What setting up AI calling actually involves
Three separate layers, and they do not take the same amount of time.
The telecom layer
A number, DLT registration, and DND scrubbing against the national registry. Entirely regulatory, mostly out of your hands once submitted, and the usual reason a launch date moves.
The agent layer
What the agent says, in which languages, how it handles being interrupted, and what it does when it does not know. Fast to build, slow to get right.
The integration layer
Where the output goes — CRM fields, a sheet, a webhook, a WhatsApp message, a live transfer to your team. Quick, provided the decision was made before testing rather than after.
Delays come from the first. Disappointing results come from the second. Most setup guides published by platforms cover only the third, because it is the part their product controls.
The seven steps, in order
This is the sequence for a first campaign. Steps two and three run in parallel with four and five, which is why the order matters less than starting the regulatory ones early.
Pick one use case
Not every call type at once. One job with a clear outcome — following up unanswered enquiries, confirming appointments, qualifying a bought list. A narrow first agent is easier to judge and faster to correct.
Provision a number
A dedicated virtual number for the campaign. Using your main business line for outbound is possible but makes the results harder to read, because inbound returns and outbound dials mix.
Register under DLT
The regulatory step, and usually the long pole. The business registers as a Principal Entity, then the headers and content templates are submitted for approval before commercial calls can run.
Design the conversation
Not a script read aloud — a goal, the questions that reach it, and what to do at each refusal. This is the part that determines whether the call works.
Connect the systems
Where captured fields land: CRM, Sheets, a webhook, WhatsApp follow-up. Decide this before testing, because changing it afterwards means re-testing.
Test on a real list
A few hundred live numbers, not internal mobiles. Colleagues know what the agent is trying to do; strangers do not, and that is what you need to hear.
Go live, then tune
The first week is data. Connect rate, average duration, where people drop, which objection repeats. The agent is edited against that, not against how it sounded in testing.
DLT, DND and permitted calling hours
India regulates commercial calling more tightly than most markets, under TRAI’s Telecom Commercial Communications Customer Preference Regulations. Getting this wrong does not produce a warning — it produces blocked traffic.
DLT registration
Commercial calling requires the business to be registered on a Distributed Ledger Technology platform as a Principal Entity, with approved headers and content templates. Unregistered outbound traffic is liable to be filtered as spam before it reaches anyone.
DND and NCPR scrubbing
Promotional lists must be checked against the national Do Not Call preference registry before each run, not once at the start. People register continuously, so a list cleaned last month is not clean.
Permitted hours
Promotional calls run between 9am and 9pm IST. Scheduling outside the window is the most common avoidable violation, and it is trivially detectable from call logs.
Promotional versus service
Service and transactional calls — an appointment confirmation, a delivery update, an OTP — are treated differently from promotional ones. Adding an offer to a service call reclassifies the whole call as promotional, with the consent and timing rules that follow.
Disclosure and recording
Calls are recorded for quality and dispute resolution, and the recording is disclosed. Being straightforward that the caller is an AI agent also tends to go better than being evasive about it — people work it out, and the conversation sours when they feel misled.
On a managed plan the registrations, scrubbing and scheduling are handled as part of setup. On a self-serve platform they are yours to arrange, and they are the reason a platform that looks ready in an afternoon is not ready to dial for considerably longer.
Designing what the agent actually says
The most common setup mistake is writing a telescript and expecting it to be read. That produces the flat, obviously-automated call everybody hangs up on.
What works is a goal, the three or four questions that reach it, and a defined response at each point the person says no. The agent needs to survive being interrupted, because people interrupt — and it needs an honest answer for anything outside its scope rather than an invented one.
Open with why you are calling
The first five seconds decide the call. A caller who states the reason immediately gets heard; one that opens with a greeting script does not.
Languages, including mid-call switching
Nine or more Indian languages, with the ability to switch when the person answers in a different one. In practice this matters more than voice quality — a caller forced into their second language gives shorter answers.
A defined exit
Transfer to a person, book a callback, or end politely. An agent without a clean exit loops, and looping is what makes automated calls infuriating.
Connecting it to where the work happens
A call that captures information and leaves it in a dashboard has not finished the job.
Decide before testing where each captured field lands — CRM record, Google Sheet, webhook into your own system, WhatsApp follow-up, or a live transfer to whoever handles qualified interest. Transcripts and recordings should land in the same place as the fields, so a disputed lead can be checked without asking anyone.
CRM, WhatsApp, SMS, webhook and Google Sheets connections sit outside standard retail plans and are quoted with the build. Enterprise or legacy-system integrations are reviewed separately.
What to check before going live
Test on strangers. Internal testing tells you the agent works; it does not tell you whether it works on someone who was not expecting the call.
Run a few hundred real numbers
Enough to hear the range of reactions, small enough that a bad opening line has not reached your whole list.
Listen to the first five seconds
Most drop-off happens there. If people are hanging up before the second sentence, the problem is the opening, not the agent.
Check an objection you did not script
Call in yourself and say something unexpected. What it does when it does not know is more revealing than what it does when it does.
Confirm the handover
Trigger a live transfer and a callback booking end to end. Transfers that fail silently are the most expensive bug in a calling campaign, because the lead was qualified and then lost.
What decides how long setup takes
The honest answer is that the agent is rarely the constraint.
Conversation design, configuration and integration are measured in days. DLT registration and template approval are external and measured in whatever the platform and operator take, which is why they should start on day one rather than after the agent is built.
If you already hold a registered header and approved templates, a first campaign moves quickly. If you are registering from scratch, plan around that step and build everything else while it clears.
What you need to supply
Four things, and the first two are usually what hold projects up.
A clean contact list
With a lawful basis for calling it. Deduplicated, correctly formatted, and with any numbers you already know are wrong removed.
The actual offer
What you want the person to agree to, in one sentence. If this cannot be written in one sentence, the call will not work either.
Who takes the transfers, and when
A qualified lead transferred to an unanswered desk is worse than no call. Agreed hours and a named fallback.
Access, if integrating
CRM credentials or API access, and someone on your side who can approve a field mapping.
Setting it up yourself versus having it done
Both are legitimate. The decision is about who operates it, not which is better.
A self-serve platform at around ₹3 a minute sells you the engine. The DLT registration, the scrubbing, the call flow, the integration and the person watching the campaign are yours to arrange. That is genuinely cheaper if someone on your payroll will do it.
A managed plan at ₹10–12 a minute includes those. Most of the difference between the two numbers is the work, not the margin — which is worth checking against your own quotes rather than taking on trust. There is a fuller breakdown in AI calling plans and in compare AI calling rates.
AI calling setup questions
The agent itself is usually days — conversation design, configuration, integration and testing. The variable is DLT registration and template approval, which is external. If you already hold a registered header, a first campaign moves quickly; if not, start that step first and build the rest while it clears.
Yes, for commercial calling. The business registers as a Principal Entity with approved headers and content templates. Unregistered commercial traffic is liable to be filtered before it reaches anyone, so this is not a step that can be skipped and fixed later.
Promotional calls run between 9am and 9pm IST. Service and transactional calls — confirmations, reminders, delivery updates — are treated differently, but adding promotional content to a service call reclassifies it.
Technically yes, but a dedicated virtual number for the campaign is better practice. It keeps outbound campaign results separate from your normal inbound traffic, which otherwise makes both sets of numbers hard to read.
Calls are recorded and the recording is disclosed. Beyond the regulatory position, being straightforward that the caller is an AI agent generally produces better conversations than being evasive — people work it out, and the call sours once they feel misled.
Yes. Nine or more Indian languages, with switching inside a single call when the person answers in a different language. In practice this matters more than voice quality, because someone speaking their second language gives shorter answers.
It should say so and offer a route — a transfer, a callback, or a message to the team. An agent that invents an answer creates a dispute; one that loops creates a complaint. Both are setup failures rather than technology limits.
Yes. A controlled live demo on a small list is the normal way to begin, and it tells you more than any specification does. You can hear recorded calls first.