The Actual Maths on an AI Receptionist

An AI voice agent answers calls, understands what the caller wants, resolves what it can and transfers the rest. Priced per AI-handled minute, the business case looks straightforward. It usually is — but the calculation people run is not the one that determines the outcome.

The four numbers

1. Calls you currently miss. The most valuable and least measured. Every business has calls that ring out, hit voicemail and are never returned, or arrive while everyone is on another line. Your phone system knows: look at unanswered inbound, after-hours inbound, and abandoned calls in queue.

2. Average value of a call. Not of a customer — of an inbound call. Total revenue attributable to inbound divided by inbound call volume. Imperfect, and good enough to make a decision.

3. Time spent on calls that need no judgement. Opening hours, directions, appointment status, "can I speak to X". Count them for a week.

4. Metered cost. At a published rate per AI-handled minute, cost is volume × average AI handling time × rate.

A worked example

A practice taking 1,200 inbound calls a month, with 18% unanswered — around 216 missed calls. Average inbound call value of $40, with a third of missed calls representing genuinely lost business:

216 × 0.33 × $40 ≈ $2,850 a month in recovered value

On the cost side, if the agent handles 800 calls at an average of 90 seconds — 1,200 AI minutes — at $0.08 per minute:

1,200 × $0.08 = $96 a month

That gap is large enough that precision in the inputs barely matters. Even if the recovered value is a third of the estimate, the case holds comfortably.

This is why metered AI voice is interesting: the cost scales with usage rather than seats, and it is small relative to the value of a single recovered customer in most businesses.

The time-saving side, honestly

The second benefit — staff time returned — is real but easier to overstate.

If the agent handles 400 routine calls a month at three minutes each, that is 20 hours. At a loaded cost of $30 an hour, $600 of time.

The honest caveat: that time is only a saving if it is redeployed. In a busy practice, 20 hours returned to a receptionist who is already at capacity means patients get better attention and fewer calls are missed — genuine value, hard to put on an invoice. In an understaffed operation it may mean you delay a hire, which is a real saving.

Do not present recovered minutes as cash unless you can say what they will be spent on.

The failure mode that destroys the return

A bad transfer is worse than a missed call.

Someone who reaches voicemail knows where they stand. Someone who spends ninety seconds with an automated system that misunderstands them, then gets transferred to the wrong place, then has to repeat everything, has had a bad experience with your business — and they will attribute it to you, not to the technology.

The economics only work if the agent is genuinely good at the narrow set of things it handles and hands over quickly and cleanly on everything else. Which means:

• Scope it narrowly at first. Hours, directions, appointment confirmation, routing by department. Add intents once each is reliably handled
• Transfer early rather than persisting. Two failed attempts to understand should route to a human, with context passed
• Pass what it already knows. A caller who gave their name and reason should not be asked again by the person who picks up
• Always allow escape. "Speak to someone" must work at any point
• Monitor transcripts weekly at first. The gap between intended and actual handling is where the tuning is

What to measure after deployment

• Containment rate — calls fully handled without transfer. Rising is good, but only alongside the next metric
• Post-transfer repeat rate — callers transferred who then need a second contact. This is the quality signal. A high containment rate with high repeats means the agent is holding calls it should be handing over
• Abandonment during AI handling — people hanging up on the agent
• Missed call rate, before and after. The primary objective
• Cost per handled call — metered spend divided by calls handled

Where it does not pay

Be clear about the cases where this is the wrong purchase:

• Very low call volume. Twenty calls a month does not justify the setup effort, however cheap the minutes
• Calls that are almost all complex. If every caller needs judgement, there is nothing to contain
• No capacity problem and no missed calls. If calls are answered promptly today, the benefit is time saving alone, which is a weaker case
• Where a bad interaction is very costly. Some businesses — crisis services, high-value relationship sales — should not put an automated layer in front of a caller at all

The decision in one line

Pull your unanswered and after-hours call counts for last month. Multiply by what an inbound call is worth and by the fraction genuinely lost. If that number is meaningfully larger than volume × 1.5 minutes × your provider's per-minute rate, the arithmetic works — and the remaining question is entirely about execution quality, not cost.