9 AI Voice Agent Prompt Mistakes (From 2,053 Real Calls)
The mistakes that break AI calls are almost all in the script and the contact list, not the voice. We read 2,053 AI call transcripts from 325 Indian businesses. First-draft agents read a placeholder like "[Company Name]" out loud in nearly 1 in 10 calls, and those agents lost twice as many callers in the first two replies. About 1 in 9 of their calls ended in "I'll check and get back to you". Here are the nine mistakes we saw most, what each sounds like on a real call, and how to fix it.
Ratnam Sachan, Co-founder · Vyora AI
Published · Last updated
From 2,053 real calls
9.6%
Of first-draft agent calls read a placeholder out loud
39% vs 18%
Callers lost in 2 replies: agents with leaks vs without
11.7%
Of first-draft agent calls ended in "I'll check and update you"
What is an AI voice agent prompt?
The prompt is the set of instructions that tells a voice agent who it is, what the call is for, what it knows and how to respond. On most platforms, Vyora included, it is split into an intro message the agent speaks word for word, plus objective, guidelines, task and FAQ sections it uses to hold the rest of the conversation. Variables in {{double_braces}} are filled from your contact list on each call. Unlike a chatbot prompt, every word may be heard by a customer, in real time, with no chance to scroll back.
Leaving placeholders in the script
On the call"Main Priya bol rahi hoon [Company Name] se." / "Namaste {{}} ji."
Why it hurtsA voice agent reads what it is given. Square-bracket placeholders, empty or misspelled variables and single braces all get spoken. It happened in nearly 1 in 10 calls from first-draft agents (9.6%), and those agents lost 39% of callers within two replies, against 18% for agents without leaks. A leak is usually a sign nobody read the script before launch, so treat it as the first thing to fix, not the only one.
FixSearch the script for "[" and "{" before every launch. Use real variables in {{double_braces}} that exist in the agent's variable list, or type the actual company name.
Writing "pick one" options into the intro
On the call"Hello Sir/Ma'am, main bol raha/rahi hoon. Good morning/afternoon/evening."
Why it hurtsFine in a template for a human telecaller, who picks one. The intro message is spoken word for word, so the agent says every option, slashes included.
FixAddress people by name ({{callee_name}} ji). Choose one gender form for the agent. Use "Namaste" or "Hello", which work at any hour.
Uploading a dirty contact list
On the call"Am I speaking with nan?" / a 10-digit phone number read out as the person's name
Why it hurtsEmpty spreadsheet cells often export as "nan", "null" or "N/A", and those words land in the call. Lists with numbers in the name column make the agent read the digits as a name.
FixBefore uploading, filter the name column for blanks, "nan", "null" and digits. Either fix those rows or remove them. A missing name costs less than a wrong one.
Not saying who is calling, or why
On the call"Am I speaking with Rahul?" … "Haan, kaun?"
Why it hurtsA name check with no company or reason makes the caller do the work. With a name-only opener, 11% of callers had to ask "kahan se bol rahe ho?" or similar, against 6% when the company was named.
FixName, company and reason in the first sentence, then one question. Scripts for eight use cases are in our opening lines guide.
Stating an amount without saying what it is for
On the call"Aapki 36,000 rupaye ki payment pending hai. Kis date tak denge?" … "Kis cheez ka?"
Why it hurtsPayment reminders without context sound like fraud. Callers answered with a question, or flatly denied owing anything, instead of giving a date.
FixLead with the invoice, product or month ({{invoice_no}}, {{product}}), then the amount, then the date question.
Giving the agent no facts to answer with
On the call"Kitne ka package hai?" … "Main check karke aapko update karungi."
Why it hurtsIn 11.7% of calls from first-draft agents (5.2% across all accounts), the agent promised to check and call back instead of answering. These were predictable questions about prices, slots, order status and package lists. The agent had a goal but no facts.
FixWrite down the 10 questions customers actually ask and put the answers in the FAQ field or a knowledge base. For anything else, offer a transfer or a callback with a specific time.
Asking for several things at once
On the call"Aapka naam, city, state aur preferred language kya hai?"
Why it hurtsCallers answer part of it, the agent asks again, and the call loops. Voice platform guidance agrees: collect one piece of information per turn.
FixOne question per turn. Confirm, then ask the next one.
Mixing up the agent's language
On the callA Marathi greeting on a Hindi agent, a Bengali opener followed by Hindi questions, Telugu typed in Roman letters
Why it hurtsThe voice is tuned for one language and the script is written in another, so pronunciation and turn-taking both suffer. Several callers had to ask to switch language mid-call.
FixWrite the script in the agent's primary language, in its native script or plain Roman, but not both in one sentence. Add a secondary language if the list is mixed.
No plan for silence, wrong numbers or voicemail
On the call"Are you still there?" two seconds into a pause / the full pitch delivered to a call-screening assistant
Why it hurtsHappy-path scripts break at the edges. We saw agents repeating the same line back to back, prompting too quickly during a pause, and pitching to voicemail and call-screening bots.
FixAdd explicit lines for "wrong number" (apologise, end), "busy" (ask for a time), and silence. Set a silence timeout that gives people time to think. Vyora defaults to a 10-second warning.
For mistake 4, start with our opening line scripts for Indian use cases. On asking one thing per turn, Deepgram's voice agent prompting guide gives the same advice.
Pre-launch checklist
- 01
Search the script for "[" and "{". Every {{variable}} must exist in the agent's variable list.
- 02
Read the intro out loud. Remove anything with a slash.
- 03
Check the first sentence has a name, a company and a reason.
- 04
Filter the contact list's name column for blanks, "nan", "null" and phone numbers.
- 05
Add answers to the 10 most common customer questions.
- 06
Add lines for wrong number, busy and call-back requests.
- 07
Call yourself in the agent's language and listen to the whole call.
How Vyora catches these before a call goes out
After reading these transcripts we added checks for the first three mistakes:
- ✓
Script check. The agent builder and editor flag unfilled [placeholders], empty or unknown {{variables}}, single-brace text, and slash options like "Sir/Ma'am" or "raha/rahi" in the intro.
- ✓
Contact list check. When you upload a CSV for a campaign, Vyora warns you about mapped columns with empty values, "nan" or "null", or phone numbers in the name column.
- ✓
Warnings, not blocks. You can still save and launch. Sometimes a bracket is intentional.
The rest still need a person to read transcripts. Our AI voice agent analytics guide covers what to look for, and how to set up AI calling walks through building the agent.
Frequently asked questions
What is the most common AI voice agent prompt mistake?
Leaving template placeholders in the script, such as "[Company Name]", an empty "{{}}" or "Sir/Ma'am". In our review of 2,053 AI call transcripts from India, first-draft agents read one aloud in 9.6% of calls, and those agents lost 39% of callers within two replies, against 18% for agents without leaks.
Why does my AI agent say "nan" instead of the customer's name?
The name cell for that contact was empty in your spreadsheet, and the export wrote "nan" (or "null") into the CSV. The agent reads it as the name. Fix or remove those rows before uploading. Vyora's campaign upload now warns you when a mapped column has empty or "nan" values.
How long should an AI voice agent prompt be?
Long enough to cover the goal, the facts and the edge cases, with the spoken parts kept short. The intro should be one sentence. Each agent reply should be one or two sentences and ask one question. Put facts such as prices and policies in an FAQ or knowledge base rather than the main instructions.
How do I stop my AI agent saying "I'll check and get back to you"?
Give it the answers. That line means the agent was asked something its script and knowledge base don't cover. Review transcripts for the questions that trigger it, add those answers, and give the agent a clear next step for anything else, such as a transfer or a callback at a set time.
Does Vyora check scripts for these mistakes?
Yes. The agent builder and editor flag unfilled [placeholders], empty or unknown {{variables}}, single-brace text and slash options like "Sir/Ma'am" in the intro. The campaign upload warns about empty, "nan" or phone-number values in mapped columns. They are warnings, not blocks, so you decide what to fix.
Build an agent with the checks built in
50 free credits, no card. Describe your agent, fix what the editor flags, call yourself.