AI answering versus voicemail versus an answering service

That is the whole decision, and most businesses get it wrong by comparing monthly prices instead of comparing what happens to a caller at 6:40pm on a Friday.

Here is the sharper way to frame it. A missed call is not a neutral event. The caller either dials a competitor or stops looking. If your average job is worth a few hundred dollars, one recovered call a week pays for almost any of these options. If your average job is worth twenty dollars, none of them make sense.

What does each option actually do to a caller?

The two columns that matter are "handles the unexpected" and "consistency." A human agent will improvise when a caller says something the script did not anticipate. That improvisation is sometimes exactly what you needed and sometimes a promise you now have to honor.

An AI does the opposite. It never improvises well, and it never freelances a price either. You are choosing between a system that will occasionally say the wrong thing creatively and a system that will occasionally say the wrong thing identically, ten thousand times.

Is voicemail really that bad?

It is worse than owners think, and better than the industry admits.

Worse, because a first time caller who has three tabs open does not leave a message. They dial the next result. You never learn that the call happened in any way you can act on.

Better, because for an existing customer base that knows you, voicemail is fine. A regular leaving a message about a delivery date is not a lead at risk.

The honest version is that voicemail fails on new business and works on repeat business. If your inbound is mostly people who already know your name, you have a smaller problem than the vendors selling you a fix would like.

What voicemail never gives you is a count, and the spread across industries is enormous. CallRail's January 2025 benchmark, built on analysis of 1.1 million de-identified conversations, put missed call rates at 32% in health care, 28% in legal, 14% in home services and 9% in real estate. CallRail's definition is narrow and worth knowing: a missed call is an unanswered inbound call that hits a detected voicemail or is detected as abandoned.

Inside a single vertical the spread is just as wide. Revmo AI, working from 12,091 real restaurant call recordings, published answer rates of 59.9% for quick service, 75.3% for fast casual, 91.0% for full service and 93.1% for pizza. "Restaurants miss X% of calls" is not a meaningful sentence. The segment decides the number, not the industry.

Yours is the only figure that matters, and you cannot get it from a benchmark.

When is a live answering service the right call?

When the calls require judgment, empathy, or a decision your script cannot pre approve.

Medical triage. Property emergencies with liability attached. Anything where the caller is upset and the outcome depends on how they are handled in the first fifteen seconds.

There is a hard line underneath that. The Doctors Company, the largest physician-owned malpractice insurer in the US, states in its telephone triage guidance that unlicensed personnel should never be allowed to triage by telephone or provide telephone advice, and that failing to follow written advice protocols may be treated as practising medicine or dentistry beyond the staff member's scope. The same guidance notes that miscommunication contributes to over 30 percent of adverse patient events in the office setting. That standard binds an answering service exactly as it binds software.

Also choose a service when your volume is low and irregular. Paying per call at ten calls a month is cheap, and building and maintaining a script is real work that will not earn back at that volume.

The tradeoff is that a service is only as good as the agent who picks up, and you do not choose the agent. Quality drifts. Your intake questions get shortened. You find out three months later that nobody has been asking for the address.

When does AI answering win?

When your calls are repetitive, your answers are knowable in advance, and volume is high enough that consistency is worth more than improvisation.

Booking, hours, service area, pricing ranges you are willing to publish, intake for a quote. All of that is scriptable, and a script that is right is right every single time.

The real advantage is calendar access. An answering service takes a message that somebody on your team has to act on tomorrow. An AI can write the appointment straight into the calendar you already use, which removes the step where leads die.

The real risk is deployment. Coval, a vendor selling evaluation tooling for voice agents, reports that roughly 95% of voice agents work in a demo and about 62% survive their first week live. Treat that as a vendor figure, but the direction is right and matches what I see. Most of the failures come from the same handful of causes, and all of them are avoidable before launch.

The second risk is latency, and it is invisible in a demo. Human conversation runs on a turn transition gap of roughly 200 milliseconds, measured across ten languages by Stivers and colleagues in a 2009 study in the Proceedings of the National Academy of Sciences. Retell publishes a target end-to-end latency of 800 milliseconds at the median and 1,200 at the ninetieth percentile. Cross-vendor benchmarking by Cekura puts real-world median turn latency at 1.5 to 2.5 seconds, which is roughly eight to twelve times the human baseline. That gap is what a caller hears as "something is off," and it is a stack decision, not a script decision.

What does the arithmetic look like at your volume?

Run it on cost per recovered call, not on monthly price. Here is the walk-through.

Start with the calls you are actually losing. Say your log shows 300 inbound calls a month and 45 of them went unanswered. That is a 15% miss rate, between CallRail's home services and legal figures, so it is a plausible starting point rather than a fantasy.

Apply a booking rate. ServiceTitan's platform data from June 2022 gives the best public anchor: a typical shop books 42% of calls, a shop with fewer than five technicians books 24%, and a shop with 25 or more books 59%. Use the small-shop figure. 45 × 0.24 = 10.8 jobs a month you are not getting.

Apply your own average ticket. At $400 that is $4,320 a month, or roughly $51,800 a year, before any discount for callers who reached you later anyway. Halve it and you still have about $25,900.

Now price the three options against that. Voicemail costs nothing and recovers none of it. An answering service billing per minute at 45 recovered calls of three minutes each is a small monthly number, and it books only what your script and the agent's attention allow. An AI at a platform fee plus usage sits in between, plus buildout.

The decision rule. Divide the annual cost of each option by the jobs it plausibly recovers. If any option costs more per recovered job than the job is worth, it is off the list. If two options survive, pick the one with the better failure mode for your industry, not the cheaper one.

That same ServiceTitan analysis notes that every 5% improvement in booking rate is worth roughly $100,000 in additional revenue for a shop with 5 to 14 technicians, achievable with less than one additional call booked per weekday. The gradient tells you what this actually is. A sub-five-technician shop books 24% and a 25-plus shop books 59%. Missed calls are a capacity problem, not a discipline problem. The small operator is on a roof or under a sink.

What about containment rates and the numbers in the pitch?

Be careful with them, because the benchmark you are being shown probably does not exist.

Vendors routinely quote a "typical containment rate" attributed to Gartner, Forrester or Deloitte. No credible containment-rate benchmark from any of those firms could be verified. Their research pages are gated and returned nothing extractable when checked. If a proposal cites one, ask for the report name, the publication date and the page number. That request ends the conversation more often than it produces a citation.

Containment is also the single most gameable metric in this category, because you improve it by making escalation harder. A system that will not transfer you looks excellent on a containment dashboard. Always read it next to escalation rate and callback volume.

What about regulated industries?

Then the comparison changes, because some answers are off limits regardless of who or what is speaking.

A dental practice can use an AI to confirm in network status and collect plan details. It must never tell a caller a procedure is covered or quote an out of pocket amount, because an AI is unlicensed personnel operating a provider approved decision tree. The full boundary for dental phone handling is narrower than most vendors will tell you, and your counsel should sign off on the script.

The volume behind that constraint is real. Patient Prism scored 11,552,668 dental calls across 8,280 locations in calendar 2025 and found that 31 of every 100 calls were abandoned before reaching an agent. That is a queueing problem, which automation solves well. The topic those callers most want to discuss is insurance, which is the one topic the automation is least allowed to speak to.

That constraint applies to human answering services too. The difference is that an AI enforces it perfectly and a human agent under pressure sometimes does not.

What to do this week

Pull last month's phone bill or call log. Count total inbound calls, then count how many were answered by a person.

Run the five step arithmetic above. That figure is your budget ceiling for solving this. If it is smaller than the cheapest option, stop here.

Ask every vendor for the source of any benchmark in their deck. Report name, date, page.

Then look at what a caller currently sees before they dial. The wording of the call to action they clicked sets the expectation for the call, and mismatched expectations produce the calls that go badly.

If you are a home services business, it is worth knowing where after hours calls currently end up before you decide who should be answering them.

Questions about which of the three fits you?

Email me at eric@seod.com with what you pay your answering service today and what it covers, or a screenshot of the quote you were given. I will tell you which parts of that scope an AI would genuinely handle, which parts it would handle badly, and which parts you are paying for twice.

I answer these myself, usually the same day, and there is no sequence attached.

The rest of the AI phone and lead response library covers the pieces around it.