Hard facts
True regardless of which vendor you pick.
Every one of these is a property of physics, telephony, human conversation or procurement — not of any product. They are the things buyers and builders keep discovering three weeks too late.
Latency
You have roughly 800 milliseconds, total
From the moment a caller stops speaking, a reply needs to begin within about 800 ms or the conversation feels broken.
Why it is true
Human turn-taking gaps average around 200 ms. People tolerate more from a phone line, but past roughly a second they assume the line dropped, repeat themselves, or talk over the reply. That budget covers everything: endpointing, transcription, the model, speech synthesis and network transit both ways.
What to do about it
Ask for p95 latency measured end to end from caller silence to first audio — not model inference time, which is the smaller half. A vendor quoting median rather than p95 is quoting the calls that went well.
Interruption handling matters more than voice quality
A bot that cannot be interrupted feels worse than one with a mediocre voice, however good the samples sounded.
Why it is true
People interrupt constantly — to correct, to hurry, to answer early. A system that keeps talking over them signals immediately that nobody is listening. Handling it means cancelling in-flight synthesis and the in-flight model call within a few hundred milliseconds, which is an architectural property, not a setting.
What to do about it
On any demo, interrupt it. Interrupt it mid-sentence, twice. That single test separates real products from pipelines assembled last month.
Audio
Telephony audio is 8 kHz, and every benchmark you have read is not
Phone calls are narrowband, compressed and lossy. Published accuracy figures are almost always measured on clean 16 kHz or better.
Why it is true
The PSTN band is roughly 300–3400 Hz, and codecs compress further. Consonants that distinguish similar words live in frequencies the phone network discards. Models trained and benchmarked on studio audio lose accuracy the moment they meet a real call.
What to do about it
Never accept a transcription or intent number that was not measured on telephony audio. Record ten of your own real calls and have any serious candidate run them.
The transcript is the ceiling
No language model can recover meaning that speech recognition got wrong. Errors upstream are permanent.
Why it is true
Everything downstream reasons over text. If the recogniser heard the wrong account number or the wrong product name, the model will answer confidently about the wrong thing — and sound completely fluent doing it.
What to do about it
Spend evaluation effort on recognition accuracy for your vocabulary — names, places, product terms — before spending it on prompt quality.
Language
Supporting two languages is not the same as handling a switch between them
A system can support Hindi and English separately and still fail on a caller who moves between them inside one sentence.
Why it is true
Most stacks pick a language per session or per utterance. A mid-sentence switch breaks that assumption: the recogniser is locked to one language model, and the words in the other come out as noise. Handling it needs multilingual recognition and a synthesiser that can pronounce both without changing voice.
What to do about it
Ask specifically about mid-sentence code-switching, and ask for accuracy measured on code-switched utterances alone — not the blended figure, which hides the failure.
Economics
The per-minute rate is not what you will pay
All-in cost per conversation is routinely two to four times the headline per-minute number.
Why it is true
The quoted rate usually covers one layer. The bill also carries telephony, speech recognition, model tokens, synthesis, platform fees, recording storage and overage — and bills for the whole call duration including silence, not just the talking.
What to do about it
Ask for total spend divided by conversations handled, for a real month of a real customer. Anyone who cannot produce that number has not been asked before.
Outbound economics are decided before the bot says anything
Connect rate and answering-machine detection move outbound unit economics far more than model quality does.
Why it is true
If only a fraction of dials reach a human, you are paying for every attempt while earning outcomes on a slice of them. Voicemail misclassified as a person burns a full call's cost for nothing.
What to do about it
Ask for cost per qualified outcome, not cost per minute, and ask how answering machines are detected and what happens next.
Compliance
Outbound calling in India carries obligations inbound does not
DLT registration, DND scrubbing and consent are the caller's responsibility, and they apply before the first dial.
Why it is true
Indian telecom regulation puts obligations on commercial communication — registration of sender and template, scrubbing against do-not-disturb registries, and calling-window rules. A voice AI vendor may help, but the exposure sits with the business placing the calls.
What to do about it
Settle who owns DLT and DND scrubbing in writing before the pilot, and budget for it. This is not a detail you can add afterwards. Take your own legal advice — this is not it.
Recording a call creates obligations in most places
Storing call audio and transcripts brings notice, retention and deletion duties, and they vary by jurisdiction.
Why it is true
Recordings are personal data. India's DPDP regime and comparable laws elsewhere require a lawful basis, a stated purpose, a retention limit and a route to deletion. Transcripts are no safer than audio — often less, because they are searchable.
What to do about it
Decide retention and redaction before go-live, not after the first access request. Where recordings are stored, and who can reach them, belongs in the vendor conversation.
Process
If a vendor needs your call transcripts, that is your timeline, not theirs
Collection, redaction and consent review of real call data routinely takes longer than the entire build.
Why it is true
Transcripts contain personal data. Getting them out of a contact centre, redacted and cleared for use crosses legal, security and operations — three queues you do not control. Vendors quote their build time, which starts after all of that.
What to do about it
Ask what training data is required before the bot performs, and who does the redaction. Then add that to the vendor's quoted timeline rather than assuming it runs in parallel.
A percentage without a sample size is not a number
'92% accuracy' is unfalsifiable. '92% over 4,182 calls in March' can be checked.
Why it is true
Without a denominator you cannot tell a real result from a good week, ten cherry-picked calls, or a benchmark run on someone else's data. The denominator is also where metrics get quietly gamed — dropping abandoned calls from a containment rate can move it twenty points.
What to do about it
For every number you are given, ask three things: over how many, in what window, and on which traffic.
Build versus buy is a question about turn-taking, not models
Every model in this stack is an API call. What you are really deciding is whether to build the conversation loop.
Why it is true
Recognition, generation and synthesis are commodities you can swap in an afternoon. Voice activity detection, endpointing, barge-in cancellation, state recovery and graceful degradation are months of engineering, and they are what separates a working demo from a system that survives ten thousand real calls.
What to do about it
If you build, budget for orchestration as the main project and the models as configuration. If you buy, evaluate the orchestration — interrupt the demo — rather than the model names on the pricing page.
A pilot with no outcome definition cannot end
Without an agreed success metric captured per call from day one, pilots get extended rather than decided.
Why it is true
At the end of an eight-week pilot somebody asks whether it worked. If nobody labelled outcomes, the honest answer is that nobody knows — and the safe institutional response is another eight weeks.
What to do about it
Agree the number and the threshold before kickoff, and make sure it is captured automatically per call. One metric, one denominator, written down.
Nothing on this page is legal advice, and the compliance items in particular are a prompt to ask your own counsel rather than a substitute for one.