You have received the emails already. “I noticed your company is growing and wanted to reach out about how we help businesses like yours.” Five a day, all slightly different, all obviously written by a machine, all deleted. Now a vendor says AI cold email outreach is the answer for your own sales, and the demo showed a tool writing a thousand of those in a minute.
The thousand emails are the easy part. The hard parts are knowing who to write to, having something true to say to each of them, and getting the message into an inbox instead of a spam folder. AI helps with two of those three and, used carelessly, makes the third worse.
This article explains how to use AI for outbound email in a way that gets replies from the right people without burning your email reputation: the list, the research, the writing, the sending, and the replies. By the end you should be able to look at any outbound tool or agency proposal and know whether it is built to work or built to look busy.
What this actually is
Cold email outreach means emailing people who have not asked to hear from you, because you have reason to believe you solve a problem they have. It is legal in the United States under the CAN-SPAM Act as long as you identify yourself, include a physical address, and honor requests to stop.
AI enters in two places. A language model (software that reads and writes text, the engine behind Claude, ChatGPT and Gemini) can read a company’s website, job postings and news and pull out the one or two facts that make an email relevant. It can then draft a short message around those facts. It cannot make a bad list good, or make a mailbox that has been sending junk look trustworthy to Google and Microsoft.
The ordinary-business analogy is a good field sales rep versus a leaflet drop. The rep visits twenty businesses, knows something about each before walking in, says one useful thing, and asks one small question. AI cold email done well is the rep with the research done faster. Done badly, it is the leaflet drop with a fake handwritten font.
1. Fix the list before you touch the writing
Every reply-rate problem we have looked at started with the list. If a third of the addresses bounce, no writing can save it, because the bounces tell mail providers you do not know who you are emailing. If the people do not have the problem you solve, the best email gets a polite no.
Build the list narrow: the kind of business (independent HVAC companies with ten to forty employees in three states) and the person inside it (the owner or operations manager, not “decision maker”). Verify every address before it enters the sending tool; verification means a service confirms the mailbox exists and accepts mail (AI Lead Research and Enrichment for Business Outreach). A rough rule from our own work: two hundred people who genuinely fit outperform five thousand who roughly fit.
2. Research first, write second
The AI step that earns its keep is reading. Point a model at the company’s website, careers page, news and reviews, and ask: “What is this company doing right now that suggests it might have problem X?” The answer is the hook. If there is no answer, the company comes off the list. Most tools skip that last part, because a shorter list looks like a worse demo.
Good research output is a fact, not an adjective. “They posted a job for a second dispatcher in August” is a fact. “They are a growing, customer-focused company” is filler. Give the model rules: return only things that appear on a page it read, quote the source, and say “nothing found” when there is nothing. Running this across a list and checking it on a sample is covered in AI Personalization at Scale for Outreach That Reads as Human.
3. One observation, one offer, one ask
The structure a busy owner will actually read has three parts. One observation: the fact you found, in one sentence. One offer: what you do for businesses in that situation, in one or two sentences with no adjectives. One ask: a single small question that is easy to answer with a yes, a no, or a name. That is the whole email, sixty to one hundred and twenty words. No company history, no feature list.
The ask matters more than people think. “Can we hop on a quick call?” asks for thirty minutes of a stranger’s time. “Is dispatch scheduling something you handle yourself, or does someone else own that?” asks for one sentence and tells you who to talk to next.
4. Write in plain language and sound like a person
If the email could have been written by any company to any company, it will be deleted. Plain language means short words, short sentences, no marketing vocabulary, and a first line that could only have been written to this recipient. Read the draft out loud; if you would be embarrassed to say it to the person’s face, cut it. The model produces marketing prose by default because it has read millions of bad emails. Tell it not to, give it three examples of the tone you want, and reject drafts that slip.
Sign with a real name and title, from a mailbox a real person checks. An email signed by the owner gets read differently from one signed by “The Team”.
5. Keep sending volume low and steady
This is where AI outreach goes wrong most often, and the damage is not reversible in a day. Google and Microsoft judge every sending domain by its behavior: volume, bounces, spam complaints, replies. A new mailbox that sends two hundred cold emails on its first Monday looks like a spammer and is treated like one.
Our rule, and the rule of every deliverability specialist we trust: never send more than about fifty cold emails per day from any one mailbox, and start well below that. A new mailbox spends its first two to four weeks sending a handful a day and increasing slowly. To reach more people, add mailboxes on separate domains rather than pushing one harder. Cold mail should never leave your main company domain at all; the reasoning and the setup are in Email Deliverability for Cold Outreach, Explained Plainly. Steady beats bursts: twenty a day looks like a person, two hundred on Tuesday looks like a campaign.
6. Sequence with restraint
A sequence is the set of follow-ups that go out automatically when the first email gets no reply. They work, because most people do not answer the first email, and they are where tools encourage bad behavior. Seven follow-ups over three weeks is harassment, and “just bumping this” trains recipients to hit the spam button.
Two follow-ups is usually enough, three at most, four to seven days apart, each adding something new. Any reply, including “not interested”, must stop the sequence instantly. This is the most common technical failure we see: a reply lands, the tool misses it, and the fourth email goes to someone who already said no. Test it before the first real send.
7. Handle every reply like a lead, because it is one
A reply is a warm lead that cost almost nothing, and the clock starts when it lands. A useful response within minutes, not days, decides whether the conversation continues (Automated Lead Follow-Up That Replies in Minutes, Not Days). A person, not the AI, writes it.
Sort replies as they arrive: interested, question, referral, not now, not interested, unsubscribe, bounce. A model classifies these well and can draft a reply for a human to approve. Interested gets a personal answer the same hour. Referral gets a fresh email to the named person. Not now gets a note in the CRM (the software where you track contacts and deals) with a date to check back. Not interested goes on the do-not-contact list permanently. Every reply lands in the CRM against the contact (AI CRM Automation That Makes the CRM Do the Work for You).
8. Know the law in every country you send to
CAN-SPAM in the United States is lenient: identify yourself, include a physical mailing address, no misleading subject lines, honor opt-outs within ten business days. Canada’s CASL generally requires consent or an existing business relationship. The UK and EU, under GDPR and the ePrivacy rules, treat an individual’s work email as personal data and require a documented legitimate interest and an easy way to object. Australia requires consent in most cases.
The practical rule for a US business: send to US addresses within CAN-SPAM, and check every other country specifically before sending there. Have a real opt-out, honor it immediately, and keep records.
9. Measure the numbers that predict trouble
Reply rate is the number everyone watches and the least useful one early on. The numbers that predict a block are bounce rate (keep it under two percent, and stop if it climbs), spam complaints (any on a small list is a warning), and open rate collapsing across a mailbox, which means messages are landing in spam. Check these daily in the first month, and keep a seed inbox (a mailbox you own at Gmail and one at Outlook) on every list so you can see where messages land.
Positive reply rate, meaning interested replies and real questions, is the number to optimize once deliverability is healthy. On a well-built list with genuine research, low single digits is normal and profitable.
Picture a business like this one
The business below is a composite of the kind of company that writes to us, not a client. The numbers describe the shape of the problem, not a case study.
Picture a business like this one: a commercial cleaning company with forty staff, grown on referrals, that wants to add office buildings and medical clinics in two neighboring cities. The owner tried an agency that sent four thousand emails in a month from the main domain. Replies were rare and mostly angry, and by week three the company’s invoices were landing in customers’ spam folders. The agency suggested a bigger list.
What a business like this would build instead:
- A separate sending domain, a close variant of the company name, with two mailboxes warmed up for three weeks before any cold mail.
- A list of about three hundred property managers and clinic office managers, built from public directories and verified, with a do-not-contact list seeded from every existing customer and everyone who replied to the agency campaign.
- A research step where a model reads each company’s site and job postings and returns one concrete fact or “nothing found”. Roughly a third come back empty and are removed.
- A template with three slots: the fact, a one-sentence offer, and a question about who handles vendor selection. The owner rewrites the first twenty drafts by hand and that tone becomes the model’s examples.
- Twenty sends per mailbox per day, two follow-ups, replies classified by the model and answered by the operations manager within the hour, everything written to the CRM.
What changes: the numbers are smaller in every direction, which is the point. Two hundred emails go out over two weeks instead of four thousand, and the main domain is untouched. Replies come from people who run buildings, several are “talk to our facilities lead, here is her email”, and the conversations that follow are about square footage and start dates.
What it costs to run
A second domain runs around ten to fifteen dollars a year. Mailboxes on Google Workspace or Microsoft 365 are a few dollars a month each; check the current pricing pages. A sending tool such as Instantly or Smartlead is typically thirty to one hundred dollars a month at small volumes, warm-up included. Verification costs fractions of a cent per address. Lead data tools such as Apollo or Clay have free tiers and paid plans that climb with volume; budget fifty to a few hundred dollars a month.
The AI research step costs a fraction of a cent per company with current models, so a thousand companies is a few dollars. A small server or workflow tool adds ten to twenty dollars a month. The real cost is human time: reviewing the research, approving the tone, and answering replies, a few hours a week for a small program.
The mistakes we see most
Sending from the main company domain. This costs the most and takes longest to undo, because invoices and support emails suffer too.
Buying a list and sending the same day. Purchased lists are full of dead addresses and people every other buyer has already emailed. Verify everything and expect to discard a large share.
Letting the model write freely. Without rules, examples and a review sample, the output reverts to the average of every marketing email ever written.
Follow-up sequences that never stop. Seven emails with no reply detection is how a business ends up on a blocklist and in a screenshot on social media.
Treating a reply as a task for the tool. The moment a human answers, it is a sales conversation.
When to bring in help
An owner with patience can run a small program alone with off-the-shelf tools: a separate domain, a Workspace mailbox, Instantly or similar, a verification service, and a spreadsheet. Do the research by hand at first, twenty companies at a time; it is the best way to learn what a good hook looks like before automating it.
A developer makes sense when the research needs to run across hundreds of companies a week with rules you can trust, when replies must land in your CRM automatically, when the do-not-contact logic is enforced by software rather than memory, and when you want logs of what was sent to whom and why. Those are the pieces where hand-built setups quietly fail.
Levelbrook builds outreach systems like this for businesses, the research pipeline, the sending rules, reply classification and the CRM integration, at a fixed price from a written scope. Everything runs in accounts you own. The form below is how a conversation starts.