Voice AI for businesses that do not run around the clock

Most voice AI proposals start in the wrong place. They pitch the daytime: fewer people on the phones, faster handling, lower cost per call. That is a hard argument to win. It asks a business to accept that software will do a job a person is currently doing well, it invites a quality comparison the agent will sometimes lose, and it quietly makes everyone who answers phones an opponent of the project.

There is an easier case, and most businesses already have it. It is the phone ringing at nine in the evening.

The gap nobody is defending

During business hours, a voice agent competes with a trained person who knows your customers. After hours it competes with a voicemail box nobody checks, or with a phone that simply rings out. That is a different bar entirely.

It also changes the internal politics. Nobody on your team is protective of the calls they are not there to take. An after-hours agent does not threaten anyone's job, which means the project gets evaluated on whether it works rather than on what it implies.

The loss it addresses is real but close to invisible. A caller who reaches nothing at nine in the evening does not usually wait until morning. For anything commodity or urgent — an appointment, a repair, a quote — they work down the search results and call the next number. You never learn that this happened. It does not appear in any report you run. It looks exactly like a call that was never made.

Measure before you model

Almost every business estimates its after-hours volume, and almost every one of those estimates is wrong. The data already exists: your telephony provider keeps call detail records, and they will tell you precisely what you need to know.

Pull twelve months and answer four questions.

1. What share of inbound calls arrive outside your working hours? Include weekends and public holidays. Businesses are often surprised — a nine-to-six operation can find a fifth of its call volume lands outside that window.
2. What happens to those calls now? Voicemail with a listened-to rate you can actually check, a diverted mobile, or nothing at all.
3. How many are new enquiries rather than existing customers? This is the number that decides everything downstream, and it is the one people guess at most confidently.
4. What is a converted new customer worth to you? Not revenue on the first transaction — lifetime value, or your best honest approximation of it.

If you cannot answer the fourth question, stop. Every calculation below depends on it, and no vendor can supply it for you.

What an agent can do when there is nobody to escalate to

Escalation is the standard safety valve in voice AI design: when the agent reaches the edge of what it can handle, it hands the caller to a person with the context already gathered. At eleven at night there is no person. That single constraint reshapes the whole design.

In practice the work sorts into four tiers, each asking more of you than the one above it.

01

Capture, properly

Who called, what they wanted, a callback number read back to them, and how urgent it is.

Nothing beyond the agent

02

Answer what is knowable

Hours, location, documents to bring, whether you take a particular insurer.

A written knowledge base

03

Transact

Book, reschedule, cancel, check an order. Where most of the value is.

Live system integration

04

Genuine emergencies

Recognise the situation quickly and route it to a person who is awake.

A real on-call rota

Two of those deserve more than a line. Capture sounds like voicemail and is not: voicemail hands you a recording somebody has to sit and listen to, while a good capture hands your morning team a structured, prioritised list with every callback number already verified because the agent read it back. And transact is where the build cost concentrates — but it is also the only tier that earns revenue rather than merely preserving it, because a booking taken at nine in the evening is money that would otherwise have gone elsewhere.

One warning on the fourth tier. If you do not have an on-call rota, the agent must not imply that you do.

The failure mode to design against is an agent that can neither help nor hand over, and papers over the gap. If it cannot do the thing, it should say so plainly, commit to a specific callback window, and make sure that callback actually happens. A caller who is told "someone will phone you before ten tomorrow" and then is phoned before ten has been served well. A caller who is told something vague and reassuring has been served worse than by a voicemail beep.

The economics, worked

The run cost of a voice agent is reasonably predictable. On a typical Indian inbound stack — telephony, speech recognition, a language model, and speech synthesis — published vendor rates land somewhere around ₹7 to ₹9 per connected minute once every component is counted. Your own figure will differ with volume and vendor choice, and it is worth calculating properly rather than taking anyone's word for it, including ours.

Here is the shape of the calculation. The numbers are illustrative; substitute your own.

Input	Illustrative value
Calls arriving outside hours	15 per day
Average handled call length	3 minutes
Monthly connected minutes	~1,350
Run cost at ₹8 per minute	~₹10,800 per month
Share that are new enquiries	30% (~135 per month)
Conversion if answered	15% (~20 customers)

At those figures the agent needs each converted customer to be worth roughly ₹540 before the run cost is covered. If a customer is worth ₹5,000 to you, the case is not close and you should stop deliberating. If a customer is worth ₹600, the case is marginal and the build cost has not even entered the calculation yet.

That sensitivity is the point. The economics of an after-hours agent are decided almost entirely by two numbers — how many after-hours callers are new business, and what a new customer is worth. Everything else is rounding.

When it does not pay back

Be honest with yourself about these.

Your after-hours volume is genuinely small. Below a few calls an evening, the engineering effort will not earn back, whatever a vendor tells you.

Your after-hours callers are almost all existing customers with non-urgent questions. They will call back tomorrow. You are automating something that was not costing you anything.

Your callers would be content to leave a voicemail. Some sectors have trained their customers into this, and if yours has, the incremental gain over a better voicemail greeting is thin.

Your calls are long and unstructured. A twenty-minute consultative conversation is not the same problem as a three-minute booking, and pretending otherwise is how pilots fail.

Start narrower than you think

The most common mistake is trying to make the after-hours agent do everything the daytime team does. It arrives late, costs more than it should, and gets judged against a standard it was never going to reach.

Start with capture and the questions that have fixed answers. That is a small build, it is difficult to get badly wrong, and it produces something valuable immediately: a structured record of exactly what people are calling about after hours. Two months of that data will tell you which transactions are worth automating next — and it will tell you with evidence, rather than with the assumption you would otherwise have built on.