The subject line says "Quick question." It isn't a question. Under it sits "Following up," and under that one, somebody hopes this finds me well. My company is named, spelled correctly, inside a sentence that would survive swapping any other company into it. There's a compliment about a recent post that doesn't name the post. In 3 days a bump will arrive to float the thread back to the top of my inbox, fired by a sequencer as step 3 of 6, whether anyone read steps 1 and 2 or not.
I know the schedule because I've built it. I build and fix sales systems for small B2B teams, outbound included. I've configured the merge fields and picked the send windows myself. I use AI every working day, and I built the thing at the end of this article with it. So I can tell you nobody wrote the emails in my inbox this morning. I recognize the assembly.
01I build these machines for a living
A cold email system is a machine with about 5 moving parts. A list, an enrichment step that pulls your name and company into position, a template with slots, a sequencer that fires the follow-ups on schedule, and now a model that sands the sentences smooth. When one lands in your inbox, you're reading the output of a machine on the day it reached your row of the spreadsheet.
Which is why the tells that give these emails away sit below the prose. A stray merge field with the brackets still on. The compliment that names your company and knows nothing about it. The question engineered so that any reply at all counts as interest, because replies are the metric the sequence reports up. Machinery showing through the paint, and no model rewrite sands it out. The bump lands 3 days later because 3 is the number somebody typed into a delay field, once, months ago.
You're reading the output of a machine on the day it reached your row of the spreadsheet.
The volume is what you'd expect once the machines took over. Barracuda, working with researchers from Columbia and the University of Chicago, compared spam against mail from before ChatGPT existed and found 51% of it was AI-generated as of April 2025. And Google's own post announcing Gmail's Gemini features says "email volume is at an all-time high," which, with 3 billion inboxes, they would know.
02Everybody can tell it stopped working
Merriam-Webster made "slop" its 2025 word of the year, so the fatigue is official enough to have a dictionary entry. Both ends of the email have numbers on them now.
On the receiving end, BetterUp and Stanford's Social Media Lab surveyed US workers about AI-generated work, the report or deck a coworker hands over that looks finished and isn't. Receiving one cost people just under 2 hours of correcting and redoing, and 42% of them trusted the sender less afterward. Those are coworkers, people with every reason to think well of each other. A stranger's assembled email starts further behind.
That's a 20% decline in one year, reported by the people whose business is sending.
On the sending end, Belkins, an agency that cold emails for clients at industrial scale and publishes its numbers, watched replies in its own 7.5M-email dataset fall from 0.50% of sends to 0.40% across 2025. That's a 20% decline in one year, reported by the people whose business is sending.
I still use AI in outbound every day, model on the research, person on the message, and that version works fine. The other version, machines sending mail nobody spent a minute on, is the one dying in everyone's numbers.
03So I started replying
At some point I stopped deleting these emails and started answering them. Not with interest and not with anger. With a teardown: the exact lines quoted back, the stray merge field named, the bump schedule pointed out, the bait question called bait. I posted teardowns on LinkedIn, where calling out AI outreach has become its own small genre.
The logic firmed up the more of them I wrote. The sender spent no attention on me, and the sequence itself is the proof. A polite decline pretends a person asked me something. A teardown answers what actually wrote to me, and names it.
Writing a good teardown by hand takes real minutes, given to senders who had spent none on me.
The problem was the arithmetic. Writing a good teardown by hand takes real minutes, given to senders who had spent none on me. That bothered me long enough that I built a way out of it.
04The reply it deserves
Brick lives at outblox.com/brick. Paste in the cold email you got, and he does what I was doing by hand: reads it, underlines every tell he finds, the way a printout looks after someone's been at it with a red pen, and writes the 80-to-140-word reply it deserves, ready to copy and send back. Free, no signup or email gate. He's a small purple brick in a slightly tilted hard hat, drawn tired, because he reads email like this all day. I built him in 2 days with Claude Code, he runs on Claude, and a teardown costs about a cent, which is why he can stay free.
He judges what I judge: the machinery. Detectors that judge prose fail people; Stanford researchers ran 7 commercial AI detectors over essays by real people writing in their second language, and the detectors called the humans AI 61% of the time. Senders, meanwhile, can rewrite sentences until every detector shrugs. The sequence behavior is the part they can't hide, and roughly half of the 16 tells on Brick's checklist have nothing to do with AI prose at all; mail-merge templates were annoying people 2 decades before ChatGPT.
The same standard runs the other way, on purpose. If an email reads like a person actually wrote it, however clumsy, Brick says so, banner and all: "Probably a human. Brick doesn't do false accusations." He drafts a courteous reply instead and notes what read human. His eyes stay heavy-lidded through every teardown and only open fully on that verdict.
The obvious objection: an AI, built with AI, criticizing AI email. Every teardown opens by admitting it, in a fixed header: "This is an automated reply from Brick, an AI agent built by Outblox. Brick reads email flagged as AI-generated and returns a teardown. This one scored high." He signs off "Digitally, Brick." Nothing about him pretends to be a person, which is the one thing the emails he answers always do.
His closing line on every teardown is the part I'd put on the office wall, if I had an office wall: "If you want an actual response from a person, send something a tool didn't write. Until then, Brick and your sequence can keep talking, AI to AI."
The people filling your inbox aren't taking the time to write their emails. Why should you take the time to respond? Paste one into Brick and let him call it out. A real conversation gets started, or a fake one gets ended.