AI changed cold email for small teams. A person who used to send twenty careful, hand-written emails a day can now research and draft hundreds. That is a real advantage. It is also a fast way to destroy your sending reputation and fill inboxes with email that sounds like a machine wrote it, because that is exactly what happened.

This post is about using AI to do cold email well: personalised enough to earn a reply, human enough to not feel fake, and careful enough to stay out of spam. It is written for a B2B SMB with a small team, maybe one or two people doing outreach alongside everything else. If cold email is part of a bigger effort, it connects to our wider AI lead generation approach, which is where the replies need to land. This post assumes you already know who you are contacting and when. If that part is not decided yet, the B2B cold outreach strategy guide covers the ICP, the trigger signals, and the channel mix that come before a single email gets written.

Let me be direct about one thing first. More email is not the goal. Better email is. The whole point of AI here is to make each message good at scale, not to send more junk faster.

Why most AI cold email fails

Open your own spam folder and you will see the problem. AI made it cheap to send personalised-looking email, so everyone started doing it, badly.

The typical failure looks like this. A tool merges the prospect's first name and company, then adds an AI-generated line that says something vague like "I noticed your company is doing great things in your industry." That line is not real. The AI had nothing true to work from, so it guessed, and the guess is generic. Buyers have seen a thousand of these. They recognise it instantly and ignore it, or worse, mark it as spam, which hurts your reputation.

The second failure is volume. A new sender blasts hundreds of emails a day from a cold domain, spam filters notice the pattern, and the whole campaign disappears into junk folders. The email might have been fine. Nobody ever saw it.

Both failures come from using AI to do more of the wrong thing. The fix is to use AI to do the right thing, and to respect the rules that keep you in the inbox.

Personalise the thing that matters, not the name

Merging the first name is not personalisation. Every spammer does it, so buyers have learned it means nothing. Real personalisation is one true, specific observation that shows you actually looked at their business.

That can be a comment on something they recently posted, a change at their company like a new location or a new hire, a detail about how they operate that you noticed on their site, or a problem their specific type of business commonly faces. One real line at the top of the email does more than a dozen merged fields.

Here is where AI helps. It is good at two things in this workflow. First, research: given a company website or a LinkedIn profile, AI can pull out a relevant fact quickly. Second, phrasing: given a real fact, AI can turn it into a natural sentence. So the workflow is to find something true, then ask the AI to write one natural opening line about that specific fact.

The test for any opening line is simple. If you could copy and paste it into an email to a completely different company and it would still make sense, it is not personalisation. Delete it and find something real.

A workflow that keeps the human in control

The safe way to run this splits the work between AI and a person, so speed does not cost you quality.

  • AI does research. Feed it the prospect's website or profile and ask for one specific, relevant fact you could reference. Verify the fact is real before you use it, because AI sometimes invents details.
  • AI drafts the body. Give it your offer, the buyer type, and the one action you want, and ask for a short plain-text email. Keep it under 120 words with one clear ask.
  • A person writes or approves the opening line. This is the line that has to feel human and true. Even if AI drafts it from a real fact, a person should read it and fix anything that sounds off.
  • A person makes the send decision. No fully automated send-everything button. A human eyes the batch, catches the obvious misses, and controls the volume.

This is the same principle we apply across our AI marketing playbook: AI for the heavy lifting, a person on the parts where judgement and reputation are on the line. It is slower than full automation and far more effective, because the emails actually read like a human sent them.

Deliverability: how to not land in spam

You can write the best cold email in the world and it does nothing if it lands in spam. Deliverability is the part small teams skip, and it is the part that decides whether any of this works.

The rules that matter most:

  • Use a separate domain. Never send cold email from your main company domain. If a campaign gets flagged, you do not want your client email, invoices, and normal business email poisoned. Buy a lookalike domain, use it only for outreach, and keep your real domain clean.
  • Warm up before you send. A brand new domain sending hundreds of emails looks exactly like spam to filters. Ramp up slowly over several weeks. Start with a handful of emails a day and build gradually.
  • Keep volume low per inbox. Often twenty to fifty emails per sending address per day for a warmed-up domain. If you need more, spread it across several inboxes rather than pushing one inbox hard.
  • Set up authentication. SPF, DKIM, and DMARC records must all be correctly configured. Missing records are one of the most common reasons email lands in spam. This is a one-time setup that matters enormously.
  • Send plain and personal. A first cold email should look like a real person typed it, not like a designed newsletter. Avoid heavy images, many links, and spammy sales phrases in the first message.

One more that is specific to where you are. In Canada, cold email is governed by CASL, the anti-spam law, which is stricter than the rules in some other countries. In general you need a valid reason to contact someone, your message must identify you clearly, and it must include a working unsubscribe. This is not legal advice. Read the current CASL guidance or speak with someone qualified before you run a campaign, so you are compliant from the start.

Measuring what actually works

Stop leading with open rate. Privacy features, especially Apple Mail Privacy Protection, report opens that never happened, so the number is inflated and unreliable. The tracking pixel that measures opens can even hurt your deliverability.

Track these instead:

  • Reply rate. How many people responded. This is a real action, not a guessed pixel load.
  • Positive reply rate. Of the replies, how many were actually interested rather than "no thanks" or "unsubscribe." This is the number that predicts pipeline.
  • Meetings booked. The end goal. Everything above only matters if it leads here.

A campaign with a lower open rate but a higher positive reply rate is winning, whatever the open number says. Judge by replies and meetings, not by a metric that privacy tools broke.

Cold email is one channel, and it works best inside a system. New conversations still need a place to land, a way to be tracked, and a follow-up path once someone replies. Connecting outreach to your CRM and a clear nurture sequence is what turns a reply into real pipeline. If you want that whole path built, from the first personalised line to the follow-up that closes, start the conversation here.

Related guides: AI lead generation for SMBs · The AI marketing playbook · Apollo vs Clay for lead enrichment · Defining your B2B ICP · B2B cold outreach strategy