The most common reason an Agni deployment underperforms is not the AI — it's the prompt. A well-configured AI agent with a weak prompt will consistently underperform a simpler configuration with a sharp, well-structured prompt. This guide shows you what good looks like.
The Anatomy of an Agni Agent Prompt
An effective Agni agent prompt has five components:
- Identity: Who is the agent, what is its name, and what company does it represent?
- Objective: What is the single primary goal of this call?
- Context: What does the agent know about the person it's calling?
- Call flow: What is the sequence of the conversation?
- Constraints: What should the agent never say or do?
Identity: Setting the Right First Impression
The agent's identity should be specific and warm. Avoid generic names like "Customer Care Agent." Use a first name that fits the brand and the language context.
Weak: "You are a customer service agent for ABC Finance."
Strong: "Your name is Priya. You are a customer relationship executive at ABC Finance calling on behalf of the loan servicing team. You speak in warm, friendly Hindi and Hinglish, like a colleague explaining something helpful to a friend."
The difference: the second prompt gives the model specific tone guidance ("warm, friendly"), a cultural reference point ("like a colleague"), and a language instruction ("Hindi and Hinglish").
Objective: One Goal Per Agent
Each Agni agent should have exactly one primary objective. Multi-objective prompts produce confused, meandering calls.
Wrong: "This agent handles EMI reminders, account queries, and upselling loan top-ups."
Right: "Your one goal is to confirm that the customer is aware their EMI of ₹{emi_amount} is due on {due_date} and to collect a commitment to pay. Everything else is secondary."
Secondary objectives are fine — but they should be clearly ranked. "If the customer commits to paying AND has time remaining in the call, you can mention the loan top-up offer. This is optional. Do not mention it if the customer seems rushed or unhappy."
Context Variables: Making Calls Feel Personal
Agni supports dynamic variable injection from your CRM or campaign data. Use these to personalise every call:
{customer_name}— use at call opening and at least once more during the call{product_name}— the specific product or loan{amount}— the specific amount due, outstanding, or offered{due_date}— formatted in Indian style (15 June, not June 15){last_interaction}— if available, what happened on the last call
Each personalisation element you add increases the probability of the customer engaging with the call rather than hanging up.
Call Flow: Structure Without Rigidity
Indian conversations do not follow Western sales scripts. Build your flow with explicit decision points rather than a linear script:
Opening → Verify identity → State purpose →
IF customer acknowledges: → Main pitch →
IF interested: → Next step (payment / appointment / transfer) →
IF objection: → Handle objection → Return to main pitch →
IF hard no: → Schedule callback or close gracefully →
IF customer busy: → Offer callback time → Close →
IF wrong number: → Apologise and close
Write each branch as a separate paragraph in your prompt, with explicit "if the customer says X" instructions.
Constraints: What the Agent Must Never Do
For compliance and brand safety, always include a constraints section:
- "Never threaten legal action unless explicitly authorised in the script"
- "Never quote interest rates or charges unless they are in the context provided — if the customer asks, say you'll have a colleague call back with details"
- "Never promise refunds, waivers, or adjustments — offer to connect to a senior executive instead"
- "If the customer becomes abusive or distressed, de-escalate and offer to transfer to a human agent"
- "Do not continue the call if the customer clearly and repeatedly says they do not want to be called — end the call and flag for DNC list addition"
Language-Specific Instructions for Indian Calls
Add explicit language guidance for mixed-language calls:
- "Respond in whatever language the customer uses. If they speak Hindi, respond in Hindi. If they switch to English, match them."
- "Use respectful pronouns: 'aap' (not 'tum') for new contacts. Shift to 'tum' only if the customer uses it first."
- "Use 'ji' as a suffix when addressing the customer by name in Hindi — 'Rahul ji' not just 'Rahul'."
- "For numbers, use Indian number system: 'teen lakh' not 'three hundred thousand'."
Testing Your Prompt
Before going live, run at least 10 test calls across different scenarios. Test:
- A cooperative customer who wants to complete the call quickly
- A customer who raises the most common objection in your use case
- A customer who switches languages mid-call
- A customer who asks a question outside the agent's scope
- An angry or uncooperative customer
Review the call transcripts and recordings. Anywhere the agent gives an incorrect, confusing, or off-brand response, add a specific instruction to your prompt. Three rounds of prompt iteration typically produce a production-ready agent.
