Test calls
The test call never arrives
| Cause | Check |
|---|---|
| Number missing country code | Use E.164. US: +15551234567. India: +919876543210. The dashboard shows a "tested format" preview before it triggers the call. |
| Trial credit exhausted | Settings → Plan and billing → Usage. |
| Carrier rejected the call | Some carriers block calls from new numbers, more often internationally. Try a different from-number. |
| Handset spam filter | Apple and Google filters drop some calls silently with no failure signal on your side. Try a different from-number. |
The call connects but the Finn says nothing
It isn't the welcome message: the Finn always speaks it first. See agents-welcome.
The usual cause is the voice: it is not assigned, or is assigned to a voice that requires a region match (agents-voice). If you changed AI service configuration, reset it to default.
Building a Finn
Save is grayed out
A required field is empty (usually Name, Voice, or Language). If not, your browser session may have expired: copy any long text you've written, then refresh and log back in. Call flow canvas issues don't block saving; see workflow-editor.
The Finn ignores your script
| Cause | Fix |
|---|---|
| Identity field too vague | "You are Maya, a friendly receptionist" beats "You are an AI assistant." |
| Guardrails buried among other instructions | Give them their own block in the system prompt. |
| Fields contradict each other | Audit Identity, Style, Welcome, and Workflow together. |
| Knowledge base holds contradictory source material | Find the conflicting documents. Documents can't be removed from the dashboard, so add a corrected document and tell the Finn in the prompt which source wins. |
| You need a fixed order and are using free-form prompting | Use a workflow instead. |
The Finn asks the same question twice
The caller was misheard, or on a call flow the answer matched none of the step's transition conditions, so the Finn asked again. Read the transcript, then rephrase the question or rewrite the conditions in the caller's terms. If the answer was captured but not remembered, add to the prompt: "Once a question is answered, do not ask it again." Also check the call flow for a transition looping back to the same step.
Pacing is wrong
| Symptom | Fix |
|---|---|
| Too talky | Response Guidelines: "Keep responses under two sentences unless asked for more detail." |
| Too brief | Response Guidelines: "Give complete answers with examples when relevant." |
| Long silences | Lower Reminder message frequency under Call setup. See agents-conversation. |
| Calls cut off around eight minutes | When Max call duration is not set, calls stop at 500 seconds. Raising it also raises the per-call cost ceiling. |
Knowledge base
| Symptom | Cause |
|---|---|
| Answers missing from uploaded documents | The upload or indexing failed. A failed upload shows Failed with the error in the knowledge base upload list. Upload the file again. |
| Scanned PDF returns nothing | Image-only PDFs are not OCR'd. Convert to text before uploading. |
| Retrieval misses a section of a long document | Split into smaller focused files. |
| Answers are outdated | Upload the current version, or import the page again. Knowledge base documents cannot be removed from the dashboard, so the old copy stays attached; tell the Finn in the prompt which source is current. |
When answers look wrong, do not guess. Read the transcript for a real call in the Data Extractor Sidebar, then compare what the agent said with the source document. See agents-knowledge-base.
Deployments
Launched, but no calls go out
| Cause | Check |
|---|---|
| Audience is empty | Open the audience and confirm rows exist. |
| Deployment paused | Live Deployments → open the deployment → status. |
See outbound-campaigns and phone-numbers.
Everyone hangs up in the first seconds
A from-number carriers label SPAM LIKELY collapses answer rates. See spam-labeling. Beyond that, generic openings get hung up on. Name yourself, name a company the caller recognizes, and state a reason they would expect. If the list is cold, low pickup-and-stay is a list problem, not a configuration problem.
Analytics
Zero calls shown when calls happened. Usually the deployment dropdown is on a different deployment. Check the full log in call-logs.
Post-call analysis fields blank. Fields only run against calls that completed after the field was added. Existing calls stay blank permanently. Vague questions also extract poorly: "Did the caller express interest?" works, "Interest level?" does not. See post-call-analysis.
Sentiment looks wrong. Sentiment is an indicator, not a verdict. If it misclassifies consistently, add a custom post-call field with a specific question and use that instead.
Phone numbers
| Symptom | Cause |
|---|---|
| Renting a number fails with "Compliance approval required" | Your workspace's latest compliance application is not Approved. Check its status in Settings → Compliance. |
| Your own number not connecting | See byo-numbers. |
Still stuck
Email [email protected] with the deployment ID, the call ID if there is one, and what you expected versus what happened. See support.