Count Your True Missed Calls First, Then Price AI Phone Answering Against Them

If your phone rings 220 times a month and you answer 172, the 48 you didn't answer might be worth more than an AI answering setup costs — or they might be worth almost nothing. You only find out by doing the arithmetic before you buy anything.

This post is the arithmetic, plus the three call types we'd never route to a bot.

Start with the only number that matters: true misses

Not "missed calls." True misses. A true miss is a unique phone number that rang you, didn't reach a human, and never connected within 24 hours. Everything else — the customer who called back two minutes later, your supplier, the SEO spam from a 415 number — is noise that makes the problem look bigger than it is.

Block 30 minutes and pull the last 90 days from your phone system. You want five columns:

1. Total inbound calls
2. Unanswered (no answer, sent to voicemail, or abandoned in queue)
3. After-hours unanswered
4. Repeat attempts from the same number within 10 minutes (that's one person, not three calls)
5. Unique numbers with no outbound call back from you within 24 hours — your true misses

If you're running off cell phones with no system, you can get most of this from your carrier's call log, or put a single call-tracking number on your Google Business Profile and your website for 30 days and measure from there.

Opinion, clearly labeled as opinion because we won't quote a statistic we can't link: we think most owners underestimate this number, because voicemail hides the evidence. The caller who didn't leave a message doesn't appear in any report you own. Your own 90-day pull is the only figure worth trusting here — including ours.

A worked example (hypothetical — every number is a placeholder)

The following is a made-up three-truck HVAC company. Not a client, not a case study, not benchmark data. The assumptions below are invented for the sake of showing the shape of the calculation. Overwrite all of them with your own.

• 220 inbound calls/month
• 48 unanswered, of which 31 are true misses
• Assume 40% of true misses are new customers rather than existing ones chasing a part → 12 new-customer misses (you'll get your real split from step 5 above plus your CRM)
• Assumed close rate on answered new-customer calls: 45%
• Assumed average first job value: $420

12 × 0.45 × $420 = $2,268/month of first-job revenue in this made-up scenario. Before repeat work and referrals.

Now be pessimistic about the fix. No answering system converts every caller — assume some hang up on a machine. Pick a recovery rate you believe; at 60% this example yields $1,361/month. That 60% is a judgment call, not a measured figure, which is exactly why the scoreboard section below exists.

Against that, the cost side has three lines:

• Usage. Voice platforms in this category publish per-minute rates, and they change. Read the live page in the month you buy and write the date next to the number: Vapi and Retell AI both publish pricing. As illustrative arithmetic with an invented rate — not a quote from either vendor — 60 calls a month at 2.5 minutes each is 150 minutes; at a made-up $0.10/minute that's $15.
• Platform/phone number fee. A fixed monthly line. Get it in writing, dated.
• Build. One-time configuration, integrations, and testing.

Run those three lines against your own recovered-revenue figure. The point of the exercise isn't to decide whether AI is good; it's to see whether the recovered revenue clears the monthly cost with room to spare. If your true-miss number is 3 instead of 31, the honest answer is to skip this entirely and turn on a better voicemail-to-text.

What the system has to actually do to earn its keep

A bot that takes a message is an answering machine with a subscription. Our bar is higher:

• Answers inside two rings, every hour of every day. Overflow and after-hours, not one or the other.
• Sounds like a person and handles interruptions. Latency is the tell. Our view, from testing these systems: if there's a beat of silence after every sentence, callers notice, and some hang up. Make the vendor demo it on a live call to your cell, not in a recorded video.
• Writes to your calendar or CRM. Booking a slot, not promising someone will call back.
• Texts the caller a confirmation with the appointment time and a reschedule link, sent before they've put the phone down.
• Drops a transcript, recording, and one-paragraph summary into your email or Slack within a minute of the call ending.
• Transfers to a human on rules you wrote, with a fallback if that human doesn't pick up.

Ask your vendor in writing how recordings, consent notices, and data retention are handled, and check your own state's call-recording consent rules with your attorney before you go live.

The three call types you should never hand to a bot

1. Emergencies and safety. Gas smell, water coming through a ceiling, no heat with a newborn in the house, chest pain, a chemical burn in the salon chair. Build intent and keyword detection that fires an immediate warm transfer to the on-call mobile, rings a second number if the first doesn't answer in 20 seconds, and never — ever — puts that caller in a queue or takes a message. If you can't get this rule working in testing, don't launch.

2. Money in dispute. Refunds, billing complaints, chargebacks, the angry callback about yesterday's job. A bot negotiating a refund is a one-star review with extra steps. Route these to a human every time; the bot's only job is to identify them fast and hand off politely.

3. Anything clinical or regulated. Symptom triage, medication questions, "should I come in?", dosage, treatment advice. For clinics there's a second layer: a vendor that creates, receives, maintains, or transmits protected health information on your behalf is generally a business associate under HIPAA, which means a signed business associate agreement — see HHS's sample business associate agreement provisions and get the paperwork done before the first call, not after. Confirm your specific situation with counsel. A bot can book, confirm, reschedule, and give directions and hours. It should not discuss a patient's condition.

One bonus rule that costs nothing: any caller who asks for a human gets one, the first time they ask, without a script that tries to talk them out of it.

A one-week rollout

• Day 1: Pull 90 days of call data. Calculate true misses and your after-hours share.
• Day 2: Write the three escalation rules above and your top 10 FAQs — hours, service area, pricing ranges you're willing to say out loud, parking, insurance accepted.
• Day 3: Pick a vendor. Get the per-minute rate and the monthly fee in writing, dated.
• Day 4: Build it on a forwarding number only. Your main line still rings your team first; the AI catches overflow and after-hours.
• Day 5: Test 20 scenarios out loud, including all three no-go types and one caller who mumbles.
• Weeks 2–3: After-hours only. Then add daytime overflow.

The scoreboard, and the kill switch

Track five things weekly: answer rate, bookings from new-customer calls, escalation rate, how long escalations take to reach a human, and callback-within-24h on anything the bot couldn't finish.

Set the kill criterion up front: if after-hours new-customer bookings don't beat what voicemail was producing within 30 days, switch it off. A tool that doesn't clear that bar isn't worth the review risk.

What we'd build

For a single-location shop, clinic, or salon this is a small build, not a platform project: an off-the-shelf voice platform, one calendar or CRM integration, an escalation tree with real phone numbers behind it, and a nightly transcript digest so you can hear what your callers actually ask for. Our rough estimate for that scope — opinion, not a quote — is 10 to 20 hours of configuration and testing. Ongoing cost is mostly usage, which means it scales with the calls it's saving rather than with your ambitions.

The expensive part was never the software. In the example above, it was the 31 people a month who rang, got a beep, and called the next name on the list. Go find out what your number is.

---

Run your own numbers first. The worksheet above tells you what your true misses are worth. When you want the wider view — where automation would actually pay off across your operation, not just the phone — our free AI Readiness & ROI Calculator is a five-minute assessment that projects ROI over 24 months.