Voice AI demos are all impressive, because a demo is a clean line, a cooperative speaker and a rehearsed script. These are the questions that separate products, and what a good answer sounds like.
On the call itself
1 · Can I interrupt it, and what happens? Ask them to demo it live, on your phone, talking over the agent after one second. It should stop mid-word. Watch for the agent finishing its sentence after going quiet — that means buffered audio was never flushed.
2 · What is the response latency, measured from what? Any number is meaningless without the starting point. Ask whether it includes end-of-speech detection and the silence window — most quoted figures do not, and those are the larger half. Under about two seconds end to end is good.
3 · How does it handle code-mixed speech? If your callers switch between English and an Indian language mid-sentence — and they will — test exactly that. Many systems handle each language well separately and fail at the join.
4 · What happens on voicemail? It should detect and end. If it holds a conversation with an answering machine, you are paying for it, and those transcripts pollute your summaries.
On the output
5 · Show me a transcript of a bad call. Anyone can show a good one. Ask for one where the person was in a car with poor signal.
6 · Can I define my own scoring rubric? Generic sentiment scores are not a shortlist. You need your own columns, your own criteria, and a reason attached to every score.
7 · Where do recordings live and how long are they kept? Your own bucket is the right answer. A defined retention policy is the minimum.
On the economics
8 · What is the all-in per-minute cost? Telephony, speech recognition, the model and the voice should be one number. If you are assembling a bill from four vendors, that complexity is yours to manage forever.
9 · Is there a cheaper mode for one-way calls? Paying conversation rates to play a reminder is the most common overspend in this category.
10 · What happens to unanswered calls? They should bill nothing.
On the risk
11 · What stops it inventing an answer? "The model is good" is not an answer. Look for a knowledge base with an explicit instruction to defer when a fact is not in it.
12 · What does it do badly? The most revealing question on the list. A vendor who cannot name a limitation either has not run this at scale or is not being straight with you.
Our own answer: judgement-heavy conversations, anything where one wrong statement is expensive, very poor audio, cold lists, and volumes under about a hundred calls a month.
The test that beats all twelve
Ask for a trial where you write the prompt and call your own phone. Not a scripted demo — your script, your language, your number, and you being deliberately difficult. Ten minutes of that tells you more than an hour of slides.