The Dialer That Calls First
Steal Like an Engineer #2
Our calling team has 10 to 15 people. Before the dialer, the whole team got about 60 to 70 calls a week where someone picked up.
Now the same team has 130 to 150 real conversations a week.
I’m a backend engineer and team lead at Caprae Capital, an M&A firm that helps buyers find small businesses to buy. I built the dialer. This post covers what it does, the two mistakes in its first version, and a small piece of code you can reuse.
The problem: half the list goes nowhere
A cold-call list looks like a full day of work. A lot of it is empty. Kevin Hong, our CEO, has written publicly that about 20% of the numbers on a list are dead and another 30% go straight to voicemail.
That means a caller can spend half the day on rings, voicemail greetings and “this number is not in service” before a single owner says hello.
What the dialer does
Our callers work in one-hour blocks. Before a block starts, the whole list, about 1,000 numbers, goes through the dialer.
The dialer calls every number first. It listens to the first few seconds of each call and puts the number in one group:
| Group | What happens next |
|---|---|
| A person answered | Goes to a caller |
| Switchboard or front desk | Goes to a caller, who asks for the owner |
| Voicemail | Skipped for now |
| Busy or no answer | Tried again another time |
| Fax, dead or out of service | Removed from the list |
When the block starts, the callers only get numbers where a person or a front desk picked up. The hour goes to conversations.
What broke first: switchboards filed as voicemail
The first version trusted its first guess. The quick listen is good at one question: did a person answer, or a machine?
A switchboard is a machine too. “Thank you for calling, press 1 for sales” sounded like voicemail to the first listen, so it went in the voicemail pile. But a switchboard leads to a business with a front desk, and someone there can put you through to the owner.
The fix was a second listen. Every number marked voicemail, and every number the first listen wasn’t sure about, gets checked again on the full recording. If it hears “press” followed by a number, or the word “menu”, it marks the number as a switchboard.
Only the unsure numbers get the second listen, so it stays cheap.
What broke second: results on the wrong rows
The dialer writes its result next to each number in the sheet the team uploaded.
I first matched each result to its row by row number. That works when the sheet’s header is on row 1. Real sheets don’t always start there. Some begin with a title, a date or a blank line, and every result after that landed on the wrong row.
Now the dialer matches on the phone number itself. It strips the number to digits, tries it with and without the country code, and uses the row number only when a row has no phone number.
Try it: match on the number, not the row
A small, generic version of the fix, in Python:
import re
def digits(phone: str) -> str:
return re.sub(r"\D", "", phone or "")
def variants(phone: str) -> list[str]:
d = digits(phone)
# US numbers: try with and without the leading country code 1
return [d, d[1:]] if len(d) == 11 and d.startswith("1") else [d, "1" + d]
def label_for(row_phone: str, row_index: int, by_phone: dict, by_row: dict) -> str | None:
for v in variants(row_phone):
if v in by_phone:
return by_phone[v] # primary: the number itself
return by_row.get(row_index) # fallback: the old row position
Build by_phone from your results, keyed by digits. Keep by_row only for rows with no phone number.
What changed for the team
- Before: about 60 to 70 answered calls a week, across 10 to 15 callers.
- After about three months with the dialer: 130 to 150 real conversations a week, with the same team.
The callers work the same hours as before. The dialer removes the numbers that were never going to answer, so those hours go to people.
Whose phone is it?
The second listen does one more job: it checks whose phone a number is.
A voicemail greeting often says a name (“Hi, you’ve reached Mike”), and a person who picks up often says it too (“This is Mike”). The dialer compares the name it hears with the owner’s name on our list, and scores how likely it is that the number is the owner’s.
So before a caller dials, they already know whether they’re likely to reach the owner.
Where the idea came from
Google has the same problem with business listings. It makes automated calls to businesses to check their opening hours, and it even calls to map out a business’s phone menu. A machine makes the repetitive call, and a person only works with the answer. The dialer does the same job for a calling list.
Where else this works
- Email lists: check that an address exists before someone writes to it.
- Any import: match records on a stable key (an email, a phone number, an ID), never on the row they arrived in.
- Any queue a person works through: put a cheap machine check in front, and re-check only the unsure cases.
Sources
- Kevin Hong on LinkedIn, the “dial sniper” post (the 20% dead and 30% voicemail figures).
- Google Business Profile Help, About automated calls and texts from Google to your business (the hours checks and phone-menu calls).
- The team numbers (callers, weekly answered calls and conversations) come from Caprae’s calling team, November 2026.
Steal Like an Engineer #2. Stolen from: Google. More from me: linkedin.com/in/ukantjadia.