Dark Truth about Voice AI · Brendan's AI Community

Dark Truth about Voice AI

If you think, "May I grab your phone number" and "May I grab your email" in the prompt can capture the phone number and emails in real life with 95% accuracy, you are in living in dreamworld.

The real world voice ai is bloody hard. I had no idea, when I focused only on voice ai when everyone else was building 100 other automation to sell.

However, there was something about the voice ai as technology that made me quit my high paying cybersecurity job at Amazon, ignore all other automation and solely focus on the nitty gritty details of prompting, STT, TTS and retell/vapi settings.

Three months and working with niches like immigration, insurance, beauty, entertainment etc , listening to 1000+ real calls with different accent, tone, pronunciation, I can tell you that building a voice ai agent to either increase the sales or save the cost on operations/support takes 4-6 weeks of intense efforts of listening to the calls, tweaking the agent and making it 95% accurate at least.

If you are in voice AI space, would love to connect, share and learn from your exp.

https://www.linkedin.com/in/connectwithamitgupta/

Sharing the phone number capture prompt after 10+ iteration which is proving to be 97% accurate so far. Feel free to tweak and use it, if you need.

***********Prompt starts***********

#Phone Number Collection Protocol:

## Format Knowledge

**Australian Numbers (10 digits total):**

- Mobile: 04XX XXX XXX

- Landline: 02/03/07/08 XXXX XXXX

- International: +61 (drop leading 0)

**Indian Numbers (10 digits total):**

- Mobile: 6/7/8/9XXXXXXXXX

- Landline: Area code (2-4 digits) + subscriber number = 10 total

- International: +91 + 10 digits

**UAE Numbers:**

- Mobile: 05X XXXX XXX (9 digits total)

- Landline: Area code (1-2 digits) + 7 digits = 8-9 digits total

- International: +971 + local number

## Capture Process

1. **Request**: "Can I please grab your phone number for a call back?"

2. **Listen for**:

- 10 digits (Australia/India)

- 8-9 digits (UAE)

- 11-14 digits with country code

3. **Common Mishears**:

- "Fifteen/Fifty" → "One-five or five-zero?"

- "Thirteen/Thirty" → "One-three or three-zero?"

- Zero vs "oh" → Both acceptable

- Repeated digits: Double [digit] = repeat digit twice, Triple [digit]= repeat digit three times

- Unclear digit → "Could you repeat the number after [last clear digit]?"

4. **Confirm**: "Let me repeat: [number in groups]. Correct?"

- Australian: Zero-four-one-two, three-four-five, six-seven-eight (example grouping)

- Indian: Nine-one-two-three, four-five-six, seven-eight-nine-zero (example grouping)

- UAE: Zero-five-zero, one-two-three, four-five-six-seven (example grouping)

5. **If unclear after first attempt**: "Phone numbers on the phone can be tricky! Let's try once more - could you repeat your number digit by digit?"

6. **If still unclear after second attempt**: "I'm clearly having one of those days with numbers! Would you mind spelling each digit out for me slowly? I promise I'll get it right this time!"

## Validation

- Must be 10 digits (AU/IN), 8-9 digits (UAE), or include country code

- Australian mobile starts with 04, landline with 02/03/07/08

- Indian mobile starts with 6/7/8/9

- UAE mobile starts with 05X, landline with area code

- Maximum 2 attempts - use humor to keep it light if unclear

******Prompt Ends********

Dark Truth about Voice AI

6 comments