Straight answer: AI-written cold emails read like AI-written cold emails, and prospects have learned to spot them in about two seconds flat, which is why reply rates on fully AI-generated outreach are collapsing across the board. The fix isn't to stop using AI, it's to stop letting it write the send and start using it for the research instead. I tested both approaches side by side last month and the gap was not small.
The test: 300 emails, two very different approaches
In January 2026 I ran a proper split test on outreach for a client in professional services, 150 UK-based small business owners in each group, same offer, same list, same day of the week. Group A got emails written start to finish by ChatGPT, using nothing more than the prospect's name, company and industry as inputs. Group B got emails where I used AI purely to pull three specific details about each prospect (a recent LinkedIn post, a piece of news about their business, or something from their website) and then I wrote the actual email myself in under two minutes each.
Group A: 4 replies out of 150. That's 2.7%. One booked call.
Group B: 19 replies out of 150. That's 12.7%. Six booked calls.
Same offer. Same list. Nearly five times the reply rate just by changing who wrote the sentence. That's the whole argument in one paragraph, but the reasons behind it matter more than the number.
Why prospects can tell in two seconds
People have been reading AI text for three years now. They know the tells, even if they couldn't name them out loud. The compliment sandwich ("I love what you're doing at [Company], your approach to X is impressive"). The fake specificity that isn't specific at all. The three-sentence paragraph that restates the subject line. The overuse of words like "excited" and "thrilled" in a cold message from a stranger. Ironically, AI models also lean hard on long dashes to link clauses together, which is one of the quiet giveaways once you know to look for it.
There's also a deliverability side to this that most posts on cold email skip entirely. Since Google and Yahoo tightened their bulk sender rules in February 2024, requiring authenticated domains, one-click unsubscribe and a spam complaint rate under 0.3%, mass AI-generated outreach has been getting caught in exactly the filters it should be worried about. Volume without quality doesn't just annoy people, it damages the sending domain, and a damaged domain affects every email you send after it, including the ones to existing clients.
What got the replies
Looking back at the 19 replies from Group B, every single one referenced something true and specific about the person, not the company. One prospect had posted on LinkedIn about missing his daughter's school play because of a client emergency. My opening line was one sentence acknowledging that, then straight into why I was emailing. He replied within 40 minutes.
This isn't a new insight, it's the oldest one in sales. Dale Carnegie wrote about this exact principle nearly ninety years ago, and if you want the fuller version of why genuine interest in the other person still outperforms every clever tactic, I've written up the specific business lessons from Dale Carnegie that still hold in 2026. AI hasn't changed the underlying psychology of why people respond to strangers. It's just made it obvious who bothered to look and who didn't.
The 20-minute system: AI for research, human for the send
Here's exactly what I did for Group B, step by step, so you can copy it without needing a course on prompt writing:
- Pull your list of 20 to 30 prospects into a spreadsheet with name, company, LinkedIn URL and website.
- Use a scraping or enrichment tool like Clay or Apollo to grab their most recent public activity, a LinkedIn post, a press mention, a job change, anything from the last 60 days.
- Feed that raw data into ChatGPT or Claude with one instruction: "Summarise the single most interesting or human detail here in one sentence, no compliments, no fluff."
- Take that one sentence and write the opening line yourself, in your own words, in under 30 seconds.
- Write the rest of the email by hand, three to five sentences, one clear ask, no attachments.
- Send from an authenticated domain, keep volume under 50 a day per inbox, and check your spam complaint rate weekly.
That's the whole system. Twenty minutes for five emails once you've done it a few times, not the two minutes per email that full automation promises, but the reply rate more than makes up for the extra time.
If this sounds like a system you'd rather have built and running for you than assemble yourself on a Tuesday afternoon, this is precisely the kind of workflow an AI consultant for small business sets up once and then hands over, rather than something you rebuild from scratch every quarter.
The uncomfortable bit about volume
Here's the part most posts on AI outreach won't say plainly: AI didn't fix targeting, it just made bad targeting faster and cheaper to send. If your list is wrong, or your offer is weak, no amount of clever personalization rescues the reply rate. I've seen business owners spend hours perfecting an AI prompt for the opening line while sending to a list that was never going to buy in the first place. AI can't fix a weak offer. It just helps you fail at a bigger scale, faster, and it does real damage to your domain reputation while it's at it.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
The businesses doing this well in 2026 are sending less, not more. Fewer emails, better targeting, a genuine reason for that specific person to hear from them that week. Volume was never the bottleneck. Relevance was.
Where AI does help
None of this means throw AI out of your outreach entirely. It's brilliant for the research step, the summarising step, and testing subject lines against each other before you commit to one. Companies that treat data and testing as core to how they operate, rather than an afterthought, tend to build stronger habits around this. It's worth studying how a product-led company like Hotjar built its brand around watching what users do rather than guessing, because that same discipline of testing small changes and measuring the result applies directly to subject lines and opening sentences.
AI is also good at keeping your brand voice consistent once you've defined it, which matters more than people think in outreach that goes out under your name. Duolingo's marketing strategy is a good study in how a distinct, consistent voice builds recognition over time, and there's no reason a solo consultant's cold email voice can't be just as recognisable to the people who receive it regularly.
One more idea worth stealing: instead of a cold email with a pitch, send one with a useful tool attached. I've written before about how interactive calculators engage and convert more visitors on a website, and the same logic works in an inbox. A short, useful calculator link ("work out what this is costing you") gets replies that a pitch never will, because it gives the prospect something before you've asked for anything.
Frequently asked questions
Should small businesses stop using AI for cold email completely?
No, but stop letting it write the final message. Use it to research the prospect and summarise what's worth mentioning, then write the email yourself in your own voice. That single change took one client's reply rate from 2.7% to 12.7% in a direct test.
How can you tell if an email was written by AI?
The most common tells are generic compliments that could apply to anyone, fake specificity that isn't specific, overly enthusiastic language from a stranger, and a structure that restates the subject line in the opening sentence. Prospects have been reading this pattern for three years and recognise it fast.
Does sending AI-written cold emails hurt deliverability?
It can. Since February 2024, Google and Yahoo require bulk senders to keep spam complaint rates under 0.3% and use proper authentication. Generic, low-relevance emails get reported more often, which damages the sending domain for every email sent afterward, including to existing clients.
What's a realistic reply rate to aim for on cold email in 2026?
Anything above 8 to 10% on a well-targeted, personalised campaign is solid for most small business services. Fully AI-generated, unpersonalised campaigns are commonly landing under 3%, which is often not worth the domain risk it carries.
Related reading: Miro Marketing Strategy: How They Built a Brand That Wins and Grammarly Marketing Strategy: How They Built a Brand That Wins.
