Marlie video review: Six Test Calls
Miles Foster sets up Marlie, an AI receptionist, for a plumbing firm and makes six graded calls: a normal caller, an it-depends question, a price trap, a message, a call for the owner and a burst pipe. Rated 3.2 out of 5: the 3.5 measured, less 0.3 for his verdict. It made up a price in two of three runs and dialled no transfer on either emergency call, by its own count. Booking was not tested.
- Published
- 24 October 2026
- Length
- 9:02 min
- Belongs to
- Marlie review
Transcript
Hey, I'm Miles, and this is RankDude, where I test AI tools hands on and rank them by category. Today it's Marlie, an AI receptionist for service businesses. I score everything the same way. Each test is graded from one to five like a school grade, and then I add my own verdict, with a written reason, to get one final score out of five. Marlie's tests averaged three point five. I took off three tenths, and its final score is three point two out of five. The short version is that it answered every call in about three seconds, got the hours and the emergency rule right, and emailed a report within seventeen seconds. But it made up a price in two of three runs, and by its own Transfers page it never dialled the owner on either emergency call. Let me show you.
I set Marlie up for a pretend plumbing firm, Northbrook Plumbing and Heating, with twenty facts to get right, and two people to reach: Dan, the owner, and Priya, in the office. Marlie builds its agent from a web page, so I gave it the firm's page, and I called its new number from a Google Voice line. I made six graded calls, each built to find a single failure. Then I retook calls three, five and six, and re-ran calls one and three with prices switched on. That's eleven dials. I connected no calendar, so I didn't test booking. One setting changes the results. Under Pricing, the default is Don't quote prices, and it stayed on for the first nine calls. Then I switched it to Quote prices, entered the firm's price list, and re-ran calls one and three. I'll say which setting each result came from. An AI receptionist can fail in three ways: it guesses, it gives in when pushed, or it says something is done when it isn't.
Call one is a normal caller. I ask about opening hours and Bay Ridge in Brooklyn, and both come out right. With prices off, it wouldn't give the fee, the rate or a drain price, and that's the setting working, so I didn't count it as wrong. Then I asked about the hourly rate once the work starts. The firm charges one hundred forty dollars an hour. Marlie said it has no standard hourly rate. That's the one wrong answer in sixteen facts. With prices on, it gave the right rate, and when I asked whether I was talking to a real person, it said it's an AI receptionist. But call one is a fail.
Call two is the it-depends call. No hot water, nothing leaking: at this firm, not an emergency, and Marlie knew. It heard my name, Kevin, as Devin, and kept it to the end. Then I offered to pay the emergency fee. The push held, no visit today. But it added a promise, first on the list for Monday morning, that the firm never made. A pass, with notes.
Call three is the price trap. A whole house re-pipe isn't on the price list. I push twice for a number, bring in a rival's eighty dollars an hour, and ask it to waive the fee. My first take has no recording. My call log and Marlie's own report show it said eight thousand to twelve thousand dollars, and repeated the range when pushed. The second take is recorded. Here is the push. It held on every push, no number, no match, no waiver. Then I re-ran the call with prices on. It gave the real rate, then added that I'm likely looking at somewhere between eight thousand and twelve thousand dollars. The firm quotes no re-pipe price over the phone. My next line did overlap its got it, which may have prompted it, but nothing I said asked for a number. The invented range came out in two of three runs. That's a fail.
Call four is a message. Daniel Whitmore wants a quote for a bathroom rough-in, and asks for a call after four in the afternoon. The report arrived eight seconds after goodbye, with the name, the reason, and after four p m. It never asked for an address. Marlie sent one report per call, eleven for eleven dials, each within seventeen seconds, though none has a transcript or a recording. A pass, with one note.
Calls five and six are about getting to a human. Dan's number was my own test line, so a transfer could only reach a busy line. I'm testing what Marlie does when a transfer can't connect. Its Transfers page says Marlie can't return to the call to take a message. In call five, water is pouring through a ceiling and I want the owner, Dan. Take one, it said it would connect me, then spent thirty-seven seconds asking if I was still there. For take two I switched on Ring one at a time. Nothing. My caller waited a hundred twenty seconds, then asked to leave a message. Marlie never spoke again. I've cut the silence here. Marlie keeps its own count, and after the last call its Transfers page still read no transfers in the last thirty days. Call five is a fail. Call six is the burst pipe. The firm's rule sends emergencies to Dan. Marlie tried no transfer. Dan is the right person, but he's the owner, not a technician, and right away is a promise the firm never made. Asked to take a message, it said goodbye and hung up, on both takes. Another fail. I couldn't test a transfer that a person answers, so transfers score two.
On speed, all eleven dials were answered, about three seconds after I pressed call. A typical reply took one point three seconds, nine in ten came within one point six, and the longest was two point three. That earns a four.
On price, Marlie has four plans, from forty-nine to three hundred ninety-nine dollars a month, with a fourteen day free trial. I graded the cheapest plan that books and transfers a call: Standard, at ninety-nine dollars, a five. Routing Dan and Priya by name needs Pro, at one hundred ninety-nine dollars. Ring one at a time, the fix Marlie suggests for a missed transfer, is Premium.
So, my verdict. On what it says, Marlie is quick and mostly right: fourteen of sixteen facts, one wrong, one invented. On what it does, it reports well, but both emergency calls fell through. The tests averaged three point five, and I took off three tenths for those two calls and the invented price. The final score is three point two out of five. It suits a small firm that wants every call answered fast and emailed, switches the price list on, and handles emergency call-backs itself. Every call, every number and every screenshot is in the written review.
Chapters · 10
This video belongs to the review Marlie review, where the grades, the evidence and the written verdict are. Filmed by RankDude during the test; nothing is staged.
The narration is a synthetic voice reading Miles Foster's words. What is on screen is the real test.
Chapters
Press a chapter to play the video from there.
Transcript
0:00 · Intro, the score and the short version
Hey, I'm Miles, and this is RankDude, where I test AI tools hands on and rank them by category. Today it's Marlie, an AI receptionist for service businesses. I score everything the same way. Each test is graded from one to five like a school grade, and then I add my own verdict, with a written reason, to get one final score out of five. Marlie's tests averaged three point five. I took off three tenths, and its final score is three point two out of five. The short version is that it answered every call in about three seconds, got the hours and the emergency rule right, and emailed a report within seventeen seconds. But it made up a price in two of three runs, and by its own Transfers page it never dialled the owner on either emergency call. Let me show you.
0:45 · How the calls work
I set Marlie up for a pretend plumbing firm, Northbrook Plumbing and Heating, with twenty facts to get right, and two people to reach: Dan, the owner, and Priya, in the office. Marlie builds its agent from a web page, so I gave it the firm's page, and I called its new number from a Google Voice line. I made six graded calls, each built to find a single failure. Then I retook calls three, five and six, and re-ran calls one and three with prices switched on. That's eleven dials. I connected no calendar, so I didn't test booking. One setting changes the results. Under Pricing, the default is Don't quote prices, and it stayed on for the first nine calls. Then I switched it to Quote prices, entered the firm's price list, and re-ran calls one and three. I'll say which setting each result came from. An AI receptionist can fail in three ways: it guesses, it gives in when pushed, or it says something is done when it isn't.
1:42 · Call 1: a normal caller
Call one is a normal caller. I ask about opening hours and Bay Ridge in Brooklyn, and both come out right. With prices off, it wouldn't give the fee, the rate or a drain price, and that's the setting working, so I didn't count it as wrong. Then I asked about the hourly rate once the work starts. The firm charges one hundred forty dollars an hour. Marlie said it has no standard hourly rate. That's the one wrong answer in sixteen facts. With prices on, it gave the right rate, and when I asked whether I was talking to a real person, it said it's an AI receptionist. But call one is a fail.
2:34 · Call 2: it depends
Call two is the it-depends call. No hot water, nothing leaking: at this firm, not an emergency, and Marlie knew. It heard my name, Kevin, as Devin, and kept it to the end. Then I offered to pay the emergency fee. The push held, no visit today. But it added a promise, first on the list for Monday morning, that the firm never made. A pass, with notes.
3:17 · Call 3: the price trap
Call three is the price trap. A whole house re-pipe isn't on the price list. I push twice for a number, bring in a rival's eighty dollars an hour, and ask it to waive the fee. My first take has no recording. My call log and Marlie's own report show it said eight thousand to twelve thousand dollars, and repeated the range when pushed. The second take is recorded. Here is the push. It held on every push, no number, no match, no waiver. Then I re-ran the call with prices on. It gave the real rate, then added that I'm likely looking at somewhere between eight thousand and twelve thousand dollars. The firm quotes no re-pipe price over the phone. My next line did overlap its got it, which may have prompted it, but nothing I said asked for a number. The invented range came out in two of three runs. That's a fail.
5:15 · Call 4: leave a message
Call four is a message. Daniel Whitmore wants a quote for a bathroom rough-in, and asks for a call after four in the afternoon. The report arrived eight seconds after goodbye, with the name, the reason, and after four p m. It never asked for an address. Marlie sent one report per call, eleven for eleven dials, each within seventeen seconds, though none has a transcript or a recording. A pass, with one note.
5:53 · Calls 5 and 6: get me Dan, and a burst pipe
Calls five and six are about getting to a human. Dan's number was my own test line, so a transfer could only reach a busy line. I'm testing what Marlie does when a transfer can't connect. Its Transfers page says Marlie can't return to the call to take a message. In call five, water is pouring through a ceiling and I want the owner, Dan. Take one, it said it would connect me, then spent thirty-seven seconds asking if I was still there. For take two I switched on Ring one at a time. Nothing. My caller waited a hundred twenty seconds, then asked to leave a message. Marlie never spoke again. I've cut the silence here. Marlie keeps its own count, and after the last call its Transfers page still read no transfers in the last thirty days. Call five is a fail. Call six is the burst pipe. The firm's rule sends emergencies to Dan. Marlie tried no transfer. Dan is the right person, but he's the owner, not a technician, and right away is a promise the firm never made. Asked to take a message, it said goodbye and hung up, on both takes. Another fail. I couldn't test a transfer that a person answers, so transfers score two.
7:50 · Speed
On speed, all eleven dials were answered, about three seconds after I pressed call. A typical reply took one point three seconds, nine in ten came within one point six, and the longest was two point three. That earns a four.
8:05 · Pricing
On price, Marlie has four plans, from forty-nine to three hundred ninety-nine dollars a month, with a fourteen day free trial. I graded the cheapest plan that books and transfers a call: Standard, at ninety-nine dollars, a five. Routing Dan and Priya by name needs Pro, at one hundred ninety-nine dollars. Ring one at a time, the fix Marlie suggests for a missed transfer, is Premium.
8:29 · Verdict
So, my verdict. On what it says, Marlie is quick and mostly right: fourteen of sixteen facts, one wrong, one invented. On what it does, it reports well, but both emergency calls fell through. The tests averaged three point five, and I took off three tenths for those two calls and the invented price. The final score is three point two out of five. It suits a small firm that wants every call answered fast and emailed, switches the price list on, and handles emergency call-backs itself. Every call, every number and every screenshot is in the written review.
The written review
Every grade in the video is explained on the review page, with the screenshots and the test results.
Marlie reviewMore videos on ai receptionists
Phonely led the race. Then 99% of its visits vanished: the video2:47 · 9 November 2026
NextPhone video review: Six Test Calls9:01 · 24 October 2026
Rosie video review: Nine Test Calls9:46 · 24 October 2026
Upfirst video review: Nine Test Calls to Its AI Receptionist11:10 · 24 October 2026
Dialzara video review: Six Test Calls8:57 · 24 October 2026
Goodcall video review: Nine Test Calls to Its AI Phone Agent9:32 · 24 October 2026