Voice Agent Greeting Design: The First 5 Seconds (2026)

TL;DR — Callers decide in 5 seconds whether to stay or hang up. A great greeting names the brand, sets expectations, discloses AI status under proposed FCC 2026 rules, and gives the caller the floor inside one breath. CallSphere's vertical-tuned openers cut early hang-ups by 38% versus a generic "How can I help you?".

The UX challenge

Cathy Pearl, head of conversation design outreach at Google, calls the opening "the contract": you tell the user who you are, what you can do, and how to get out. Get any of those wrong and abandon-rate spikes inside 8 seconds. Production traces from CallSphere show that 52% of all hang-ups happen before the first user turn — the greeting itself is the leak.

Three failure modes dominate:

• Latency — greeting starts > 800 ms after pickup; caller thinks the line is dead.
• Identity — caller cannot tell if they reached the right business or a robot.
• Goal ambiguity — the agent asks an open-ended "How can I help?" with no hint of what it can actually do.

Patterns that work

The Google Conversation Design System's cooperative principle says give as much information as is needed — and no more. A 5-second greeting fits four moves:

1. Brand anchor — "Thanks for calling Acme Dental."
2. Identity — "This is Aria, an AI assistant" (FCC NPRM Sept 2024 + 2026 disclosure proposals).
3. Capability frame — "I can book, reschedule, or take a message."
4. Floor pass — "What can I do for you?"

flowchart TD
  PICKUP[Caller picks up] --> AUDIO{First audio < 800ms?}
  AUDIO -->|No| LEAK[High abandon - fix TTS warm-up]
  AUDIO -->|Yes| BRAND[Brand anchor: 'Thanks for calling X']
  BRAND --> ID[AI disclosure: 'This is Aria, an AI assistant']
  ID --> CAP[Capability frame: 3 verbs max]
  CAP --> FLOOR[Floor pass: short open question]
  FLOOR --> LISTEN[Open mic, VAD armed]
  LISTEN --> SUCCESS[First user turn captured]

CallSphere implementation

CallSphere ships 6 vertical-tuned greeting templates across its 37 specialized agents and 90+ tools, all backed by the 115+ DB tables that store transcripts and outcomes:

• Healthcare (14 tools) — "Thanks for calling [Practice]. This is Aria, an AI assistant. I can book, reschedule, or transfer you to the front desk — what can I do for you?"
• Salon greet — "Hi, [Studio] front desk, this is Mia. I can book a service or check availability — go ahead."
• OneRoof Aria triage — "OneRoof property line, Aria here. I can take a maintenance request or hand you to leasing — which one?"

Explore a live demo and compare current plans to find the right fit for your business.

Build steps

1. Measure first audio latency at the SIP edge — anything > 800 ms means the TTS engine is cold-starting; pre-buffer the greeting WAV.
2. Write four moves, ≤ 18 words total — brand, identity, capability, floor pass.
3. A/B against an open "How can I help?" baseline; track 8-second hang-up rate, not just CSAT.
4. Localize for time of day ("Good morning" before noon) — UXmatters notes this lifts perceived warmth ~12%.
5. Cache the brand anchor as a static audio asset; only the dynamic name field needs TTS.

Eval rubric

FAQ

Q: Should I say "AI" or "virtual assistant"? The proposed FCC 2026 rule wants "clear and unambiguous" disclosure — "AI assistant" is safest. "Virtual assistant" is contested.

Q: Does a long brand jingle help? No. UXmatters and Cathy Pearl both flag earcons over 1.5 sec as a hang-up driver. Keep it under 600 ms.

Q: Can I skip the disclosure for inbound calls? The FCC proposal applies to AI-generated voices regardless of inbound/outbound. Disclose every time.

Q: What if the caller interrupts the greeting? Honor the barge-in. If they say "human," route immediately. See our demo flow.

Sources

• Cathy Pearl — Designing Voice User Interfaces (O'Reilly)
• Google Conversation Design — Cooperative Principle
• FCC — TCPA & AI-Generated Voices
• UXmatters — Designing Voice User Interfaces (Pearl review)
• Fuselab — Voice UI Design Guide 2026