finn takes or places live calls only after you deploy it, so you can experiment freely.
You can also create Finns through the API with api-finns. Call flows and knowledge bases are dashboard-only.
1. Start a new Finn
In the sidebar, click Create New Finn. You get three choices:
| Choice | What it does |
|---|---|
| Start from scratch | Opens the editor with empty fields. This page follows this path. |
| Browse our playbooks | Pick a ready-made Finn for your industry and adjust it in a three-step setup. |
| Create with AI copilot | Describe the agent you want and the copilot drafts it. |
The editor has four tabs: Call flow, Agent handbook, Call setup and Post-call analysis. This quickstart uses the first two.
2. Set identity and the welcome message
On the Call flow tab, under Identity:
| Field | What to put |
|---|---|
| Finn agent name | An internal label, such as Demo booking. Callers don't hear it. |
| Voice persona | Pick a voice and play the sample first. See agents-voice. |
| Default language | The language your callers speak. See agents-multilingual. |
Then fill in Welcome message, the first line the Finn speaks on every call:
Hi, this is Maya from Acme Dental. How can I help you today?
See agents-welcome.
3. Write the system prompt
System prompt holds the Finn's objective, flow and rules. Be specific about who it is, what it does and what it must not do:
You are Maya, a receptionist at Acme Dental in San Francisco. You help patients book, reschedule and cancel appointments. Keep answers to one or two sentences. Never quote prices; offer to have someone call back with pricing. Confirm the date and time back to the caller before ending the call.
A vague prompt produces vague calls, and a prompt with no limits lets the model improvise. See agents-prompting and agents-guardrails.
If the call needs fixed steps, such as checking availability before booking, build a call flow instead. See agents-tasks.
4. Give it something to be correct about
A prompt alone won't make the Finn answer questions about your business. It will guess. Open the Agent handbook tab and, under Knowledge base, add one source:
| Source | How |
|---|---|
| Files | Drop or browse for PDF, DOCX or TXT files. |
| Website | Enter a public URL and click Import. |
| Text | Paste text and click Add as document. |
One focused document works better than a large, mixed one. See agents-knowledge-base.
Click Create Finn. The Finn opens on its builder page as a Draft.
5. Test on a call
On the builder page, open Test → Phone call, enter Your phone number and click Place test call. The Finn calls you. A test call is a real call, so your wallet needs credits.
Ask a question that the document you added answers. If the Finn answers from it, your setup works end to end.
Test → Simulation chat runs a text conversation in the browser. Use it to iterate on the prompt quickly, then confirm on a phone call, because chat doesn't exercise speech recognition or interruptions.
| Symptom | Likely cause |
|---|---|
| The call never arrives | The phone number is missing its country code, or the wallet has no credits. |
| The voice sounds wrong for the job | Try another voice in Voice persona. |
| The Finn ignores your document | The document is too long or off-topic. Use a narrower file. |
| The Finn invents a price or a promise | The prompt doesn't forbid it. Name the forbidden behavior explicitly. |
To change something, click Edit on the builder page, change one thing, save, and test again.
Next
- agents-testing: test edge cases before real callers find them.
- agents-tasks: calendar booking, transfers and API calls during a call.
- post-call-analysis: structured answers extracted from every call.
- quickstart-inbound or quickstart-outbound: put the Finn on real calls.