First call
This is the shortest path from an empty account to hearing your own agent hold a conversation, and then reading the transcript of it. It takes about fifteen minutes, and most of that is spent writing the prompt.
You will do it in two passes. The first pass never touches telephony at all — you test the agent in the browser, where iteration takes seconds. Only once it sounds right do you attach a phone number, because that is the step whose details depend on where your organization is.
An account with its region already chosen (sign up covers that), and — for the last two steps — a phone number you can attach. If you don't have one yet, stop after step 3: the browser test call is genuinely useful on its own.
Create a workflow
From an empty workspace, create a workflow. It does not start blank: a new workflow arrives with two nodes already wired together.
Start Call is the single entry point, and every workflow must have exactly one. It cannot be deleted. End Call closes the conversation. Everything you build from here sits between those two, and the line joining them is an edge — the rule that decides when the conversation moves on.
Give the workflow a name you will recognise in a list of thirty. "Test 2" is a decision you regret in a fortnight.
Write the prompt
Open the Start Call node. Two fields matter on the first pass.
The greeting is the literal first thing the caller hears. It is deliberately separate from the prompt, because the opening line is the one moment where you want exact words rather than a model's interpretation of them.
The prompt is everything else: what the agent is trying to accomplish at this step, and how it should behave while doing it. There is no separate "personality" setting elsewhere in the product — tone lives in the prompt, and so does judgment.
Keep the first one small. A prompt covering four goals produces an agent that drifts between them, and you cannot tell which of the four is failing. One clear objective, tested, then add the next as its own Agent Node.
Writing one enormous prompt that describes the whole call. It reads sensibly and it behaves badly. Several small nodes joined by edges give you a conversation whose shape you can see on the canvas — and, when a call goes wrong, one specific node to fix.
Test it in the browser
Before touching telephony, use the builder's test-call panel to place a call over WebRTC, straight from the browser.
This is not a simulation. It runs the exact same pipeline a phone call would — the same speech recognition, the same model turn, the same speech synthesis, the same turn-taking — with the phone network swapped out for your microphone and speakers. Latency will feel slightly better than a real call, because there is no carrier leg, but everything you are testing at this stage behaves identically.
Iterate here. Change a line of the prompt, call again, listen. Thirty seconds a cycle, no number required, no per-minute cost.
Attach a phone number
Once it sounds right in the browser, connect it to a real number. This is the step that differs by region, because number sourcing is a compliance question before it is a technical one.
Import a number you already own from your existing carrier. Woise does not sell or provision US numbers directly — you bring the number, and Woise connects it to the workflow. You will need your carrier's credentials; see BYO telephony for exactly which fields each carrier asks for.
Once your organization has cleared identity verification, Woise provisions a managed +91 number on your behalf and attaches it to the workflow. The KYC step gates this, so start it early if you have not — see India.
Full detail on both paths lives in Numbers.
Place a real call and read the transcript
With a number attached, call it — or, for an outbound agent, have it call you.
Every call is a workflow run, and the distinction matters more than it sounds. The workflow is the definition; the run is one execution of it. Editing the workflow afterwards does not rewrite the runs that came before, which is what makes a transcript worth trusting as a record.
Each run carries four things:
| What | Where it comes from |
|---|---|
| Transcript | Both sides of the conversation, turn by turn |
| Recording | The call audio, stored once the call ends |
| Cost | Metered from the call's actual duration and services used |
| Extracted variables | Whatever your nodes were configured to collect |
The recording and transcript appear alongside the workflow's other runs once the call completes — Recordings covers where to find them and what they hold.
What to do next
You have a working agent with one step. The three most useful directions from here:
| If you want to… | Read |
|---|---|
| Give the agent more than one step | Nodes — the ten types and when each applies |
| Let it look things up or take action | Tools and Knowledge base |
| Call a list of people rather than wait | Campaigns |
If something in the walkthrough did not behave as described, Concepts is the page that defines every term used above precisely.
Common questions
What is Woise?
Woise is a managed platform for voice AI agents. You design an agent in a drag-and-drop workflow editor, connect a phone number, and it handles real inbound and outbound calls over the phone network or in the browser. The infrastructure, carrier setup and scaling are run for you.
Do I need a phone number to start?
You bring a number you already own and import it — that is the primary path today, and it means you keep your carrier relationship and your existing number. Agents can also take and place calls in the browser over WebRTC with no number at all.
How does billing work?
You are billed for the call minutes your agents actually use, metered as credits against your account. There is no seat licence and no per-agent fee, so an agent that takes no calls costs nothing to keep.
Can an agent connect to my CRM or calendar?
Yes. Agents can read from and write to your CRM, calendar and support tools through the integrations marketplace — you authorise each connection once with OAuth, and no credentials are stored in your workflow.
Do I need engineers to build an agent?
The workflow editor is visual, so building and changing an agent needs no code. There is also a REST API, a Python SDK and a TypeScript SDK if you would rather drive the platform programmatically, and an MCP server so coding agents can build workflows for you.
What happens to call recordings and transcripts?
Recordings and transcripts belong to your workspace and are available through the dashboard and the API. Raw transcripts, recordings and phone numbers are never handed to third-party analytics or error tracking without redaction.
What infrastructure do I have to run?
None. Woise runs the speech models, the language model turn, the call recording pipeline and the telephony legs for you, and scales them with your call volume. There is nothing to provision, patch or keep running.
Still stuck? Talk to our team.