A step-by-step view of which parts of writing cold outreach AI can do, which are plain automation, where a person must review and how to run a small pilot.
7 stepsTypical mix: AI candidateIllustrative analysisUpdated
Yes, AI can write most of a cold outreach message, but a person should still decide who gets it and approve what is sent. AI does the research summary, the first draft and the sorting of replies well; rules handle the list, the timing and the opt-outs; and sales leadership keeps the choice of segment and offer. Cold outreach automation works best as a draft-and-review loop, not as a button that mails thousands of people.
What each step needs
01● AI candidate
Researching the prospect and the company
A language model can read a prospect's website, recent news, job posts and public profile, then return three or four facts that are actually relevant to your offer. This is the step where AI saves the most time, because the manual version is slow and gets skipped when the list is long. Ask for a short brief with a source link per fact, so a person can check any claim in seconds. Do not let the model fill gaps with guesses about the company; an empty field is better than an invented one.
⛨Works when you can feed it public sources and a clear description of what you sell. Keep the brief to facts with links, not opinions.
02● Keep human
Deciding who to contact and what to offer
Choosing the segment, the role you write to and the reason they should care is a commercial decision. AI can suggest segments from your closed deals, but it cannot know which relationship you want to build or which offer you are willing to stand behind. If this step is vague, every message that follows will be vague too, however well it is written.
⛨Owned by sales leadership. Revisit it every quarter or when reply rates drop for a whole segment.
03● Standard automation
Building and cleaning the contact list
Removing duplicates, checking that an address is formatted correctly, excluding current customers and anyone who opted out, and respecting a contact limit per account are rules, not judgement. A normal automation or your CRM does this reliably and cheaply. Using AI here adds cost and the occasional surprise for no gain.
⛨Needs a clean suppression list and one source of truth for who is already in a conversation.
04● AI candidate
Drafting the first message
Given the brief and your approved positioning, a model can write a short first email that opens with something specific to the prospect, states one clear reason to talk and asks one easy question. It can produce several variants for testing. The risk is generic fluency: messages that sound polished and say nothing. Give it examples of your best past replies and a hard limit on length.
⛨Works with a one-page message guide: who you are, what you offer, claims you may make and phrases you never use.
05● Human review
Reviewing tone, claims and legal basis
A person should read what goes out, at least for the first weeks and for any new segment. They check that every fact is true, that the tone fits the person and that you may contact them at all, since rules on unsolicited messages differ by country and by whether the recipient is a company or an individual. Over time you can move to sampling instead of reading every message.
⛨Start with 100 percent review, then sample once errors are rare. Keep a log of what was changed and why.
06● Standard automation
Sending, sequencing and follow-ups
When to send, how many days between follow-ups, when to stop after a reply and how to include an unsubscribe line are all fixed rules. Your sales engagement tool or CRM handles this. AI should not decide to send extra follow-ups on its own, because the cost of being annoying lands on your brand and your sending domain.
⛨Needs proper email authentication (SPF, DKIM, DMARC), a daily sending cap and a visible way to opt out.
07● AI candidate
Sorting replies and booking the meeting
A model can label replies as interested, not now, wrong person, out of office or unsubscribe, and draft a short answer or route the right person's name to the right rep. Unsubscribe requests must be honoured by a rule, not left to a model's reading. The conversation itself, once someone is interested, belongs to a person.
⛨Works when replies land in a shared inbox or CRM where a person can overrule the label in one click.
A sensible first experiment
Pick one segment of 100 prospects and one offer. Let AI produce a short research brief and a first draft for each, and have one salesperson review every draft and fix what is wrong. Send half with the AI-assisted draft and half with your current message. After three weeks, compare replies that are positive, replies that are complaints or unsubscribes, and the minutes spent per message. Only widen the pilot if the AI-assisted half is at least as good on both reply quality and complaints.
The trap to avoid
The biggest trap is volume. Once drafting is cheap, it is tempting to send ten times as many messages, and the result is more complaints, a damaged sending domain and a worse name in your market. Set a cap on messages per day and per account, track unsubscribes and spam complaints weekly, and treat a rise as a reason to slow down, not to rewrite the template.
Questions teams ask
Can AI send cold emails on its own?
Technically yes, but you should not let it. Sending is a rules job: your sales tool should control timing, caps and opt-outs. AI is best used before sending to research and draft, and after sending to sort replies, with a person approving the messages, at least until errors are rare.
Is cold outreach with AI legal?
The tool does not change the law; the recipient and the country do. Rules differ for companies and individuals and between the US, the UK and the EU, and some countries require prior consent even for business email. Check the rules for each market with a qualified adviser, and always include a working way to opt out.
Will AI-written outreach get a lower reply rate?
It can, if the messages are generic. Replies depend mostly on relevance and on the offer, not on who typed the words. Messages built on a checked, specific fact about the prospect tend to do better than polished but empty ones, so test your own segments against your current message instead of trusting averages.
What should a person always do in cold outreach?
Choose the segment and the offer, review the messages for truth and tone, and handle every real conversation. They should also watch complaints and unsubscribes, because a model will not notice that your sending domain is losing its reputation.
Illustrative workflow guidance by Arcgent. Each business needs its own assessment. No integration or savings claim has been verified for your systems.