AI receptionist vs answering service: the honest math

If your prospect already pays an answering service, you are not selling them their first phone solution. You are selling a cheaper, faster, more consistent front end that hands the hard calls to the humans they already pay. A human answering service runs about $3.45 to $5.00 a minute at Ruby's published rates and $7 to $11.50 a call at Smith.ai's. The raw infrastructure for an AI receptionist runs under 15 cents a minute. That gap is the whole business, and the way to win the deal is to stop pitching replacement and start pitching the order the two things run in.

This is the comparison I actually run on calls. Published prices on both sides, where the AI is better, where the answering service is better and you should say so, and how to price against an invoice your prospect is already paying.

What does an answering service actually cost?

Start with real numbers, because most prospects cannot tell you their effective rate off the top of their head.

Ruby publishes four virtual receptionist plans: $250 for 50 minutes, $395 for 100, $720 for 200, and $1,725 for 500. Divide it out and you get $5.00, $3.95, $3.60 and $3.45 a minute. Their chat product is priced separately, starting at $143 for 10 chats. Ruby states no setup, activation or customization fees, and includes 24/7 live answering on every plan.

Smith.ai prices by call instead of by minute. Their published plans are $300 for 30 calls, $810 for 90, and $2,100 for 300, which works out to $10.00, $9.00 and $7.00 a call. Overage runs $11.50, $10.50 and $8.50 a call depending on plan. No setup fee, month to month, 10 percent off for annual prepayment.

Both of those are honest, well-run businesses. Notice what the pricing shape does to your prospect, though. On a per-minute plan, the four minute call where someone asks about hours and price costs the same as the four minute call that lands a $9,000 job. On a per-call plan, a wrong number that gets picked up still burns an included call. The service has every reason to answer professionally and no reason to care whether the call was worth answering.

Two more things to pull out of the invoice when you get your hands on one. First, per-minute plans bill in increments, so the 40 second call is rarely billed as 40 seconds. Second, the monthly block is use-it-or-lose-it, which means a business with seasonal volume pays for the peak and eats the trough. A roofing company in February is subsidizing a roofing company in June.

What does the AI side cost to run?

Published component pricing, not a guess. Vapi lists $0.05 a minute for platform hosting, plus pass-through on the pieces: $0.0095 to $0.0099 a minute for Deepgram transcription, $0.0077 to $0.0452 for the model, and $0.0146 to $0.0238 for ElevenLabs voice. Add it up and a minute of voice agent costs roughly 8 to 13 cents in raw spend, before anybody's margin.

Retail is a different number, and it should be. On Lumina builds, voice runs $0.35 a minute at wholesale on top of the build's setup and monthly, agreed on a 30-minute call. Set your retail above that and the spread is yours. We broke the per-minute stack down in more detail in what a voice AI agent costs per minute, including why the quoted platform rate is never the whole bill.

Here is the comparison that matters on a call. Take a three minute inbound call, the length of a normal "are you open Saturday and how much is a service visit" conversation:

• Ruby Business plan: about $10.80
• Smith.ai Starter plan: $10.00 flat, or $11.50 on overage
• AI receptionist at $0.35 a minute wholesale: $1.05

Do not lead with that in the room. A 10x cost gap sounds like a trick to a business owner who has been pitched by four AI guys this quarter. Lead with coverage and consistency, let them discover the cost difference when they look at their own invoice, and price your service against the invoice rather than against your cost. Our post on margin math walks through what you keep when you price that way.

Where the AI genuinely wins

Four places, and they are all mechanical rather than aspirational.

It knows the business. An answering service operator is reading a script and a FAQ sheet, and they are handling twelve accounts that shift. An agent built on a vertical pack knows the service list, the price bands, the objection a caller raises right before they book, and the booking language that trade actually uses. That is the difference between "let me take a message and have someone call you back" and "a standard diagnostic visit runs $89 and I have Thursday at 10 or Friday at 2." We ship 100-plus trade packs for exactly this reason, and a generic prompt loses demos on this point alone.

It answers on the first ring, at any volume. An answering service has operators on shift. At 2pm in August, when every HVAC phone in the county is ringing, that queue is real. An AI receptionist has no queue and no eleventh-hour staffing problem. Ten simultaneous calls get ten answered calls.

It books instead of taking a message. This is the one owners care about most and the one they least expect. A message in an inbox is a lead that still needs a callback. A calendar slot is a job. Most answering services will offer scheduling, but the operator is working inside a system they do not live in, which is why message-taking is the default.

It does the thing the business keeps meaning to do. Missed-call text-back, follow-up on a quote that went quiet, a nurture sequence that runs whether or not anybody remembers. A human answering service answers calls. It does not run your follow-up.

Where the answering service still wins, and you should say so

Skip this part and you will lose the deal to the first person who says it out loud.

The weird call. A caller whose house is flooding, a caller who is furious, a caller who is confused and in the wrong account entirely. A good operator reads that in two seconds and does something sensible. An AI agent handles the ones you anticipated. On the ones you did not, the best outcome is a clean handoff, not a clever recovery.

Callers want a person available. Gartner surveyed 3,566 B2B and B2C customers in February and March 2026 and found 87 percent say access to a human agent is essential when a company uses generative AI for service. Half said the interaction was easier with AI, which is the interesting half, but the 87 percent is the constraint. Gartner's Eric Keller put it directly: service leaders should not use AI as a mandatory first step for every issue. Build the escalation path or you are selling your client a wall.

There is still a timing gap you can hear. AssemblyAI's engineering writeup on voice latency is worth reading before you promise anyone the AI is indistinguishable. Human conversation turns over in about 300 milliseconds, and under that nobody notices the gap. Past 500 milliseconds a caller starts to wonder if they were heard. Past a second they repeat themselves. Their own voice agent pipeline lands around one second end to end for a full conversational turn, and that is a team optimizing for this specifically. So the honest claim is not that callers cannot tell. It is that the agent is good enough that most callers do not care, which is a different and more defensible thing to say. We went deeper on that in does an AI receptionist sound human.

Names, addresses and spelling. Transcription on a bad cell connection with a hard surname is where AI receptionists embarrass themselves. Design around it. Confirm back, spell out, text the booking confirmation so the caller can correct it in writing.

I would rather hand a prospect this list myself than have their office manager find it in week two.

The setup that actually wins the deal

Stop selling replacement. Sell order of operations.

The AI answers every call first. It handles hours, services, pricing bands, directions, appointment booking, rescheduling and the sixty percent of calls that are the same eight questions. It escalates on a defined trigger: the caller asks for a person, the caller says an emergency word, the call is about an existing complaint, or the agent fails twice on the same question. Escalation goes to the business's cell during hours and to the answering service after hours.

What happens to the answering service bill? It drops, because the AI absorbs the short repetitive calls that the per-minute and per-call models both overcharge for, and the service keeps the small number of calls where a human is worth $10. Your prospect is not being asked to fire a vendor on your word. They are being asked to put something in front of it and watch the invoice shrink.

This framing does three useful things. It removes the risk objection, because nothing gets turned off. It makes the first month measurable, since the answering service invoice is the before number. And it quietly makes you the owner of the front door, which is where every future upsell lives.

What about just hiring someone?

Sometimes they raise this instead. The Bureau of Labor Statistics puts median pay for receptionists at $38,010 a year, or $18.27 an hour, across 947,500 jobs in 2025, with employment projected to fall 2 percent through 2035 and about 105,100 openings a year from turnover. BLS names the reason for the decline plainly: organizations keep automating or consolidating administrative work.

Run the coverage math out loud. A full time hire covers 40 of the 168 hours in a week. That is under a quarter of the week, before vacation, sick days and lunch. The calls that actually go unanswered at a service business are mostly the ones outside those 40 hours, which is the exact window the hire does not cover and the AI does.

The honest version includes the other side. A receptionist who has been there four years knows the customers, catches the thing nobody wrote down, and handles the walk-in. If your prospect's front desk person is good, do not pitch against her. Pitch the nights, the weekends and the overflow.

Who this comparison does not help

Three cases where I would not run this pitch.

A business doing fifteen calls a month does not have a phone problem worth a monthly system. Forwarding to a cell is fine, and selling them anyway is how you earn a cancellation in month three.

A business where every call is genuinely consultative, with custom scoping and real negotiation, gets a worse experience from an agent answering first. Some commercial and specialty work is like this. The answering service, or a person, should stay in front.

And a business that will not give you hours, services, price bands and a calendar cannot be built for. If they will not hand over inputs, the agent will improvise, and an improvising agent on a live line is a liability. That is why we gate the build on intake rather than hoping it shows up later, and why the agent should be tested against real call recordings before it takes a single live call.

How to price it against their current bill

Ask for the last two answering service invoices before you quote anything. You are looking for the monthly total, the included block, the overage line and whether they blow through the block in busy months. That total is your anchor.

A client doing $720 a month with Ruby and complaining about message-taking is not comparing you to $49. They are comparing you to $720 plus the jobs they believe they lost. You price against that number, not against your wholesale cost. On our side a top tier build runs $1,150 to $1,400 setup and $315 to $460 a month plus $0.35 a minute on voice, agreed on a short call. You set retail. The pricing page has the platform limits and the tiers in full.

One more mechanic that helps here. Because there is no build fee on the platform side, you can put a working demo in a prospect's hands before anybody signs anything, and the demo answers the "will it sound like my business" question faster than any slide. Build it in their trade, use their business name, and let them talk to it themselves.

FAQ

Can a client keep their answering service and add an AI receptionist? Yes, and that is usually the right first configuration. The AI takes every call first and escalates to the answering service after hours or on trigger conditions. Their answering service minutes drop because the AI absorbs the short repetitive calls, and nothing has to be cancelled for the client to see a result in month one.

Do we have to tell callers it is an AI? Depends on the state and the use case, and the rules move. We wrote up where disclosure rules stand for voice agents, and the practical answer is that a brief natural identification at the start costs you almost nothing in booking rate and removes a category of risk. Published Lumina demos always carry a sandbox disclosure, which is the one piece of branding your prospect will see.

What happens when the AI gets a name or address wrong? Design for it. Confirm the spelling back on the call, send a text confirmation the caller can reply to, and route any booking with a failed confirmation to a human for review. The failure mode you want is a flagged booking, not a silently wrong one.

How many minutes will a small service business actually use? It varies more than any vendor average will tell you, so measure instead of guessing. Pull their answering service invoice for the included block and the overage line and you have a real baseline for their business. Note that the $49 platform plan includes 100 live voice minutes a month for your own demos, which is a separate budget from client builds.

Is the AI cheaper than an answering service for everyone? On cost per minute, yes, by a wide margin at published rates. On value, no. A business with low call volume, or one whose calls are all complex, is better served by a person. Say that out loud and you will close more of the deals where the answer is actually yes.

Your next step

Pick one prospect who is already paying an answering service. Ask them for last month's invoice on the pretext of benchmarking, which is true. Work out their effective per-minute or per-call rate, find the short repetitive calls in it, and build a demo in their trade with their business name on it. Then get them to call it themselves at 9pm.

The pitch writes itself at that point. They are paying $10 a call for the questions a machine can answer in twelve seconds, and the machine can hand the rest to the humans they already trust. Start your workspace and have the demo live before the call.