AI Prospecting That Does Not Sound Like AI: A Research Workflow That Earns Replies

Picture a sales rep who opens their inbox on Monday and finds five cold emails. All five open with some version of “I came across your company and was impressed by your growth.” Two mention the rep’s first name in a way that does not quite fit. One is clearly the second email in a sequence that began a week ago, and says so. The rep deletes all five in under a minute. (This is an illustration, not a case study.)

That rep is also, on other days, the one sending those emails. AI made it cheap to write and send a thousand messages, so a thousand messages is what a lot of teams send. This guide takes the opposite position: AI cold email works when it does the research and the first draft, and a person decides what goes out. You will get a five-stage prospecting workflow, five sources of personalization that go beyond “I saw your website,” four prompts to start from, four cold email templates with a worked example, sending rules, and a way to measure the only number that matters: replies that turn into real conversations.

A note on what this is. It is a framework built from public sources (listed at the end, with the date we read each one) and from common sales-operations practice. We have not run a prospecting experiment of our own, so there are no reply-rate results claimed here. The prompts and templates are starting points, not proven scripts. Every number inside a worked example is made up to show the arithmetic and is labeled as such. This is general information, not legal advice.

Why mass AI prospecting fails

Before AI, writing a cold email cost a few minutes, which put a natural ceiling on volume. Now the cost of one more message is close to zero, and the ceiling is gone. Four things follow from that, and none of them needs a statistic to see.

1. Everyone sends, so the inbox is full of the same email

When a tool can produce a plausible opening line for any company, thousands of senders use the same tools with the same prompts and get the same kinds of lines. Recipients do not need to know the lines are AI-written. They only need to have seen the shape of the email before: a compliment that could apply to anyone, a one-line pitch, a request for fifteen minutes.

2. The personalization is cosmetic

Mail-merge personalization (first name, company name, job title) is easy to spot, and “I saw you are the VP of Marketing at Acme” tells the recipient nothing they did not already know. Cosmetic personalization shows that a database was queried. Real personalization shows that someone looked at the recipient’s situation and has a reason to write.

3. Mailbox providers measure complaints, and complaints are a rate

Google’s published sender guidance, as summarized by Mailsuite (we could not open Google’s own page, see Sources), tells senders to keep their spam complaint rate low: below 0.10% as a target and never reaching 0.30%. A rate is a ratio, so the arithmetic is unforgiving. Three people marking your message as spam out of a thousand sent is already 0.30%. The same three complaints out of one hundred sent would be 3%, which is why small senders with narrow targeting have an easier time staying out of trouble than large senders with broad lists, but also why a rushed batch to the wrong audience can hurt quickly.

4. Your sending reputation is shared

If you send cold outreach from the same domain you use for customer email and invoices, a bad run of cold email can affect messages that matter. A common practice is to use a separate domain, or at least a separate mailbox, for outbound prospecting, and to keep volume low while it builds a history. Treat that as operations advice, not a rule. Whatever you choose, authenticate the domain (SPF and DKIM) before you send anything.

The twenty-versus-two-hundred arithmetic

The promise of this guide is a workflow that produces twenty deeply personalized messages a day instead of two hundred generic ones. The honest reason is not that twenty beats two hundred in every case. It is that twenty is what one person can actually review, and review is the step that separates a draft from a message. Here is the arithmetic with invented numbers, so you can see the shape and then replace the assumptions with your own.

We looked for a trustworthy published benchmark for cold email reply rates to anchor these assumptions, and did not find a primary source we could read. Figures in circulation vary widely with the list, the offer and the sender, and many are published by vendors selling outreach tools. So this guide does not quote one. The only benchmark that will ever be reliable for you is your own, measured the way Section 7 shows.

If your pipeline already holds leads that went quiet, you may not need new prospects at all. Our guide to the missed-lead follow-up system covers how to recover revenue from conversations that already exist, and it is usually the cheaper place to start. And if you would rather have the research workflow built around your own CRM and sending tools, the Request a Custom Workflow page is where you describe it and get a quote.

The five-stage research workflow

The workflow has five stages, and each stage has a gate: a condition that must be true before the account moves on. Gates are what keep volume honest. An account that fails a gate is skipped, and skipping is a normal, healthy outcome.

Stage 1: The target list, kept small and specific

Start from a written description of who you can help and why now. Not “SMBs in the US,” but something like: marketing agencies with five to twenty staff that hired an operations role in the last quarter. The narrower the description, the easier it is to find something specific to say. Build a Google Sheet with one row per account and these columns: company, website, contact name, contact role, contact email (verified, see Section 6), source of the lead, status, and notes.

Gate 1: the account matches your written description, and the contact is a person who plausibly owns the problem you solve. If you have to stretch to make it fit, it does not fit.

Before you load any list, make sure the data in it is clean. Duplicate contacts, outdated roles and wrong emails turn a careful message into an embarrassing one. Our Zoho CRM workflow for automated data quality audits shows the kind of check that catches those problems before they reach an outbound batch.

Stage 2: Research one account at a time, with sources

This is the stage AI is best at and where it most often goes wrong. The job is to collect a short set of public facts about the account and the person, each with a source link and a date. The output is a research brief, not an email. A good brief has three to five facts, each one traceable.

The single most important rule at this stage: the model may only report facts that appear in the sources you gave it. Language models will fill gaps with plausible-sounding details if you let them. A made-up detail in a cold email, such as a funding round that did not happen or a podcast the recipient was never on, is worse than a generic email, because it is a lie addressed to a stranger. The research prompt in Section 4 forces every fact to carry a source URL and tells the model to return “no verifiable fact found” when that is the truth.

Gate 2: the brief contains at least one fact with a source you can open and a date from the last twelve months (or a longer window you chose on purpose). No verifiable fact means the account is marked Skip. It can be revisited when something new happens.

Stage 3: Find one angle that connects the fact to your offer

A fact alone is not a reason to write. “Congratulations on the funding round” is a fact plus a greeting. An angle is a fact plus a connection: this thing happened, which usually creates this problem, which is what we help with. The model proposes up to two angles. A person picks one or rejects both.

A useful test for an angle is to remove your company name from the email. If the message would still make sense to the recipient as a note from a thoughtful stranger who read about their situation, the angle is sound. If it only makes sense as a pitch, the angle is a pretext.

Gate 3: the angle names a specific situation and connects it to one thing you can do. If you cannot say it in one sentence, do not write the email.

Stage 4: The AI draft, under strict rules

Now the model writes a short first draft from the chosen angle and the research brief. It receives rules, not freedom: a word limit, a ban on certain phrases, one call to action, no claims that are not in the brief, and an instruction to write the way a person writes to a colleague. The drafting prompt in Section 4 shows the shape.

A second, separate model call can then check the draft against the brief and the rules, and flag any sentence that makes a claim the brief does not support. Keep the checker separate from the writer. A model asked to both write and grade its own work tends to approve it.

Stage 5: Human review, the step that makes it work

A person reads every message before it is sent. Not skims: reads, with the research brief open next to it. The reviewer asks four questions:

1. Is every claim true? Open the source link for at least the main fact.
2. Would I be comfortable if the recipient saw my research? If the answer is no, the personalization is crossing a line (more on that below).
3. Does it sound like a person? Read it aloud. If you would not say it to a colleague, rewrite it.
4. Is the ask small and clear? One question, easy to answer yes or no.

The reviewer then marks the row Approved, Edit, or Reject. Only Approved rows move on. This is where twenty a day comes from: with research and a first draft already done, a reviewer spending a few minutes per account can handle about twenty in under two hours. Your own timing will differ, and the real number is the one you measure.

What this looks like inside n8n

If you use n8n (may be an affiliate link), the five stages map to a short chain. This is a description of a pattern, not a published LeenLabs workflow, so treat the node names as suggestions:

• Schedule Trigger each weekday morning.
• Google Sheets (may be an affiliate link) read: the next 20 rows with status “Ready.”
• Search step: a news or web search service returning recent results about the company and the contact (the service is your choice, check its terms and cost).
• AI step 1: build the research brief as structured JSON, with a source URL and date for each fact. A condition node routes “no verifiable fact” rows to Skip.
• AI step 2: propose up to two angles, then draft the email from the chosen one under your rules.
• AI step 3: a separate check against the brief and the rules.
• Write to a Review sheet: brief, angle, draft and check result, with an empty Decision column for the reviewer.
• A second workflow reads rows marked Approved and creates a draft in Gmail (may be an affiliate link). It does not send. A person sends, or schedules, from the mailbox.
• Log: the status, the date and the template used are written back to the row so you never contact someone twice by accident.

The closest published example on this site is the LinkedIn lead scoring and outreach workflow. As described on its page, it reads a Notion table, uses a search service, has Claude grade each lead from 1 to 10, drafts two messages, puts the result in a Google Sheet for review, and never sends anything on its own. Its page also includes a section on legal and account risks, which is worth reading before you build anything that touches LinkedIn. Check its page for the details that decide whether it fits your case.

Five sources of personalization that are not “I saw your website”

“I saw your website” is the weakest hook there is, because the recipient controls what is on their website and knows everyone has seen it. Better sources share three traits. They are recent, they are specific to this person or company, and they hint at a need your offer can address. Here are five, with where to look, what to extract, and an invented example of the kind of line they can produce.

Source 1: A job posting the company published

Job posts are public, dated and revealing. A company hiring a first operations manager, a RevOps lead or a customer support lead is telling you where it feels pain. Where to look: the company’s careers page and its public job listings. What to extract: the role, the date, two or three responsibilities that overlap with your work, and any tools named in the requirements. Made-up example: “I saw you are hiring your first operations manager and the posting mentions reconciling data across three systems. That is usually the point where a spreadsheet stops scaling.”

Watch out for: the posting being old or filled, and for inferring too much. A job post says what a company wants to hire for, not what is going wrong inside.

Source 2: A recent company announcement

A product launch, a new market, a partnership, a funding announcement, an office opening. Where to look: the company blog, press page, and its own social accounts. What to extract: what was announced, the date, and one practical consequence for the people running it. Made-up example: “Congratulations on opening the Austin office. Teams that add a second location often find that the way leads were routed when everyone was in one room stops working. Is that on your radar?”

Watch out for: announcements that are three years old and the temptation to congratulate without adding anything. A congratulation is a greeting. The sentence after it is what earns the reply.

Source 3: Something the person wrote or said

A LinkedIn post, a blog article, a conference talk, a podcast appearance. This is the strongest source because it is the recipient’s own words. Where to look: their public posts, their company’s blog, podcast and event pages. What to extract: one specific idea they stated, quoted accurately, plus the date and a link. Made-up example: “In your post last month you wrote that most onboarding problems are handoff problems. We see the same pattern in how data moves between a CRM and the billing system, and I had a specific question about how you handle it.”

Watch out for: quoting something out of context and flattery that is clearly generated. If your reference could apply to any post on the same topic, it is not specific enough. Include a detail only someone who read the piece would know.

Source 4: A change in tools or technology

Public signals that a company adopted, dropped or migrated a tool: a job post that lists a tool, an announcement of a new integration, a case study on a vendor’s site, a changelog mention. Where to look: job posts, vendor case-study pages, the company’s own technology pages. What to extract: the tool and the evidence, with a date. Made-up example: “I noticed your team is moving to a new CRM, according to the posting for your RevOps role. Migrations usually leave a pile of duplicate and incomplete records. We build small automations that catch those after the move.”

Watch out for: guessing a tech stack from a scraped list. Tools that infer technology from website code can be wrong, and the cost of a wrong guess is that you sound careless.

Source 5: What customers and the community say publicly

Public reviews, community threads and customer comments about the company. Where to look: public review sites, forums, and the company’s public social accounts. What to extract: a theme, not a quote. Made-up example: “A few public reviews mention slow replies after hours. I do not know whether that matters for your business, but if it does, here is one way other teams handle it.”

Watch out for: using a complaint as a weapon. Criticizing a company to a stranger who runs it is a bad first impression. Phrase it as a question, never an accusation, or leave it out.

A rule for all five: the relevance test

Before a fact makes it into an email, it passes two checks. First, could the recipient verify it in ten seconds? If you cannot link to the source, do not use it. Second, does it lead to your offer without a leap? A fact the recipient might find interesting but that has nothing to do with what you sell is just a distraction. Dropping a fact is better than forcing a connection.

Get the Automation Opportunity Scorecard

A spreadsheet that ranks your repetitive tasks by what is worth automating first. List the task, rate how often it happens, how long it takes, how error-prone it is and how many tools it touches, and the sheet orders your list. Enter your email below and we will send the download link.

Please enable JavaScript in your browser to complete this form.

Name *

First

Last

Name Email

Email *

Submit

Research and drafting prompts to start from

Below are four prompts, one per AI step. They are starting points we wrote for this guide and have not tested on live campaigns. Expect to change them. They work with any capable language model, for example Claude (may be an affiliate link) or a model from OpenAI (may be an affiliate link). Replace the text in square brackets with your own information.

Three design choices run through all of them. They forbid the model from using anything that is not in the sources it was given. They give it a way to say “I found nothing” instead of inventing something. And they ask for structured output, so an automation can check it and route it.

Prompt 1: The research brief

You are a research assistant preparing notes for a salesperson. You will receive search results and page text about one company and one person.

RULES
- Use ONLY information that appears in the provided sources. Do not use your own memory of the company.
- Every fact must include the source URL and the date of the source. If the date is not shown, write "date unknown".
- Prefer facts from the last 12 months.
- Do not report anything private or personal: family, home, health, personal social media.
- If no useful fact is present, return {"facts": [], "verdict": "no verifiable fact found"}.

OUTPUT (JSON only)
{
  "company": "",
  "person": "",
  "facts": [
    {"type": "job_post | announcement | own_writing | tool_change | public_comment",
     "fact": "one plain sentence",
     "source_url": "",
     "source_date": "",
     "why_it_might_matter": "one sentence, or empty"}
  ],
  "verdict": "usable | weak | no verifiable fact found"
}

COMPANY: [company name and website]
PERSON: [name and role]
SOURCES:
[search results and page text go here]

The verdict field is what lets an automation make a decision. In n8n, a condition node can send “no verifiable fact found” and “weak” rows to a Skip status without any human involvement.

Prompt 2: The angle

You help a salesperson decide whether there is a good reason to write to a prospect.

OFFER (what we do, in one sentence): [your offer]
WHO IT HELPS: [your written target description]

Given the research brief below, propose at most TWO angles. An angle is: one fact from the brief, plus the situation it usually creates, plus one thing our offer can do about it.

RULES
- Use only facts from the brief. Cite which fact each angle uses.
- If no angle connects to our offer without a stretch, return {"angles": [], "recommendation": "skip"}.
- Do not flatter. Do not congratulate without adding something useful.

OUTPUT (JSON only)
{"angles": [{"fact_used": "", "situation": "", "connection_to_offer": "", "one_sentence_version": ""}],
 "recommendation": "write | skip"}

RESEARCH BRIEF:
[brief JSON goes here]

Prompt 3: The draft

Write a first-touch cold email from [sender name], [sender role] at [company], to [recipient name], [recipient role] at [recipient company].

USE ONLY: the chosen angle and the research brief below. Do not add any fact that is not in them.

STYLE
- Maximum 90 words in the body. Plain text, no formatting, no links in the first email.
- Write the way a person writes to a colleague they respect. Short sentences.
- Open with the specific fact, not with a greeting or a compliment.
- One idea. One small question at the end that can be answered with a few words.
- Subject line: 2 to 5 words, lowercase, no clickbait, no "quick question", no recipient name.

BANNED
- "I hope this email finds you well", "I came across your company", "impressed by", "game-changer", "leverage", "unlock", "revolutionize", "synergy", "reaching out", "just circling back", "10x".
- Any claim about our results, customers or savings.
- Any statement about the recipient's feelings, problems or intentions that is not in the brief. Use a question instead.

OUTPUT (JSON only)
{"subject": "", "body": "", "fact_used": "", "source_url": ""}

ANGLE: [chosen angle]
RESEARCH BRIEF: [brief JSON]

Prompt 4: The checker

You are a strict editor. Compare the draft email to the research brief and the rules.

For each sentence in the body, answer:
1. Is it supported by the brief, or is it a general statement, a question or a call to action?
2. Does it make any claim about the recipient or their company that is not in the brief?
3. Does it contain a banned phrase or a vague compliment?
4. Would it make sense if our company name were removed?

OUTPUT (JSON only)
{"unsupported_claims": [], "banned_phrases_found": [], "sounds_generic": true | false, "word_count": 0, "verdict": "pass | needs_edit | reject", "notes": ""}

DRAFT: [draft JSON]
RESEARCH BRIEF: [brief JSON]

Keep the checker’s verdict visible to the human reviewer, but do not treat a “pass” as permission to skip the review. A checker catches unsupported claims and banned phrases. It does not catch tone, timing or a message that is technically accurate and still a bad idea.

What to do when the prompt output is not good enough

• Too generic: your angle prompt is too loose, or your target list is too broad. Tighten the list before touching the prompt.
• Too long: lower the word limit and add “cut the first sentence if it can be cut.”
• Invented details: check whether the model had a source for the detail. If not, add a stricter line to the research prompt and make the checker reject unsupported claims.
• All emails sound the same: vary the angle types across the batch and keep a small library of your own sentences, written by you, as style samples in the prompt.

Cold email templates and examples that read like a person wrote them

These cold email examples and templates are starting points, not proven scripts. They show structure: a specific opening, one idea, one small question. Their value is in the pattern, so do not paste them. Every name, company and fact in the examples is invented.

A few rules that apply to all of them. Keep the first email short enough to read on a phone without scrolling. Put the fact first. Ask a question that is easy to answer. Leave out links and attachments in the first message, which also helps deliverability. Write the subject line the way a colleague would: short, lowercase, specific to the message, never a trick.

Template 1: The job post

Template 2: The announcement

Template 3: Their own writing

Template 4: A tool change

The follow-up, kept short and different

A follow-up should add something, not repeat the first email. If you have nothing new to add, the follow-up is a reminder, and reminders are what people mark as spam. Examples of what to add: a relevant fact you found since, a one-line answer to the question you asked, or a direct, polite way to say no.

What to avoid, side by side

Sending rules: platform, volume, timing and follow-ups

Good research can be undone by careless sending. This section covers the practical rules: how to choose a sending setup, how much to send, when, how often to follow up, and what the law requires of every commercial email.

Choosing a cold email platform or setup

You can send from a normal mailbox, or use a tool built for outbound email. Examples of such tools include Instantly (may be an affiliate link), mentioned here as an example of the category. We have not evaluated it and this is not a recommendation. Whatever you pick, judge it against criteria that match this workflow:

• Human approval fits: can you load approved drafts and send them yourself, or schedule only what a person approved?
• Opt-outs are handled automatically and applied across every campaign, so nobody gets a second message after saying stop.
• Reply detection: a reply, an out-of-office and a bounce should stop the sequence for that person.
• Limits you control: a daily cap per mailbox that you set, not one that pushes you to send more.
• An API or webhook so your n8n workflow can write status back to your sheet.
• A clear export of your data, so you are not locked in.
• Price you can carry: a flat monthly fee that you can stop. At twenty emails a day you do not need a plan built for ten thousand.

At twenty reviewed emails a day, a plain mailbox plus a sheet is enough to start. A dedicated platform becomes worth its cost when you need multi-mailbox rotation, built-in unsubscribe handling or reporting you cannot easily build yourself. Start with the simplest setup and move up when you hit a real limit.

Authenticate your domain and keep complaints low

According to Mailsuite’s summary of Google’s sender guidance (read on November 3, 2026), all senders to Gmail should set up SPF and DKIM for their sending domain, use valid forward and reverse DNS and a TLS connection, and keep the spam complaint rate below 0.30%, with 0.10% as the figure to aim for. Senders of more than 5,000 messages a day to Gmail face stricter rules: authentication with SPF, DKIM and DMARC and one-click unsubscribe. The same summary states that senders below that volume are not required to include an unsubscribe link, but recommends one.

Two cautions. First, that summary is not Google’s own page, and it contains internal inconsistencies about the thresholds, so read Google’s current sender guidelines yourself before you rely on a number. Second, even though you are far below the bulk threshold, include a simple way to opt out in every message. The law requires it (see the box below), and it protects your sending reputation.

Volume: a conservative ramp

Mailbox providers pay attention to sudden changes. A mailbox that has sent ten emails a month and then sends two hundred looks different from one that grows steadily. There is no single correct ramp, and we found no source that gives a safe number for every sender. This is a conservative policy to use as a starting point:

Notice that the cap is reviewer capacity, not a provider limit. If you can only review fifteen drafts properly in a day, fifteen is your number.

Timing

Advice on the best day and hour to send cold email is everywhere, and most of it comes from vendors with their own data and their own lists. We did not find a primary source we could trust. A reasonable default is a weekday morning in the recipient’s time zone, since that is when business email is read. Then test it yourself: send half of a batch in the morning and half in the afternoon, and compare qualified replies over several weeks. With small volumes, a difference of one or two replies is noise, so do not change the policy after a single week.

Follow-ups: few, spaced out, and with a clear exit

A person who has not replied to the first email may have missed it, may be busy, or may not care. A well-judged follow-up helps with the first two. It does nothing for the third, so it needs to be easy to refuse. A policy to start with:

• At most two follow-ups, three touches in total, spread over about three weeks.
• Each follow-up adds something: a new fact, an answer to your own question, or a way to say no.
• The last message is a polite exit. “I will not write again unless you ask” is a legitimate, respectful ending.
• Stop immediately on any reply, even a negative one, and on any opt-out.
• Do not re-enter a contact into a new sequence for several months, and not at all after an opt-out.

This is a policy, not a measured optimum. We looked for a source for how many touches a cold sequence should have and did not find a reliable one. Your first month of data will tell you more. If your sequence runs on warm leads rather than cold prospects, the cadence in our missed-lead follow-up guide uses a different logic, because those people already know you.

What the law requires of commercial email in the US

LinkedIn: a different set of rules

LinkedIn is governed by its own terms, not by CAN-SPAM. Its help center has a page titled “Prohibited software and extensions” and another on automated activity. We could not open either when preparing this guide, so read them on LinkedIn (may be an affiliate link) yourself. The practical advice for this workflow is simple: use the research to write messages, but send them by hand from your own account. A tool that logs in as you and sends messages in bulk puts your account at risk, and an account restriction costs more than any campaign gains. The LinkedIn workflow on this site never sends on its own for the same reason.

Verify addresses and keep a suppression list

• Verify addresses before you send. Use a verification service or confirm through a reliable source. Bounced emails hurt your reputation, and guessed addresses bounce.
• Keep a suppression list: every opt-out, every bounce and every “please do not contact me” goes on it, and the workflow checks it before every send.
• Record where each address came from. If anyone asks how you found them, you can answer.

Measuring qualified replies, not volume

Emails sent is the easiest number to track and the least useful. It measures effort. The number you want is qualified replies per week: replies from the right person that show real interest or a real fit, such as a booked call, a question about how it works, or a referral to the right person.

Define the terms before you count

Write the definition of “qualified” down before the first send and do not change it afterwards. If you redefine it every time a week looks bad, the number loses meaning.

A tracking sheet you can set up in ten minutes

Add these columns to the same Google Sheet that runs the workflow: date sent, template used, source type (job post, announcement, own writing, tool change, public comment), angle used, reviewer edits (none, light, heavy), reply (yes or no), reply type (positive, question, referral, not now, no, opt-out), qualified (yes or no), and days to reply. Then a weekly summary tab computes counts and rates by source type and by template.

What small numbers can and cannot tell you

At twenty emails a day you will send about a hundred a week. That is enough to see large differences and not enough to see small ones. If one source type brings three qualified replies in forty emails and another brings one in forty, that is suggestive, not proof. Keep going for several weeks before you drop anything. And pay attention to the reviewer edits column: if heavy edits correlate with better replies, your prompt is not yet doing enough of the work.

A weekly ten-minute review

• Read every reply and opt-out, yourself, not only the counts.
• Check bounces and complaints and remove the sources that caused them.
• Write one sentence: what we learned this week, and one thing to change next week.
• Change only one variable at a time (a prompt, a template, a source), so you know what caused what.

If you want the summary tab built and the weekly numbers pushed to a message each Monday, that is a small automation. You can describe it on the Request a Custom Workflow page.

A fourteen-day launch plan

The whole thing looks large. It is not, if you start with the smallest version and automate only after you have done it by hand. Each day below fits in a short work block.

If a step feels too big, do the first ten minutes: open the sheet and write the target description. The rest becomes easier once the first line exists.

When it is time to hand it to someone

You can run all of this yourself with a spreadsheet, a mailbox and an AI chat tool. Many people do, at least for the first month. At some point the manual parts stop being worth your time: copying search results into a prompt, moving drafts between tabs, remembering who was contacted. That is when a small, specific workflow pays for itself.

At LeenLabs we build n8n workflows for exactly that kind of task. The process is deliberately plain. You describe the task once on a single form, including your sheet, your CRM and your mailbox. We review it and send a quote. You pay 40% to start, and the remaining 60% when we deliver the workflow. There is no bidding round to run and no profiles to compare. We do not promise results from your outreach, because those depend on your offer, your list and your market. What we deliver is a workflow that does what the request describes.

If you want to learn how these workflows are put together first, browse the free workflows, which come with JSON you can import, or read more about us.

Frequently asked questions

Can AI write cold emails that get replies?

AI can draft cold emails, and some of them will get replies. What decides the result is the research, the list and whether a person reviews the message. A model with no real facts to work from will produce the generic email the recipient has seen a hundred times. This guide uses AI for research and first drafts and keeps a person responsible for what is sent.

How do I write a personalized cold email at scale without sounding like AI?

Limit the scale to what a person can review, and personalize on specific, recent, verifiable facts: a job post, an announcement, something the person wrote, a tool change, a public comment. Ban the usual generic phrases in the prompt, cap the length, and read every message before it goes out.

Is AI cold email legal?

In the US, the FTC’s CAN-SPAM guide that we read applies to commercial email whoever or whatever wrote it, and it has no exception for business-to-business messages. We did not find a rule in that guide about AI authorship. Other countries and states have their own rules, and some are changing, so get legal advice for your own situation. This is general information, not legal advice.

How many cold emails should I send per day?

As many as a person can properly review, which for most people is around twenty. Ramp up slowly on a new mailbox, as in the table in Section 6, and stop raising volume if you see bounces or complaints. The right number is the one that keeps your qualified reply rate healthy, not your sent count high.

How many follow-ups should I send?

We found no reliable source for an ideal number. A conservative policy is two follow-ups spaced over about three weeks, each adding something new, ending with a polite exit. Stop at once on any reply or opt-out. Adjust after you have a few weeks of your own data.

What is the best AI cold email platform?

It depends on your volume and workflow, and we have not evaluated platforms against each other. Judge a platform on whether it lets a person approve messages, handles opt-outs across campaigns, detects replies and bounces, gives you control of your daily cap and lets you export your data. At about twenty emails a day, a normal mailbox and a sheet are enough to start.

Should I use LinkedIn automation for outreach?

Check LinkedIn’s own rules first. Its help center has pages on prohibited software and automated activity, and we could not open them from our environment. The safer pattern is to use AI for the research and drafting, and to send messages by hand from your own account.

What should a human check before an AI-written cold email goes out?

Four things: that every claim is true and has a source, that you would be comfortable if the recipient saw your research, that it sounds like a person, and that the ask is small and clear. If any answer is no, edit or reject the draft.

Sources

• Federal Trade Commission, CAN-SPAM Act: A Compliance Guide for Business. Read on November 1, 2026. Seven requirements, 10 business days to honor opt-outs, no exception for business-to-business email, penalty of up to $53,088 per violating email (the amount is adjusted periodically).
• Mailsuite, a summary of Google’s email sender guidelines (mailsuite.com/blog/google-sender-guidelines). Read on November 3, 2026. Second-hand: we could not open Google’s own page (support.google.com/a/answer/81126), and the summary gives inconsistent spam-rate thresholds (0.10% and 0.30%). Check Google’s current guidelines before relying on a figure.
• LinkedIn Help, “Prohibited software and extensions” and “Automated activity on LinkedIn.” Not read by us: both pages blocked our access when we tried. Read them directly before using any automation near LinkedIn.
• Not used: published cold-email reply-rate benchmarks and follow-up-count statistics. We could not find a primary source for them, and many come from vendors of outreach tools.

This article is general information and not legal, tax or financial advice. Laws and rules differ by country and state and change over time. Check with a qualified professional about your own situation.