Start Call

Exactly one Start Call node exists per agent, it has no incoming edges, and every call begins on it. It does three jobs in order: fetch anything the agent should already know, say the disclosure and the greeting, then run the first conversational turn.

Greeting and prompt are not the same field

This is the one thing worth getting right before anything else.

	Greeting	Prompt
What it is	Fixed text, spoken verbatim	Instructions to the model
Who writes the words	You	The model, guided by you
Varies per call	Only through {{variables}}	Yes, it responds to the caller
Costs a model call	No	Yes

A greeting is played and then the conversation begins. It is the right place for anything that must come out identically every time — a company name, a regulatory line, a fixed opening. The prompt handles everything after that.

Leave the greeting empty and the agent says nothing until the caller speaks. On an inbound call that is often correct; the caller rang you and will say hello. On an outbound call it is nearly always wrong — the person who answered hears silence and hangs up.

Fields

namestringdefault: Start Call

Short label shown on the canvas and in call logs. No runtime effect.

greeting_typetext | audiodefault: text

Whether the greeting is spoken by the voice model from text, or played from a pre-recorded audio file.

greetingstring

The text spoken at the start of the call. Supports {{template_variables}}. Empty means no greeting.

greeting_recording_idrecording

The audio file played instead, when greeting_type is audio.

recording_disclosure_enabledbooleandefault: on

Speak a short disclosure before the greeting. Unset follows the platform default, which is on. See Recording disclosure.

recording_disclosurestring

What the agent says. Blank uses the platform default wording.

promptstringrequired

The system prompt for the opening turn. Supports {{template_variables}} from pre-call fetch and the initial context.

allow_interruptbooleandefault: false

Whether the caller can talk over the agent mid-utterance. Leave it off for a node whose words are a disclosure that must be heard in full.

add_global_promptbooleandefault: true

Prepend the Global node’s prompt to this one at runtime.

delayed_startbooleandefault: false

Wait before speaking after the line is answered.

delayed_start_durationnumberdefault: 2.0

How long to wait, in seconds. Between 0.1 and 10.

extraction_enabledbooleandefault: false

Run a variable-extraction pass on this node’s exchange. See Context and variables.

tool_uuidsstring[]

Tools the agent may call during the opening turn.

document_uuidsstring[]

Knowledge Base documents available on this node.

pre_call_fetch_enabledbooleandefault: false

Look the caller up before the call starts. See Pre-call data fetch.

Recording disclosure

Every call on Decibyl is recorded, and three separate regimes expect the person on the other end to be told:

• India — under the DPDP Act a call recording is personal data collected from the person who answered, and telecom practice expects disclosure.
• Two-party-consent US states — twelve of them require consent from everyone on the call, not just the party recording.
• Your own customers’ expectations, which are usually stricter than either.

So it defaults to on, and stays on until somebody deliberately turns it off. The setting lives on the node rather than in organisation settings, which means it travels with the agent — versioned alongside everything else, and readable at call start without a database round-trip on the critical path.

Note

If you write your own disclosure wording, write it to be heard, not read. “This call may be recorded for quality and training purposes” takes two seconds and everyone recognises it. A paragraph of policy gets talked over.

Delayed start

An outbound call connects the moment the network says it did, which is often a fraction before the person has the phone at their ear. An agent that starts talking immediately loses its first few words.

Switching on Delayed Start waits — two seconds by default — before speaking. It costs you nothing on the model side, because nothing is generated during the wait.

Tip

Delayed start is for outbound. On inbound, the caller has been listening to a ring tone and is already waiting; a pause reads as a broken line.

Pre-call data fetch

With Pre-Call Data Fetch on, Decibyl POSTs to a URL of yours before the call begins, and merges the JSON it gets back into the call context as template variables. The request body carries the calling and called numbers, so your endpoint can look the person up.

The difference it makes:

Without

Agent:  Thanks for calling Sunrise Dental. Can I take your name?
Caller: Priya Menon.
Agent:  And do you have an appointment with us already?

With

Agent:  Hello Priya — I can see you're booked with Dr Rao on Thursday.
        Is that what you're calling about?

Two turns saved on every single call, which on a few thousand calls is real money as well as a better experience.

The full setup, including authentication and what happens when your endpoint is slow or down, is in Pre-call data fetch.

Rules

• Exactly one Start Call node per agent. The editor will not let you save a second, or none.
• No incoming edges. Nothing transitions to the start of a call. If you need to return to the opening behaviour later, use an agent node.
• At least one outgoing edge, unless the whole conversation is the greeting.

See also

Agent node

Every conversational step after this one.

Global node

Rules that apply on every turn, not just this one.

Pre-recorded audio

Playing a file instead of synthesising the greeting.

Pre-call data fetch

Knowing who is on the line before they say a word.