You can build an AI system that reads every cold email reply, classifies whether the person is interested, annoyed, or just out of office, and fires back a contextual response in under two minutes — all without you lifting a finger. I know because I built one, it’s been running for a month, and it’s changed how I think about outbound entirely.

Updated April 2026

Why did I build this in the first place?

I was doing what most people who run cold email do — checking replies manually a few times a day, trying to respond quickly to the interested ones, and inevitably missing the ones that came in at 9pm or over the weekend. The maths bothered me more than anything else.

According to research published by Harvard Business Review, responding to a new lead within five minutes makes you 100x more likely to actually reach them compared to waiting 30 minutes. That’s not a typo — one hundred times. And the data from InsideSales.com’s lead response study backs it up: 78% of deals go to the company that responds first. So every time I was asleep or in a meeting or just having dinner when a reply came in, I was haemorrhaging conversion.

The other problem was classification. When you’re running campaigns to a few thousand people, you get a lot of replies, and they’re not all the same. Some are interested. Some want you to go away. Some are out of office auto-replies. Some are the wrong person but they’ve helpfully told you who to contact instead. And some are just confused. Manually sorting through all of that multiple times a day was eating 30-45 minutes I didn’t have.

So I built something to handle it.

What does the system actually do?

The whole thing runs on three layers, and I want to be specific about this because most “I automated X” posts are frustratingly vague about what the system actually does.

Layer 1: Ingestion. The system polls my cold email platform’s API every few minutes for new replies. When it finds one, it pulls the full reply content, the original email that was sent, and any metadata about the lead — their name, company, what campaign they were part of. All of that goes into a processing queue.

Layer 2: Classification. This is where Claude comes in. The AI reads the reply and classifies it into one of five categories: Interested, Not Interested, Out of Office, Wrong Person, or Do Not Contact. But it’s not just slapping a label on — it also extracts key information. If someone says “I’m interested but not until June,” the system captures the timing. If they say “you want my colleague Sarah, her email is sarah@company.com,” it captures the referral.

According to Woodpecker’s 2025 cold email benchmark report, average cold email response rates sit between 1-5%, and of those responses, roughly 30-40% express genuine interest. The rest are a mix of opt-outs, auto-replies, and wrong-person redirects. Having AI sort that automatically means I only see the ones that matter.

Layer 3: Response. For interested replies, the AI generates a personalised response that pushes toward a call booking. It’s not sending a canned template — it reads what the person actually said and responds to it, like a human would. The response goes out in under two minutes from when the original reply landed. For unsubscribe requests, it honours them immediately and removes the lead. For wrong-person referrals, it logs the new contact for follow-up.

The thing that surprised me most wasn’t the speed — it was the consistency. Every single interested reply gets a thoughtful response within two minutes, whether it comes in at 2pm on a Tuesday or 11pm on a Saturday. That consistency is something I could never achieve manually, and it turns out consistency matters more than cleverness in cold outbound.

How does the AI know what to say?

This is the part most people get wrong when they think about automating replies. They imagine setting up a few templates and having the AI pick between them, like a glorified autoresponder. That’s not what this is.

The AI has context about the original campaign — what value proposition was pitched, what industry the lead is in, what their role is. When it reads their reply, it generates a response that acknowledges what they actually said and moves the conversation toward a call. If someone says “this sounds interesting, but we already have something in place,” the response addresses that specific objection. If someone says “tell me more,” the response gives them the next layer of information and suggests a time to talk.

I built this with Claude’s API, and the key design decision was keeping the responses short and human. No one wants to get a 500-word essay in response to their two-line email. The system aims for three to five sentences — acknowledge what they said, add one piece of value, suggest a specific next step. That’s it.

According to Boomerang’s email response study, emails between 50-125 words get the highest response rates at around 50%, so the AI is optimised to stay in that range. Anything longer and you’re actually hurting your chances.

What happened when I turned it on?

The first week was nerve-wracking, not going to lie. I kept checking the responses the AI was sending, half-expecting it to say something bizarre to a prospect. But it didn’t. The responses were better than what I was writing at 7am before my first coffee — clearer, more consistent, and always hitting the right note.

Here’s what the numbers looked like after the first month of running the system:

The speed-to-lead improvement was the biggest factor. A study by Drift in 2025 found that the average B2B company takes 42 hours to respond to a new lead. Forty-two hours. Even my manual 3-4 hour average was dramatically better than that, but getting it down to two minutes put me in a category most companies can’t touch.

I had a prospect reply at 10:47pm on a Thursday saying “this could work for us, can we talk?” The AI responded at 10:48pm with a contextual reply and a Calendly link. The prospect booked a call for Friday morning before I even saw the email. That call turned into a client. If I’d responded the next morning — even first thing — the moment might have passed.

What about the replies that go wrong?

They happen, and being honest about that matters more than pretending the system is perfect. In the first month, I’d say maybe 5-8% of the AI’s responses could have been better. Not wrong, exactly, but not quite what I would have said. Usually it was a tone thing — being slightly too formal when the prospect was casual, or vice versa.

The critical thing is that none of the mistakes were catastrophic. The system never sent something offensive, never promised something it shouldn’t have, and never misclassified a “do not contact” as an interested lead. The guardrails I built in were straightforward: any reply that the AI isn’t confident about classifying gets flagged for human review instead of getting an automatic response.

According to Gartner’s 2025 AI in Sales report, 74% of sales teams using AI for email automation report that the AI’s responses are “acceptable or better” without any human editing. That matches my experience — most responses are good to go, a small percentage need a tweak, and a very small percentage need rewriting.

I’m going to address this head-on because it matters. B2B cold email in the UK is legal under PECR regulations as long as you have a legitimate interest basis, include a clear way to opt out, and identify yourself properly. The ICO’s guidance is clear on this — unsolicited B2B emails are permitted as long as they’re relevant to the recipient’s professional role. Nothing about automating your reply process changes any of that.

The ethical question is slightly different: is it okay to have AI reply to people without telling them they’re talking to an AI? I thought about this a lot. My position is that the AI is acting as my assistant — it’s responding with my intent, in my voice, pushing toward a conversation with me (the human). It’s no different from having a VA handle your inbox, which nobody considers deceptive. The moment someone books a call, they’re talking to me, Matthew Lowe, in person.

That said, I built the system to be straightforward. The responses don’t pretend to have experiences they don’t have. They don’t make up stories. They respond to what was said and move the conversation forward. That feels right to me.

What would I change if I built it again?

Three things. First, I’d start with a narrower classification system. Five categories felt right but I ended up adding sub-categories pretty quickly — “interested but not now” versus “interested and ready” require different responses, and the original system treated them the same.

Second, I’d build in better tracking of conversation threads from day one. The system handles first replies brilliantly but when a conversation goes back and forth three or four times, keeping the context tight requires more thought than I initially gave it.

Third — and this is the Matthew Lowe operator-brain talking — I’d have turned it on sooner. I spent weeks tweaking and testing when I could have just shipped it and corrected as I went. According to CB Insights’ analysis of failed automation projects, 42% fail because of over-engineering before launch, not because the core idea was wrong. Speed to live matters more than perfection at launch.

Should you build one of these?

If you’re sending cold email at any real volume — say 50+ emails per day — and you’re handling replies manually, you’re leaving money on the table. Not because manual replies are bad, but because manual replies are slow, inconsistent, and they stop when you stop working.

The Zero Hire Method exists because I believe small businesses should have the same operational advantages as companies with 50-person sales teams. An AI reply system is one of those advantages — the kind of thing that a well-funded startup has an SDR team doing around the clock, but that a founder-led business can replicate for the cost of an API subscription.

The tech is ready. The question is whether you’re willing to trust it, and after a month of watching it work, I can tell you the trust builds faster than you’d expect. Not because the AI is perfect, but because it’s consistent, and consistency in outbound sales is worth more than occasional brilliance.