A few weeks ago, two cold emails landed in my inbox eleven minutes apart. One was pitching warehouse management software. The other was from a dog groomer two neighborhoods over. Both opened with “I hope this email finds you well.” Both listed three tidy bullets about streamlining operations. Both invited me to hop on a quick call.
I sat there wondering whether one guy had written both. Turns out the same thing had. It just wasn’t a guy.
Quick context, because I’m not coming at this from the anti-AI corner. I run StartMit, I spend my weeks wiring automation and AI into small businesses, and I put together a no-hype playbook on AI automation for SMBs because this stuff earns its keep when it’s pointed at real problems. If you’re curious who’s talking, I wrote up my own background and how StartMit got started a while back.
But customer-facing copy is where I’ve started drawing a line, and it’s a line I keep having to defend at dinner parties. When everyone’s brand voice comes out of the same handful of models, you don’t have a voice. You have a costume. Customers feel it even when they can’t name it, and it’s quietly one of the most expensive AI content pitfalls nobody warns you about.
The ‘AI smell’ is real, and customers clock it fast
Read enough generated text and you develop a nose for it. My team calls it the AI smell. Customers can’t articulate it either, but they react, usually by clicking away or by filing your email under “later,” which is where sales go to die.
The tells are consistent. Sentences arriving in neat groups of three. Adjectives like “robust” and “unparalleled” attached to products no human has ever described that way out loud. An opening paragraph that restates the question, then a closing paragraph that summarizes what you just read, in case you blacked out. Those long dashes everybody suddenly loves. And a strange universal politeness, like the text is scared to hold an opinion in case a hypothetical reader gets upset.
None of it is grammatically wrong. That’s what makes it sneaky. It reads clean, ticks the SEO boxes, and contains nothing a competitor couldn’t paste onto their own site tomorrow morning.
Last year a client of mine, an HR software company selling into dental practices, let me run an experiment on their demo follow-up sequence. We took 960 leads who’d requested a demo but never booked, split them down the middle, and sent two versions of the same email. Version A was written by ChatGPT (GPT-4 at the time), fed their complete brand guidelines and tone-of-voice doc, the works. Version B I wrote myself in about twenty minutes, and it opened with the founder’s sister, who ran a dental practice in Sheffield and once lost an entire Friday to a paper scheduling mess.
Version A pulled a 1.7% reply rate. Version B pulled 4.6%. Same list, same offer, same send window.
I’ll be honest about the limits of that test, because I’d rather be accurate than right. It’s one email in one niche, and I wanted version B to win, which is exactly the sort of thing that should make me distrust my own conclusions. Maybe the story would have worked no matter who typed it. Maybe that list was just tired of polished emails that week. I don’t fully know why the gap was that wide, and not knowing has bugged me for months.
The shortcut mentality is the real conversion killer
What worries me more than any model’s writing quirks is the mentality that creeps in once a business discovers that decent-looking copy is free and instant.
Copy is just where it shows up first. Shortcut the part of your business that literally talks to customers and the habit spreads. Automated replies that don’t quite answer the question asked. A chatbot deployed to deflect tickets rather than to help a single human. Proposals generated from a template with the competitor’s name half-swapped out. Each one saves an hour and quietly spends trust you can’t buy back.
The spend is invisible on dashboards, which is the trap. Nothing breaks. Open rates hold steady. Support volume drops and it looks like efficiency, when actually people have stopped expecting help. Then one quarter a competitor says something specific and human about a problem your customers genuinely have, and you’re left wondering where the pipeline went.
There’s a parallel from the ops side of my work. An automation that fails silently is worse than a person who forgets, because at least the person knows they forgot. It’s why I treat workflow reliability as its own discipline instead of an afterthought, and “is this trustworthy?” is the same question your copy has to answer, not just your Zaps.
I’m not lecturing from a mountaintop either. In 2023 I did this to myself. I published 34 blog posts in five weeks, all drafted by GPT-4 from one templated prompt, lightly edited by a VA I paid $450. Organic sessions climbed from about 2,100 a month to 6,800, and I felt like a genius. Demo requests fell from 11 that quarter to 4.
For another two months I let it run, because the traffic graph looked so good in client reports. I knew better and did it anyway. Wanting the line to go up got the better of me, and no tidy lesson arrived on schedule afterward. Mostly I just felt a bit stupid, and honestly I’d run a smaller version of that experiment again tomorrow, which tells you how unresolved it still is in my head.
Use AI for structure, keep the soul human
The split that’s held up for me over two years is simple to say and annoyingly easy to drift from: AI gets the invisible work, humans get anything a customer will actually read.
Concretely, a piece like this starts with customer calls recorded in Otter. I dump the transcripts into Claude and ask it to cluster the recurring complaints, which is useful precisely because it’s working from real words real people said. Outlines, research summaries, internal recaps, all fine. That’s skeleton work. Then I write the sentences myself. Badly some weeks, but mine.
Behind the scenes, go nuts. Agents summarizing support tickets, chasing invoices, triaging leads at 2am, all fair game, and if you want the mechanics, this is how AI agents actually work when they’re set up with proper guardrails. The plumbing barely matters, whether you’re on Zapier, Make, or n8n (we compared all three platforms). Internal work has no audience. Nobody forms an opinion of you from your invoice-chasing agent.
Customer-facing words are the opposite. They’re the product, or close enough. That’s the line I’d draw: automate what nobody sees, be a bit precious about what everybody does.
A five-minute audit that catches most AI content pitfalls
Before anything AI-touched ships under your name, run four checks. Five minutes once it’s a habit, and it catches the sameness problem early, while it’s still cheap to fix.
1. The specificity check
Does the piece contain at least one detail a competitor couldn’t produce? A customer’s exact phrasing, a real number, a mistake you made, the name of a street. Generic text carries no fingerprints. If there’s nothing in there that could only have come from you, it’s a costume, and readers feel costumes even when they can’t explain why.
2. The out-loud test
Read one paragraph out loud, ideally to an actual human. Would you say it to a customer standing in front of you? “We streamline operations through robust, scalable solutions” fails instantly, because nobody talks like that and everyone knows nobody talks like that. Polished generated copy lands the way a flyer lands: it arrived, nobody asked.
3. The competitor swap
Imagine the same words on your closest competitor’s homepage. If they’d work there word for word, you’ve written something that belongs to nobody. Your voice should break when it gets moved. That’s kind of the point of having one.
4. The haircut
Delete the first paragraph and the last one. Generated drafts open with throat-clearing and close with a tidy summary that insults nobody and helps nobody. Nine times out of ten the piece gets better. The tenth time, put them back and don’t sweat it.
None of this means handwriting everything and setting fire to your API keys. The models keep getting better at sounding like people, and pretending otherwise is its own kind of laziness. Last month a friend’s fully generated ad beat his handwritten one on clicks by about 30%, and I believe him. It doesn’t contradict anything above either, because his generated ad was built from forty pages of his own customer call transcripts. The machine had something real to say because he fed it something real.
Anyway. Customers forgive a typo from a person. They don’t forgive being processed by something pretending to be one.
I still get the balance wrong some weeks. A client recently caught a paragraph in my proposal that I’d clearly phoned in with Claude. She was right, I rewrote it, and I’m still a bit embarrassed. That’s the job, honestly. Catch it, fix it, keep the weird human parts in.
Talk soon,
Damian
