Nodes

Pick the right step for each moment in a call, and know what each one can be configured to do.

A node is one step in a conversation. Ten types exist and the list is fixed — you create new instances, never new types. That constraint is deliberate: every type maps to something the voice runtime knows how to execute on a live call, so a workflow can never describe behaviour the engine cannot perform.

They fall into three groups:

• Conversational — talk to the caller and decide where to go next. Start Call, Agent Node, Global Node, End Call.
• Input — collect keypad digits. Keypad Capture, Keypad Menu.
• Side-effect — reach outside the call, to your systems or to Woise's own analysis, without the caller hearing anything. Webhook, Integration Action, API Trigger, QA Analysis.

Conversational nodes

Start Call

Every workflow has exactly one, and it is the only entry point. It cannot be deleted, and no edge can point into it. A workflow without one will not run.

It owns two things nothing else does.

The greeting is what the caller hears first, and it is separate from the prompt. Greeting Type chooses between speaking text (Greeting Text) and playing an audio file you have uploaded (Greeting Recording). Use a recording when the first impression matters more than flexibility — a recorded human voice still beats synthesis on the opening line.

Pre-Call Data Fetch is the other. Enable it and Woise calls Endpoint URL before the conversation begins, then makes the response available to the agent, so the greeting can already know who is calling. Attach a stored credential through Authentication rather than putting a key in the URL.

Delayed Start holds the agent silent for Delay Duration (seconds) after the line opens. This is not cosmetic. On some carrier paths audio connects slightly before the caller has the handset to their ear, and a greeting spoken into that gap is simply missed.

Agent Node

The workhorse. Everything between the greeting and the goodbye is Agent Nodes, each with one focused Prompt describing what to accomplish at that step.

Three attachments change what a node can do:

• Tools lets the agent call out mid-conversation — check availability, look up an order.
• Knowledge Base Documents grounds its answers in your content rather than the model's training data.
• Variable extraction turns speech into structured data. Enable it, then define Variables to Extract, each with a Variable Name, a Type and an Extraction Hint telling the model what to look for. Extracted values persist for the rest of the run.

Allow Interruption decides whether the caller can talk over the agent. Leave it on for anything conversational. Turn it off only for content that must be heard in full, such as a legal disclosure.

Resist writing one enormous Agent Node. A node's prompt is the model's entire instruction for that turn, and a prompt covering four goals produces an agent that drifts between them. Several small nodes joined by edges give you a conversation whose shape you can see on the canvas — and, when a call goes wrong, a specific node to fix.

Global Node

Not a step. A Global Node holds a Global Prompt that applies across the whole workflow — tone, persona, standing rules like never quote a price. No edge touches it.

Individual nodes opt in through Add Global Prompt, which is why that setting lives on Start Call, Agent Node and End Call rather than on the Global Node itself. A node with it switched off sits deliberately outside the global instruction, which is occasionally what you want for a scripted passage.

End Call

Closes the conversation. Like Agent Node it takes a Prompt and supports variable extraction, which matters more here than it looks: a final extraction pass is where you capture the outcome — did they book, did they decline, what was the reason — and that is what lands in your reports.

A workflow may have more than one End Call. Different endings usually deserve different closing lines and different extracted outcomes.

Input nodes

Speech recognition is strong on language and unreliable on long digit strings. When a caller reads out a sixteen-digit account number, use the keypad.

Keypad Capture

Collects a sequence of digits and stores it under Store As, the key later nodes and webhooks read.

Three settings decide whether it feels reliable:

• Expected Length tells Woise the sequence is complete without waiting.
• Termination Key (typically #) lets the caller signal they are done when the length varies.
• Inter-digit Timeout (seconds) governs how long to wait between presses before assuming they have finished.

Max Retries caps the re-ask loop. Set Expected Length whenever the length is genuinely fixed — it is the difference between a capture that completes instantly and one that always waits out the timeout.

Keypad Menu

The classic press 1 for sales branch. Menu Prompt reads the options, and each option becomes an edge on the canvas, so the branching is visible rather than buried in configuration.

Invalid / No-input Reprompt is what the caller hears after an unrecognised press or silence, and No-input Timeout (seconds) sets how long silence has to last first. Write the reprompt as a real second attempt. Repeating the menu verbatim is the most common reason callers give up.

Side-effect nodes

Webhook, Integration Action and QA Analysis all run after the workflow completes, not during the conversation. Put them on the canvas for what should happen once the call is over. For something the agent needs while the caller waits, attach a tool to an Agent Node instead.

Webhook

Sends an HTTP request to your own endpoint once the run finishes. Set HTTP Method and Endpoint URL, attach a stored credential through Authentication, and add any Custom Headers as Header Name / Header Value pairs.

Payload Template is the field that matters: it decides what you send. The template can reference workflow_run_id, initial_context, gathered_context, annotations and the call's metadata, so a webhook can hand your system everything the agent gathered.

Integration Action

The same idea without the plumbing. A Webhook talks to an endpoint you host; an Integration Action talks to a system Woise already knows how to authorise against — see the integrations catalogue. It runs once after the workflow completes, durably, with retry and dead-letter handling. Prefer it when the target is listed, and keep Webhook for everything else.

API Trigger

Not a step in the conversation but an entry point into it. Trigger Path defines a path your systems can call to reach this point in the workflow from outside.

QA Analysis

Grades completed calls. System Prompt defines what you are grading for, and the rest controls what gets graded: Sample Rate (%) for the share of calls scored, Minimum Call Duration (seconds) to skip hang-ups, and Include Voicemail Calls for whether answering machines count.

By default it uses the workflow's own model (Use Workflow's LLM). Turn that off and QA LLM Provider, QA Model and credentials appear, so grading can run on a different — usually stronger — model than the one holding the conversation. That separation is worth having: the model judging a call does not have to be fast, only accurate.

How nodes connect

Nodes are joined by edges, and an edge is a rule about when the conversation moves on. It might fire when the model judges the caller's goal met, when a specific key is pressed, or when a condition on an extracted variable holds.

This is the part worth internalising: the graph is the agent. There is no configuration behind it and no hidden state between nodes beyond the variables you explicitly extract. What you can see on the canvas is what runs on the call.

Next steps