How do you write a cold email with AI that a human would reply to?
Write the inputs before you write the prompt: who the reader is, one fact about them with a URL, the problem that fact usually creates, what you do in one line, one proof you are allowed to use, and one ask with a day and time. Then tell the model what it may not do (no praise, no claims you did not give it, no filler openers, one ask, under 80 words), because a forbid list beats "sound human" every time. The draft that comes back is raw material: run a five-step edit (facts, swap test, tells, ask, length) and only then does it go out.
Updated
Why does an AI cold email read like an AI cold email?
The model writes the average cold email very well, and that is the one your buyer deletes. With no facts, it fills the hole with what it has seen most.
Buyers punish that. Gartner asked 632 B2B buyers (survey fielded August to September 2024, released June 25, 2025), and 73% of them said they actively avoid suppliers that send them irrelevant outreach. A generic AI draft is exactly that, just produced faster.
The tells are predictable:
| Tell | Why the model writes it |
|---|---|
| Pleasantry opener | Most emails in its training start that way |
| Compliment nobody checked | The prompt gave it nothing true to say |
| Trend sentence ("In today's competitive market...") | It needs a bridge and has no real implication |
| Invented numbers | Asked to persuade, with no proof supplied |
| Feature bullets | Bullets are how it shows thoroughness |
| Two or three asks | Hedging reads as polite to a model |
| Marketing vocabulary (leverage, seamless) | The most common verbs in the genre |
Each row is fixed twice: in the prompt, then in the edit.
What do you give the model before you ask for a draft?
The model is only as specific as your inputs, so fill a six-line sheet first. If a line has no source, leave it blank and tell the model it is blank.
| Input | What goes in the cell | Source | If empty |
|---|---|---|---|
| Reader | Role, what they own, short company name | Contact record | Stop: wrong list |
| Fact | One checkable thing about them, with the URL | Press page, job post, pricing page (sales trigger events) | No one-to-one AI email: send a segment email and say it is one |
| Problem | What that fact usually causes in their job, hedged | Your customers' stories | Cut the fact too |
| What you do | One sentence, no adjectives | Your own words | n/a |
| Proof | One result, one customer, permission confirmed | The customer | Leave it out |
| Ask | 15 minutes, a day, a time | Your calendar | n/a |
Filling the sheet is the research (time budgets). Paste nothing private or under NDA.
What do you forbid in the prompt?
"Write it like a human" has no test. "No sentence that could go to another company" does. Give the model seven concrete bans.
| Rule in the prompt | What it prevents |
|---|---|
| No fact I did not give you | Invented triggers and stats |
| No praise | The unchecked compliment |
| Open with the fact | The pleasantry |
| One ask, a question, with a day and time | Hedged double asks |
| No bullets, links or attachments | The feature dump |
| Under 80 words | Scrolling on a phone |
| Plain verbs, no marketing words | The vocabulary tell |
On the 80-word rule: Instantly's 2026 benchmark, built on its own platform data from January 1 to December 18, 2025, found that its elite performers (the top 10% of senders by reply rate) keep first-touch emails under 80 words on average. It is Instantly's own data and shows a pattern among its users, not a cause and not a market average. Our reason for the cap is simpler: less room for the model to pad. Cutting order is in how long a cold email should be; the full anatomy is in cold emails that work.
What does the prompt look like?
Fill the brackets from your input sheet, nothing else.
Write a cold email to [READER: role, what they own, short company name].
Use only these inputs:
Fact: [ONE CHECKABLE FACT] (source: [URL])
Problem it usually causes: [ONE SENTENCE, hedged]
What we do: [ONE SENTENCE]
Proof: [ONE RESULT, ONE CUSTOMER] or "none"
Ask: 15 minutes on [DAY] at [TIME]
Rules: no greeting line before the first sentence. No praise. No fact, number or claim I did not give you. One ask, as a question, with the day and time. No bullets, links or attachments. Plain verbs; never use leverage, seamless, transform, unlock, robust. Under 80 words. Give me three versions: the first opens with the fact, the second with the problem, the third with the proof (if Proof is "none", make the third a shorter version of the first).
Three angles, you pick one. The prompt is ours and uses the role, input, output, rules shape from AI prompts for sales prospecting.
What does the human edit pass look like?
The prompt gets you a draft; the edit wins the reply. Run the steps in order, because a cut fact changes the length.
| Step | What you check | Pass test |
|---|---|---|
| 1. Facts | Every claim traces to a sheet cell | Zero unsourced claims |
| 2. Swap test | Replace their company with one of its competitors | Any sentence that still works is generic: cut it |
| 3. Tells | Run the seven-row table | Zero present |
| 4. Ask | Exactly one, a question, with day and time | Answerable in one word |
| 5. Length | Count the words | Under 80, cutting from the sender's side first |
The model can help with steps 1 and 2 if it only lists and never rewrites: ask which input each claim came from, then which sentences survive the competitor swap.
Cut anything about you beyond the offer and the proof, anything not on the sheet, and every adjective that is not a fact. Our guidance on time: the sheet 3 to 15 minutes by tier, the prompt a minute, the edit two to three. If the edit outlasts the sheet, the sheet was thin.
What does a bad AI draft look like next to the edited email?
Everything below is invented, for illustration. Lena sells Tallyard, software that matches marketplace payouts to orders. Priya is VP Finance at Mossbrook Home, a roughly 60-person home goods brand whose press page says it went live on a second marketplace in August. Customer Fenmoor Goods closed its books three days faster the quarter after switching, and agreed to be named.
The lazy prompt Lena typed: "Draft an outreach email for a VP of Finance at an ecommerce brand. We sell Tallyard, payout reconciliation software."
Draft one (122 words, greeting and sign-off included):
Subject: Transform Your Financial Operations with Tallyard
Hi Priya,
I hope you're having a great week! Love what the Mossbrook Home team is building in the home goods space!
In today's competitive ecommerce landscape, finance teams are under more pressure than ever to close the books quickly and accurately. That's where Tallyard comes in. Our cutting-edge platform leverages AI to seamlessly reconcile marketplace payouts, helping brands like yours:
- Reduce manual reconciliation time by up to 70%
- Eliminate costly errors
- Gain real-time visibility into cash flow
I'd love to schedule a quick 30-minute demo to show you how Tallyard can transform your month-end close. Are you available sometime next week? Alternatively, feel free to check out our website for more information.
Looking forward to hearing from you!
Best regards, Lena
All seven tells from the table, zero facts about Mossbrook. With the sheet filled (the August marketplace launch, the hand-matching problem, Fenmoor Goods, 15 minutes Tuesday at 11), the edited email is 74 words, greeting included:
Subject: payouts from the new marketplace
Hi Priya, your press page has Mossbrook live on a second marketplace since August.
When a brand adds a channel, finance usually matches two sets of deposits to orders by hand, and month-end slips.
Tallyard matches marketplace payouts to orders automatically. Fenmoor Goods, a home goods brand about your size, closed three days faster the quarter after switching.
Could you spare 15 minutes Tuesday at 11 to see if your close could move too?
What changed: ad subject to trigger, pleasantry to sourced fact, three bullets to one sentence, the model's "up to 70%" to one real customer, two asks to one. Our math: 122 - 74 = 48 words cut, and 48 / 122 = 39% shorter; tells 7 to 0.
When should you not let AI write the email?
When should you skip the AI draft entirely?
- Your top accounts. You already spent up to 15 minutes on the sheet; writing 74 words yourself beats editing 122. Use the model for the fact check and swap test only.
- An empty fact cell. No fact, no one-to-one email. Send an honest segment email instead, and never let the model paper over the gap.
- When AI is about to raise your volume. A 30-second draft tempts a rep to send 300. Google tells senders to keep the Postmaster Tools spam rate below 0.10% and never reach 0.30%; per its FAQ, since June 2024 bulk senders over 0.3% lose eligibility for mitigation until they stay under 0.3% for 7 consecutive days. Identical-shaped mail at volume is how you get there. See spam complaint rate and why cold emails go to spam.
Unsure which angle to send? Split test with A/B testing cold emails and judge by positive replies.
Sources
Research and vendor pages accessed October 6, 2026.
- Gartner Sales Survey Finds 61% of B2B Buyers Prefer a Rep-Free Buying Experience
- Instantly, Cold Email Benchmark Report 2026
- Google, Email sender guidelines
- Google, Email sender guidelines FAQ
- The input sheet, forbid list, edit pass, time budgets, prompts and the Tallyard example are our own guidance, invented for illustration.