AI Voice Agent Best Practices for Indian Businesses
The best practices for an AI voice agent in India are operational, not technical: give it one job, write short scripts in the caller's own mix of language and English, respect calling windows and DND, define when it hands over to your team, and read a sample of transcripts every week. Teams that do these five things usually reach a stable agent within a month. Teams that skip them tend to blame the technology.
Outpero publishes this post and sells an AI voice agent. Facts about other vendors come from their own websites, read on 2026-10-14, and are labelled as the vendor's claim where they are claims. The post gives criteria and a test method so you can check any vendor, Outpero included, for yourself.
Why best practice is mostly about operations
An AI voice agent is software that makes or answers phone calls and holds a spoken conversation. The speech, language and voice components are bought, not built, so two businesses on the same product can get very different results. The difference is almost always how the agent was briefed, tested and reviewed.
Think of it as onboarding a new telecaller who never tires. A human recruit gets a job description, a script, a first week of shadowing and a manager who listens to calls. An AI employee needs the same four things, and most failures trace back to one of them being skipped.
The ten practices below follow the life of a deployment: before launch, on the call, after the call, and the conversations to keep the agent away from altogether.
Before launch
1. Give the agent one job
A first agent should do one thing: call new leads within a minute and book a site visit, confirm cash-on-delivery orders, remind patients about tomorrow's appointment, or answer the business line after hours. Every extra job multiplies the number of conversations you must test.
Start with the job where speed or volume already hurts. A real-estate team that loses leads because nobody calls within an hour has a clear first job. A clinic that loses a share of its slots to no-shows has another. Once the first agent is stable, clone it for the next job rather than stretching it.
2. Write the goal as an outcome, not a script
State what a successful call produces: a booked slot with a date and time, a yes or no on the order, a callback request with a preferred hour. Then list the fields to capture, such as budget, location, language preference and the reason for any refusal.
Outpero stores the recording, transcript, summary, sentiment and the custom fields you asked for in the dashboard, and sends them to your webhook, so define those fields before you build. An agent that captures nothing structured turns every call into a transcript someone has to read.
3. Test in the caller's real language, with the English mixed in
Indian callers rarely speak one clean language. A Hyderabad caller says the budget in English and the locality in Telugu, a Pune caller answers in Marathi and asks the price in English, a Chennai caller drifts into Tanglish halfway through a sentence. The agent must follow the switch without restarting the sentence or changing voice.
Run at least ten test calls in the language of your customers, from different phones, with at least one tester whose accent differs from the person who built the agent. Include numbers: phone digits, amounts in lakh and crore, and dates. Number handling is where agents most often slip.
Outpero speaks Hindi, Telugu, Tamil, Marathi, Kannada, Malayalam and Indian English at the same per-minute price in each. Bengali, Gujarati, Punjabi and other languages are not on its list, so a business that needs those should check each vendor's own language list before testing anything else.
On the call
4. Keep the opening to fifteen seconds
The first fifteen seconds decide whether the caller stays. Name the business, say why you are calling in one clause, and ask one easy question. Do not read a paragraph about the company.
नमस्ते, मैं Sai Estates से बोल रहा हूँ। आपने अभी हमारी site पर 2BHK के बारे में enquiry डाली थी। क्या अभी दो मिनट बात कर सकते हैं?
Never deny being an AI if the caller asks. A short, honest answer such as "I am the business's AI assistant, and I can book this for you right now" keeps most callers on the line, and it avoids the worse outcome of a caller who feels tricked.
5. Ask one question at a time and confirm what matters
Spoken conversations are not forms. Ask for one thing, wait, then move on. Read back the items that cost money if they are wrong: the appointment time, the delivery pin code, the phone number to call back. Confirm digits in groups, the way a person would.
Short agent turns also help the pace of the call. Outpero's end-to-end conversational latency is 0.8 to 1.3 seconds on average from the moment the caller stops speaking to the moment the agent starts, and a short reply keeps the exchange feeling like a conversation rather than a recording.
6. Respect calling windows, DND and consent
Want to hear this working, not just read about it?
Outbound calling in India is regulated. TRAI's rules on commercial communication, administered through DLT registration and the DND registry, apply to promotional calls. Keep calls inside sensible hours, usually mid-morning to early evening, screen lists against DND, and stop contacting anyone who asks to be removed.
Outpero bulk campaigns have calling windows, retries and DND screening, and the calling line is DLT registered. Those controls only help if you set them. Treat the calling window as part of the script: an excellent opening line at 10 p.m. is still a complaint.
Compliance depends on the kind of call (service, transactional or promotional) and on your sector. Confirm the rules with your telecom provider and a lawyer, because this post is operational advice and not legal advice.
The practices at a glance
| Practice | What goes wrong without it | How to check it before launch |
|---|---|---|
| One job per agent | Dozens of untested branches; wrong answers in the corners | Write the single outcome in one sentence |
| Outcome plus fields | Transcripts nobody has time to read | List the fields the CRM or sheet needs |
| Language and mixing tests | Agent restarts or goes stiff when the caller switches to English | Ten calls, three accents, numbers and dates included |
| Fifteen-second opening | Callers hang up before the first question | Read it aloud with a stopwatch |
| One question at a time | Callers answer the wrong question | Listen for talk-over in test recordings |
| Windows, DND, consent | Complaints, blocked numbers, regulator risk | Set the window and DND screening, then dial a test list |
| Handover rules | Angry or high-value callers trapped with software | Test each handover trigger once |
| Retry and callback logic | Leads called five times in one hour, or only once | Review the retry schedule on paper |
| Weekly transcript review | Silent drift; nobody notices a bad branch | Put the review in a calendar |
| Clear no-go list | Agent improvises on grief, legal or medical matters | Write the list and test it |
After the call
7. Decide the handover rules before the first real call
Outpero can transfer a call to your team when a person should take it, as an opt-in setting per employee. The practice is to decide the triggers in advance. Typical triggers are a caller who asks for a person, a negotiation above a threshold, a complaint, or two failed attempts to understand the caller.
Do not promise a human on every call, and do not hide the option either. Make sure someone is actually available to take transferred calls during the window, or have the agent offer a callback instead. A transfer that rings out is worse than an honest callback offer.
8. Design retries and callbacks like a sales process
Most outbound calls are not answered the first time. A workable starting pattern for leads is a first call within 30 seconds of the enquiry, a retry a few hours later, a retry the next day at a different time, and then a stop. Unanswered calls are not billed by Outpero, so retries cost you only when someone picks up.
Offer a callback whenever the caller is busy. A scheduled callback at the hour the caller chose usually serves you better than a fourth cold retry, and it shows respect for their time. Caller memory helps here: the agent recognises a returning caller and picks up from the last conversation instead of starting over.
9. Read transcripts every week
Every call produces a recording, transcript and summary, so there is no need to guess. Sample twenty calls a week, mixing successes, failures and short calls. Short calls are the most informative, because they show where callers leave.
- —Listen to the first ten seconds of five short calls. Is the opening wrong?
- —Read three failed bookings. Did the agent miss an objection, or was the slot genuinely unavailable?
- —Check two transfers. Did the person who took the call have the context?
- —Note any question the agent could not answer and add it to its knowledge, or to the handover list.
- —Check the call log for credits used per call, so a spike in call length is caught early.
Fix one thing per week and re-test. Changing six things at once makes it impossible to know which change helped.
Keep the agent away from some conversations
10. Write the no-go list
Some calls need a person from the first word: bereavement, a medical emergency, a legal dispute, a regulated advice conversation, a customer who has already complained three times. The agent should recognise the signal, say a colleague will call back or transfer the call, and stop.
For lenders, insurers and clinics, add the regulator's constraints to the list. RBI, IRDAI and NMC rules differ and change, and the business remains responsible for what is said on its behalf.
A 30-day rollout that works
- 01Days 1 to 3: choose the job, define the outcome and fields, and write the opening line and five likely objections in the caller's language.
- 02Days 4 to 7: build the employee and run the ten test calls. The 50 free credits cover about 10 minutes of conversation at the Standard rate of ₹5 per minute, so keep the first tests short.
- 03Week 2: call a small real list of 50 to 100 contacts, with handover on and your team on standby. Read every transcript.
- 04Week 3: fix what the transcripts showed, widen the list and set retries and the calling window.
- 05Week 4: move to full volume, start the weekly review and compare cost and outcomes against your previous process.
The cost of this pilot is small. At Standard, ₹5 per connected minute, 100 two-minute conversations cost about ₹1,000 plus 18% GST, with no setup fee and no monthly subscription. A dedicated +91 business number is ₹649 per month and includes one channel, which means one call at a time. Each extra channel is ₹649 per month.
Mistakes that repeat
- —Launching on the whole database on day one, then discovering a script problem thousands of calls later.
- —Using a formal textbook register that no real caller uses in a sales call.
- —Measuring calls placed instead of outcomes: bookings, confirmed orders, qualified leads.
- —Forgetting that the calling line, DND screening and time window are part of the caller's experience.
- —Treating the agent as a replacement for judgement. It handles the repeatable part of a flow so people can handle the rest.
None of this needs a technical team. Outpero is self-serve: describe the job, choose a voice tier, connect your lead source or upload a CSV, and start with the free credits.
Frequently asked questions
What are the most important AI voice agent best practices?
+
Give the agent one job, write short scripts in the language your customers actually use, respect calling windows and DND, decide handover rules before launch, and review a sample of transcripts every week. These five practices explain most of the gap between a stable deployment and a disappointing one, whichever vendor you use.
How many test calls should I run before going live?
+
Run at least ten test calls in the customers' language, from different phones, with at least one tester whose accent differs from the builder's. Include phone numbers, amounts and dates. Then call a small real list of 50 to 100 contacts with your team on standby before opening the full database.
Should an AI voice agent say it is an AI?
+
It should never deny it when asked. Many businesses open with the agent's role, such as the business's assistant, and answer honestly if the caller asks directly. Callers tolerate an honest AI far better than one they feel was disguised, and honesty avoids complaints and a loss of trust later.
How do I keep an AI voice agent compliant with TRAI and DND rules?
+
Use a DLT-registered calling line, screen outbound lists against the DND registry, keep calls inside sensible hours, honour opt-outs immediately and keep records. Rules differ for service, transactional and promotional calls and by sector, so confirm the specifics with your telecom provider and a lawyer.
When should an AI voice agent hand over to a person?
+
Hand over when the caller asks for a person, when the conversation is a complaint or a high-value negotiation, when the agent fails twice to understand, or when the topic is on your no-go list. Decide the triggers in advance and make sure staff are available, or offer a callback instead.
How often should I review AI call transcripts?
+
Weekly during the first month, then at least fortnightly. Sample about twenty calls, including short ones and failures, and fix one issue at a time. The call log shows the credits each call used, which helps you spot unusual call lengths early, before they affect your spend.
The Outpero team builds AI voice agents for Indian businesses, in Hindi, Telugu, Tamil, Marathi, Kannada, Malayalam and English. Figures are from Outpero's own product unless another source is named.