Steps Overview

Step Types

Voxworks supports three step types:

Type	Purpose	Interactive	Awaits Response	Caller can talk over it
Voice	Speak to user and listen for response	Yes	Yes	Yes by default. Set per step with Interruptions.
Tool	Call external services	No	No	Not while the tool runs. Its “thinking” line follows Interruptions.
Flow	Navigate to a sub-flow	No	Depends	Not applicable. The sub-flow’s own steps decide.

Common Step Properties

All steps share these properties:

Property	Description	Default
`id`	Unique identifier for the step	Required
`type`	Step type (voice, tool, flow)	Required

Voice and Tool steps also have a script property for spoken output. Flow steps do not have a script, they simply navigate directly to the sub-flow.

Step-Specific Settings

Each step’s settings live on the step’s menu on the Script tab. A step with a non-default setting shows a badge for it.

Menu item	Options	Default for a new step	Available on	What it does
Step Type	Voice, Tool, Flow	Voice	All steps	Changes what kind of step it is.
Interruptions	Normal, Uninterruptible, Uninterruptible turn, Allow skip, plus a Verbatim threshold	Normal, threshold 80	All steps	Whether the caller can talk over the step. See Interruptions.
Fast Response	Auto, Force, Disable, Variable	Auto	All steps	Whether the step waits for a reply or moves on. See Fast Response.
Silence Tolerance	Hurried, Patient, Extremely Patient, Infinite	Hurried	All steps	How long to wait before prompting a silent caller. See Silence Tolerance.
Response Eagerness	Keen, Normal, Patient	Keen	All steps	How quickly to respond after the caller stops speaking. See Response Eagerness.
Thinking Effort	Fast, Normal, Deep	Fast	All steps	How much the AI reasons before replying. See Thinking Effort.
Models	The models your team can use	The included default	Voice steps in Instruct mode	Which AI model writes the step’s reply. See Models below.
Classifiers	Script, Flow and Step classifiers	None	Voice steps, and tool steps other than composite tools	Background data extraction that runs alongside the step. See Classifiers below.
Key Terms	Edit…, Overfit	None	Voice steps	Words the speech recogniser should listen for while this step runs. See Key Terms.
Prompt Mode	Verbatim, Instruct	Instruct	All steps	Whether the line is spoken word for word or used as an instruction. See Writing Scripts.
Live Handoff	On or off	Follows the script’s setting on voice and flow steps; off on tool steps unless turned on for that step	All steps	Whether the caller can ask for a person during this step. See Live Handoff.
Intent Pathways	The script’s pathways	Per pathway	All steps	Which Intent Pathways can trigger during this step.

In the stored script, Response Eagerness is saved as high / medium / low (Keen / Normal / Patient), Thinking Effort as low / medium / high (Fast / Normal / Deep), and Silence Tolerance as low / medium / high / infinite. You’ll only see these values in exported script files.

Models

The Models menu chooses which AI model writes a voice step’s reply. The tooltip reads: “Which model writes this step’s reply. The default is included with your plan; the rest run on your team’s own provider key, so a group appears once that integration is connected. Thinking Effort applies to the GPT-OSS models — every other model has its settings fixed by us.”

Models are grouped by provider. Included comes first and holds the default. The others (SCX, Anthropic, OpenAI and Google AI) appear once your team has connected that provider under Integrations, using your own API key.
Some models show a speed tag (fast or medium) as a guide to how quickly they reply.
The menu appears on voice steps in Instruct mode. Verbatim lines are spoken word for word, so no model is involved.
If a step is pinned to a model whose provider is disconnected or that is no longer offered, the step shows an amber Model unavailable badge. The step runs on the default model until you fix it. The badge is a warning only and doesn’t stop you publishing.
Thinking Effort only applies to the default GPT-OSS models. Other models run with fixed settings.

The list of available models can change over time. Leave steps on the default unless you have a reason to choose another model.

Classifiers

Classifiers are background AI extractions that run at the same time as a step and write what they find into variables. They never change where the call goes next. The tooltip reads: “Background LLM extractions that run alongside this step, capturing data into variables without changing where the call goes next. Script and flow classifiers are shared — editing one changes it everywhere it is enabled; the tick only turns it on or off for this step.” Open the step’s menu and choose Classifiers. The menu has three sections, each with Add Classifier:

Section	Scope
Script	Defined once for the whole script. It can be turned on for any step in any flow.
Flow	Defined once for the current flow. It can be turned on for any step in that flow.
Step	Belongs to this step only.

A Script or Flow classifier is shared: you write it once and tick it on for each step that should run it (the tick shows “Turn on for this step” / “Turn off for this step”). Editing a shared classifier changes it everywhere it is turned on, and the editor says so: “This classifier is shared. Changes apply to every step it is enabled on.” Its Label field sets how it is listed in the step menu. Shared classifiers are also listed under Settings → Shared Classifiers, where you can rename or delete them and choose which flows or steps run them. The menu shows a count of the classifiers active on the step. There is no limit on how many a step can run.

Only the LLM Data and LLM Custom tools can be used as a classifier.
Composite tool steps can’t carry classifiers.
On a voice step the classifiers run on the caller’s answer, alongside the assistant working out its reply. They get about two seconds; one that takes longer is dropped for that turn and its result is recorded as false, so a slow classifier adds at most a fraction of a second before the reply.
On a tool step they finish before the step moves on.
Whatever they extract lands in the variables you map them to, ready for later steps to read.

Typical uses: classifying what the caller just said, capturing a detail mentioned in passing, or scoring the call — all without adding a step or a pause to the conversation.

Voice Steps

Voice steps are the primary interactive component. They:

Speak a message to the user
Wait for the user’s response
Evaluate conditions to determine the next step

Voice Step: "Would you like to schedule a demo?"
   Condition: "User agrees" → Booking step
   Condition: "No" → Thank user and say goodbye step
   Condition: "Otherwise" → Loop back to current step

It’s recommended to use “Otherwise” to loop back to the current step, allowing the assistant to ensure step conditions have been met before continuing. Learn more about Voice Steps →

Tool Steps

Tool steps call external services. They:

Execute integrations (SMS, calendar, webhooks)
Return results that can be used in conditions
Can map outputs to variables

Tool Step: Check calendar availability
   Condition: true (slots available) → Speak options step
   Condition: false (no slots) → Apologize and offer alternative

Learn more about Tool Steps →

Flow Steps

Flow steps navigate to sub-flows. They:

Invoke another flow as a module
Variables are automatically available in the sub-flow
Return success or failure to the parent flow

Learn more about Flow Steps →

Step Execution Order

Steps execute in sequence unless conditions redirect:

Enter step — Step settings are applied (eagerness, silence tolerance, etc.)
Execute — Step-specific action runs
Evaluate — Conditions are checked (for voice/tool steps)
Navigate — Move to the next step based on conditions

Conditions

Most steps include conditions that determine navigation:

Step Type	Condition Basis
Voice	AI analysis of user response
Tool	Boolean result of tool execution
Flow	Result returned from sub-flow

Where a condition can send the call

Each condition’s Next step… dropdown offers every step in the flow, plus four outcomes:

Outcome	What happens
End Flow (Success)	Ends the current flow with a success result. In a sub-flow, control returns to the parent step.
End Flow (Failure)	Ends the current flow with a failure result. In a sub-flow, control returns to the parent step, which can route on the failure. In the main flow, both End Flow outcomes end the call.
End Call	The assistant finishes what it is saying and the call ends — but it leaves a short window in which the caller can speak and keep the call alive. Use it for a normal goodbye.
End Call (Strict)	The assistant finishes the sentence it is on and hangs up. The caller cannot talk it out of it. Use it when the call must end — abuse, a wrong number, or a compliance stop.

End Call and End Call (Strict) are available on any condition, including otherwise and the routing table on a tool step. Both record the end reason AI Hangup; the call’s System notes show which of the two fired.

Best Practices

Keep steps focused — Each step should do one thing
Use clear scripts — Write naturally, as you would speak
Plan your conditions — Consider all possible user responses
Set appropriate dynamics — Adjust eagerness, thinking effort, interruptions and fast response for each step
Test step transitions — Verify conditions work as expected

Next Steps

Voice Steps — Design interactive conversations
Tool Steps — Integrate external services
Flow Steps — Navigate between flows

Get Started & Build

Calls, Contacts & Batches

Numbers & Routing

Tools, Automations & Integrations

Account, Billing & Security

Step Types

Common Step Properties

Step-Specific Settings

Models

Classifiers

Voice Steps

Tool Steps

Flow Steps

Step Execution Order

Conditions

Where a condition can send the call

Best Practices

Next Steps

​Step Types

​Common Step Properties

​Step-Specific Settings

​Models

​Classifiers

​Voice Steps

​Tool Steps

​Flow Steps

​Step Execution Order

​Conditions

​Where a condition can send the call

​Best Practices

​Next Steps

Step Types

Common Step Properties

Step-Specific Settings

Models

Classifiers

Voice Steps

Tool Steps

Flow Steps

Step Execution Order

Conditions

Where a condition can send the call

Best Practices

Next Steps