Notes on running conversations at scale
Opinionated, specific, and mostly about the parts that go wrong. Written by the people building Convarza AI.
Latest
What actually drives your WhatsApp bill
Meta stopped charging per conversation in July 2025. It bills per delivered template now, and the category you pick moves the number by an order of magnitude.
ReadWhat an AI voice agent actually does on a call
Not a recording, not a menu, not a chatbot reading aloud. Here is what happens between the ring and the hang-up, and which parts you control.
ReadAI calling
7An AI voice agent is not an IVR with a nicer voice
Press 1 for sales is a menu. An AI voice agent is a conversation that can end in a booked appointment. The difference shows up in your abandon rate.
ReadWhat an AI voice agent actually costs to run
Three meters run at once and vendors bundle them differently, which is why per-minute comparisons rarely compare the same thing. How to price a ninety-second call.
ReadWhat an AI receptionist should answer after hours — and what it should not
The calls you miss between 7pm and 9am have a floor of zero, which makes them the safest place to start. The trick is knowing where to stop.
ReadLatency is the whole experience in voice AI
A one-second pause feels broken on a phone call and fine in chat. Where the delay comes from, what is fixable, and what you trade away to fix it.
ReadBooking an appointment by voice: what has to be true first
An agent that offers a slot it cannot actually hold is worse than one that takes a message. Availability, conflicts and least-notice come before the voice.
ReadSpeech recognition and Indian English
Code-switching mid-sentence, place names no model has seen, and numbers spoken in lakhs. What breaks, and what to test before you buy.
ReadAn AI receptionist for a clinic, specifically
Front desks lose calls during consultations and after close. What an agent can safely handle in a clinic, and the clinical line it must never cross.
ReadMessage templates that get approved, with examples
Six templates you can adapt today — order updates, reminders, callbacks, re-engagement — with the category each should be filed under and why.
ReadWhat a WhatsApp shared inbox actually needs to do
Most tools give you a list of chats. The ones teams keep using give you assignment, an AI that knows when to stop, and the contact record beside the thread.
ReadWhatsApp Business API in India: what you are actually buying
Everyone resells the same API. What differs is the inbox on top, who owns the number, and what happens when a template gets rejected at 9pm.
ReadWhatsApp messaging limits and quality rating, explained
Your number has a tier and a quality score. Both move on their own, and the way most businesses discover them is by hitting one.
ReadWhy your WhatsApp template keeps getting rejected
Six reasons, in the order you should check them. Most rejections are one of the first three, and none of them require talking to support.
ReadMoving your WhatsApp number to another provider
It is a migration, not a settings change. What you need to own before you start, and the one question that decides whether it is even possible.
ReadSales
4Five minutes: the number that decides who gets the customer
Buyers overwhelmingly go with whoever answers first, not whoever is best. Most companies answer in hours. Here is how to close that gap without hiring.
ReadHow to qualify a lead on a phone call, and hand the script to an agent
Four questions, asked in an order that survives a real conversation. Then how to give the same script to an AI agent without it sounding like a form.
ReadThe enquiry that went quiet is not lost yet
Someone asked a question, you answered, and nothing. What to send, when, and how many times before it becomes harassment.
ReadHow many times should you follow up?
Most teams stop after two attempts. Most answers come later than that. Where the honest line is, and how to stop before it becomes harassment.
ReadOperations
4Appointment reminders that actually reduce no-shows
One reminder is worth sending. Three is nagging. The timing, the channel and the one word that makes people reply instead of ignoring it.
ReadThe five call metrics worth a dashboard
Most call dashboards measure what is easy to count. Five numbers that change decisions, and three popular ones that mostly do not.
ReadYour missed calls are a bigger leak than your website
You measure form conversion to two decimal places and have no idea how many people rang and gave up. Here is how to find out, cheaply.
ReadFirst contact resolution, measured honestly
Almost every FCR number is flattering, because it counts tickets closed rather than problems solved. A definition that survives contact with reality.
ReadAI agents
4Stopping an AI agent inventing things
A model that does not know reaches for the nearest plausible answer. Five controls that turn that into an honest refusal instead.
ReadHow to write a prompt for an agent that talks to customers
Not prompt engineering. Job description writing — scope, tone, refusals, and the three sentences that prevent most embarrassing transcripts.
ReadHandover: the moment your AI agent should stop talking
Most disappointing deployments fail at the seam, not the answer. Name the conditions in advance and the handover stops being a failure.
ReadA knowledge base your AI agent can actually use
Your help centre was written for people who are already reading it. An agent needs the same facts arranged very differently.
ReadCRM
2Call outcome and lead stage are different things
One describes an attempt. The other describes a person. Conflating them is why your pipeline reports never reconcile.
ReadChoosing a CRM when your team lives on the phone
Most CRMs are built around a form and treat the call as an attachment. Seven questions that separate the ones that were not.
ReadCompliance
2Consent, DND and calling hours: the checks before you dial at volume
An autodialer will happily break rules you have not read. What to confirm with your carrier and your counsel before the first campaign goes out.
ReadWhat your AI agent does with customer data
Your customers' conversations pass through a model someone else runs. Nine questions to get answered in writing before they do.
Read