Monday, 7:41. Luis is under a kitchen sink with a flashlight in his teeth. The shop line is on speaker in the van. A homeowner with water on the linoleum hears the recording he cut in 2019: thank you for calling, press 1 for service, press 2 for billing, press 3 for our hours. She does not press anything. She hangs up and taps the next listing.
Same minute, another homeowner is on the website. The bubble offers three buttons - New service, Existing customer, Hours - then a form. Name, phone, how can we help, submit. None of the buttons say water on the floor. She closes the tab and calls the company that answered.
Two doors. One recorded tree, one form that cannot act. Both were built to sort a person into a pile someone else would handle. Neither of them handled her.
The phone tree was built to sort calls, and it now sorts callers away. The contact form, and the menu-bot on the site, do the same job with a submit button. An AI receptionist is a different object: it answers, it acts, and it hands off with the work already done.
What the menu was built to do
An IVR is a sorter. Press 1 for sales is not a conversation. It asks the caller to file themselves, then waits until a person matches the folder. The website copy is the contact form, and the widget that is only a menu in a speech bubble. Pick a department. Fill the fields. Wait. The form does not know your price book, cannot see an open slot, and cannot tell a leak from a billing question unless the visitor already knew which box to tick.
Call-center math often never sees that sort. Call abandonment rate, as Geckoboard defines it, is the share of inbound calls where the customer hangs up before an agent answers. A rate around 5% is commonly called acceptable, and Geckoboard, citing HubSpot, puts typical contact-center abandonment between 5 and 8%. The line most owners skip: callers who drop while they are still in a self-service menu often do not count toward that rate at all. They count against the menu, if anyone is counting the menu. A form start that never submits is the same invisible exit.
On 1.8 million calls through VeloCalls' own IVR flows, January through May 2026, abandonment before an agent tracked the depth of the tree. One menu: 6.1%. Two: 12.4%. Three: 18.7%. Four or more: 27.3%. Their book is pay-per-call operators in home services, legal intake, and insurance - not a bank, and not your shop. One menu is a question. Three menus is your org chart, read aloud. Every extra required field on the form is another level nobody asked for.
If the phone rings out before anyone speaks, that is a different leak - the one-ring standard, and what happens after the beep. This is the other one: a machine greeted them.
Three exits the log already has
Export one week. Tag every call that hit the tree, and every form start or widget session, with one of three exits.
- Abandon. They entered and left before a person, a booking, or a usable message. A hang-up between options. A form start with no submit. A widget that dies after "I didn't understand." Your abandonment rate may not count these, because no agent was offered.
- Wrong menu. They pressed something, or picked a subject, and landed in the wrong pile. Press 2 for billing when the garage is flooding. "Existing customer" when they needed a same-day slot and have never been one.
- Zero. Nothing was captured. No keypress, or they hit 0 for a person and nobody was free - the old escape, which only works if a human is actually there. A first name and a blank problem. A widget that collected "hi" and closed.
Add the three. That sum is the week the menu sorted away. A call answered by a tree is not an answered call. A submitted form is not a booked job. For dollars, multiply abandon plus zero by your average ticket and your usual close rate.
An AI receptionist is a different object
A shorter greeting is still a sorter. So is a speech tree that maps "billing" onto the same dead branch as press 2. Replacing the IVR with a nicer recording does not change the object. It changes when the thing that picks up can finish a piece of the job before it needs a person.
On the phone, that is a voice agent. It answers inbound, and it can dial outbound when the job is a reminder or a follow-up. It picks up on the first ring. The caller says the water is on the floor. The agent asks what your dispatcher would ask, qualifies the job, books a real slot or follows the on-call rule you wrote, and if a person has to take it, makes a warm transfer or leaves a message with the intake already done. Voice runs in 25+ languages, not a second tree of language buttons.
On the site, the same job is a chat agent, not a second phone tree in a bubble. It lives on the website, and on WhatsApp, email, and in-app, from a one-line embed. The visitor types the problem. The agent answers from your knowledge, books or opens a ticket, and hands off with the transcript attached. Chat covers 100+ languages. The phone and the website do not share one language limit.
Those surfaces are voice agents on the line and chat agents on the pages people already have open. Answer. Act. Hand off. Keep a tree only where a tree still fits - a recorded status, or a keypress an audit must replay.
The one-week audit, on both logs
Pick seven days with a weekday stack and a weekend. A quiet Tuesday hides the exits that happen when nobody is at the counter. Put both logs on one sheet and write five counts.
- Phone, tree hits. Calls that reached the greeting, split into abandon, wrong menu, and zero. If the system only shows "answered" and "missed," tag fifty recordings, or the option-selected field, by hand.
- Phone, false success. Of the calls that pressed a digit and connected, how many still needed a person to ask what the menu was supposed to settle? Count how many pressed 0 and then sat in a queue the tree had already delayed.
- Site, form and widget. Same split. A start with no finish is abandon. A subject your team had to recategorize is wrong menu. A close with nothing a dispatcher can use is zero. A button menu that apologizes is abandon, not containment.
- Overlap. A phone call from someone already on the site, or a form that arrived after they hung up on the tree. A note that says "I tried calling" is enough.
- Worked pile. Of everything left, how many ended as a booked job, a quoted fee, or a lead with a name, a number, an address, and a symptom? A transfer without the facts is still a sort.
If abandon plus wrong menu plus zero is most of the week, the tree and the form are doing what they were built to do.
Start small, measure it
You do not have to rip the phone system out on a Thursday. Take the exit that dominated the sheet - the after-hours service branch, or the form that only keeps a first name - and put an agent on that surface for the next week. Leave the rest of the menu alone.
Then run the same counts again, on the phone log and on the site log. Either those three moved, or they did not. A voice agent on the line and a chat agent on the page replace the sorter. If they book, qualify, and hand off with the message or the transcript already filled, the sheet shows finished work instead of a digit.
Verlingo runs both surfaces - the number and the site - so the week can move from the tree to a thing that acts. Start with the exit that hurt. Measure the next seven days against this one. The menu already told you what it was built to do. The labels are how you decide whether to keep paying it to do that.