A form fills in at 9.40 on a Tuesday. Your sales team is in a review meeting until eleven. By the time someone opens the shared inbox, the enquiry is ninety minutes cold and the person who sent it has filled in two more forms elsewhere. This is the gap an AI voice agent is supposed to cover, and most of the arguments about it are arguments about how far that cover should extend.
The honest answer is: not very far. An AI voice agent is good at one narrow job, which is making contact quickly and finding out three things. It is bad at almost everything after that. Teams get into trouble when they let the tool keep talking past the point where it should have handed over.
What the first call is actually for
A first call is not a sales call. It is a confirmation call. You are confirming that a real person sent the enquiry, that they want the thing you sell, and that there is a sensible next step. That is the entire brief.
If you accept that brief, the design of the AI voice agent becomes obvious. It needs to be short, polite, clear about what it is, and quick to get out of the way. It does not need to handle objections. It does not need to explain your pricing model. It does not need to be charming.
The reason to automate this at all is timing, not labour. A lead called back within a few minutes gets a different reception from the same lead called back the next afternoon. That is the argument made in more detail in the piece on fixing first response time, and the voice agent is one way to close that window when your team physically cannot. It is not the only way, and it is not the best way when your team can.
Disclose in the first ten seconds
Say what it is. Not buried, not implied, not hidden behind a human-sounding name.
The opening line should contain three pieces of information: the company name, the fact that this is an automated call, and the reason for the call. Something in the shape of “This is an automated call from Sharma Industries about the enquiry you submitted this morning.” Then a pause, then the first question.
There are three reasons to disclose, and only one of them is about ethics.
The practical reason is that people can tell. Indian callers in particular are used to recorded voices and IVR trees, and the moment someone suspects they are being tricked, they stop cooperating. An undisclosed AI voice agent gets short answers and early hangups. A disclosed one gets more patience, because the person knows what they are dealing with and adjusts their expectations.
The second reason is reputational. If the first interaction anyone has with your business involves a small deception, every later interaction inherits that. Discovering it later is worse than knowing it upfront.
The third is that disclosure lets you be shorter. A human caller has to do social work: how are you, sorry to disturb, is this a good time. A disclosed automated call can skip all that without seeming rude, because nobody expects small talk from a machine. Disclosure buys you brevity.
Also state the exit. “You can say stop at any point and I will end the call.” And honour it immediately, without a retention attempt.
The three questions worth asking
Most first-call scripts try to gather ten fields. Cut it to three. Every additional question raises the chance of a hangup, and the marginal information is usually something your salesperson would rather ask themselves.
One. Is this still relevant? Phrase it as a confirmation, not an interrogation. “You enquired about warehouse racking. Is that still something you are looking at?” A surprising share of enquiries are already dead on arrival: someone browsing, someone comparing, someone who solved it internally. Finding this out in fifteen seconds is worth more than any other data point, because it stops a salesperson spending forty minutes on a lead that closed itself.
Two. What is the rough timeline? Not budget. Timeline. “Are you looking to move on this within the month, or is it further out?” Timeline is the single most useful thing for routing, because it decides urgency without requiring judgement. Someone who says “this week” goes to the top of the queue. Someone who says “maybe after Diwali” goes into a nurture cadence and stops distorting your pipeline. People also answer timeline questions honestly, which is not true of budget questions on a first call with a machine.
Three. Who should we speak to, and when? This is the handoff question. It captures the correct name and the correct window. It is the only question that directly makes the next call better. “Is there a good time tomorrow for one of our team to call you?” If they name a slot, you have a booked call rather than a callback attempt.
That is it. Three questions, under ninety seconds, and everything the agent learns lands on the lead record so that whoever picks it up is not starting from zero. On BizvoraOne the answers attach to the lead in the CRM alongside the source, so the human call opens with context rather than with “so, tell me what you need.”
What you should not ask: budget, technical specifications, decision-making authority, competitor comparisons, or anything requiring a number. Not because those do not matter. They matter enormously. They just cannot be collected reliably by a script that cannot ask a follow-up.
The handoff rule
Here is the rule, and it is one sentence.
The moment the caller asks a question the agent was not built to answer, the call becomes a handoff.
Not a deflection. Not a “let me note that down”. A handoff. Either transfer live if someone is available, or say plainly that a person will call, name the window, and end the call.
The failure mode to avoid is the agent that keeps going. Someone asks about GST treatment on a particular line item, and instead of stopping, the agent produces a plausible general answer. The caller now believes something that may be wrong, and your salesperson has to unwind it before selling anything. Every wrong confident answer costs you two conversations instead of saving you one.
So build the stop conditions explicitly:
- Any question about price, discount or payment terms
- Any question about delivery dates or capacity
- Any technical or compliance question
- Any expression of frustration, however mild
- Any request to speak to a person, first time of asking
- Any answer the agent has to ask twice to parse
Hit any of those, and the script’s job is to close politely and create a task with a named owner and a deadline. A handoff without a named owner is not a handoff. It is a lead sitting in a queue, which is the same problem you started with.
The same discipline applies elsewhere in the stack. There is a version of this argument about drafting, covered in the piece on keeping AI out of the scope section: the tool does the first pass, a person owns anything that commits the business.
Four situations you never automate
Some calls should never be placed by an AI voice agent, whatever the response time argument says. Configure these as exclusions before you switch anything on.
A complaint or a service failure. If the record shows an existing customer with an open issue, an automated call reads as contempt. Someone whose order is late does not want a machine confirming their timeline. This is the exclusion most people forget, because complaint records and lead records often live in different places.
A referral or a warm introduction. Someone was recommended to you by a person they trust. The value of that introduction is the human relationship behind it, and an automated first call spends that value for nothing. Call these yourself, same day.
High-value or named accounts. Set a threshold that suits your business. Above it, a person calls. The cost of a partner-level or director-level call is trivial next to the deal, and the signal it sends is the whole point. In professional services and manufacturing enquiries of any size, this exclusion covers most of what matters.
Anything sensitive. Health, legal trouble, debt, employment, family matters. Education enquiries about a struggling child. A person deals with those, always. If you are not sure whether a category is sensitive, treat it as sensitive.
There is a fifth situation, less absolute: repeat contact. If the agent has already called this number once and got no useful answer, do not call again automatically. Switch to a human or to email. Automated persistence is just nuisance calling with better infrastructure.
What to do first
Write the ninety-second script on one page. Disclosure line, three questions, close. Read it aloud. If it takes longer than ninety seconds at a normal pace, cut a question rather than speeding it up.
Then write the exclusion list against your own lead sources, so you know which of your fifteen sources are eligible for automated first contact and which are not. Referrals almost certainly are not.
Then run it on one segment for two weeks. Listen to twenty recordings yourself. You are not listening for whether the agent sounded natural. You are listening for the moment it should have handed over and did not. Fix those, and the rest of the design mostly takes care of itself.
