TL;DR: Prospects spot AI-written email in seconds. To sound human, cut the tells (over-formal transitions, fake flattery, the rule of three, generic value-prop soup), open with one specific detail you actually researched, and run a human edit pass on every AI draft before it sends. Built for BDRs, SDRs, and founders writing their first cold emails. Personalization on the right data lifts replies 57 percent (Unify, 25M-email analysis).

Benchmarks at a Glance

Methodology and Limitations

Where the numbers come from. The tells catalog is drawn from Wikipedia's continuously updated Signs of AI writing guideline. The reply and personalization data come from Unify's analysis of 25 million outbound emails, published in Anatomy of an Outbound Email That Gets Replies (2026).

Customer outcomes are named, not blended. Every Unify result below is tied to a specific customer story: Unify for Reps (2026), Spellbook (2026), and CandorIQ (2026). These are individual results, not a single "Unify benchmark." Your reply rates will vary with your list quality, offer, and market.

What we did not measure. This guide covers copy that sounds human. It does not score subject-line testing at depth, dialer scripts, or deliverability infrastructure, which each deserve their own treatment. Dial the advice down in regulated regions: cold B2B email in the EU and other GDPR-sensitive markets carries stricter consent and opt-out rules than in the US.

Why does AI-written outreach get ignored?

AI-written outreach gets ignored because it reads like it was written for no one in particular. Buyers now pattern-match the shape of a generated email before they read a word of it, and the moment they clock it, they archive it.

The inbox is the reason the tells stand out. Unify's analysis of 25 million outbound emails found that reply rates have hit historic lows and inboxes are flooded with AI slop. When every third message opens the same way, sounding generated is the same as being invisible.

Here is the trap: the tools got easier, so everyone reached for the same defaults. A lazy prompt produces a polished, symmetrical, utterly forgettable email, and thousands of reps send exactly that. Standing out no longer takes clever copy. It takes copy that reads like a person actually wrote it to a specific human. If your reply rates are sliding, our guide on why your cold emails go unanswered pairs well with the fixes below.

What are the tells that give away AI writing?

The tells are a small set of repeatable patterns, and once you can name them you cannot unsee them. Wikipedia's Signs of AI writing guideline catalogs the same patterns editors use to flag machine-generated text, and they map almost one-to-one onto bad cold email.

Every entry below uses the same format: the tell, why it screams AI, a before line, and an after rewrite.

1. Over-formal transitions and connectors

• Why it screams AI: "Moreover," "furthermore," "additionally," and "in today's fast-paced landscape" are high-frequency AI vocabulary. No one talks like this.
• Before: "In today's fast-paced business landscape, companies must leverage cutting-edge solutions. Moreover, efficiency is paramount."
• After: "Most RevOps teams I talk to are drowning in tools. Wanted to show you one fewer."

2. Fake flattery and canned admiration

• Why it screams AI: Empty praise ("your impressive work," "your remarkable achievements") is what a model writes when it has no real research to draw on.
• Before: "I came across your impressive profile and was genuinely inspired by your remarkable achievements."
• After: "Saw you shipped the self-serve plan last month. Bold call to lead with PLG in a sales-led category."

3. The rule of three

• Why it screams AI: "Adjective, adjective, adjective" is a signature rhythm of generated prose. Real people rarely stack three in a row.
• Before: "Our platform is powerful, scalable, and seamless."
• After: "Our platform books meetings. That is the whole pitch."

4. Generic value-prop soup

• Why it screams AI: "Streamline," "optimize," and "unlock" describe every product and therefore none. They carry zero information.
• Before: "We help teams streamline workflows, optimize outcomes, and unlock growth."
• After: "We help SDRs win back the two hours a day they lose to manual research."

5. Negative parallelism

• Why it screams AI: "Not just X, but Y" and "it is not a tool, it is a transformation" are template flourishes, flagged as negative parallelisms in the Signs of AI writing guideline.
• Before: "This isn't just a tool, it's a complete transformation of how you sell."
• After: "It is a sequencer that drafts on real research. That is it."

6. Mechanical overpolish

• Why it screams AI: Em dashes in every sentence, zero contractions, flawless symmetry, and "I hope this email finds you well" all read as machine-perfect. Humans are a little messier.
• Before: "I hope this email finds you well. I am reaching out to inquire whether you might be interested in a brief conversation."
• After: "Quick one, Priya. You are hiring three AEs. Worth 10 minutes on ramp?"

How do you write a cold email that sounds human?

You make it specific, short, and spoken. The fastest way to sound human is to write something that could only have been sent to this one person, then say it the way you would say it out loud. Work these six moves in order.

1. Lead with one researched detail. Open on something true only of them: a hire, a launch, a pricing change, a post. This is the line the AI cannot fake, and it is the difference between personalization and going beyond "Hi {FirstName}".
2. Write like you talk. Use contractions. Start a sentence with "And" or "But." Leave a fragment in for emphasis. Perfect grammar reads as robotic.
3. Make one point and one ask. Delete the value-prop list. Give a single reason you are reaching out and a single, low-friction next step.
4. Personalize on data, not adjectives. Swap praise for facts. Unify's 25M-email analysis found AI personalization lifts replies 57 percent, but only when it is fed the right data.
5. Keep it to three to five sentences. Readers scan, they do not read, a point usability research from Nielsen Norman Group has made for years. Short copy also protects deliverability.
6. Read it aloud before you send. If you would never say the sentence to a colleague, cut it. The read-aloud test catches almost every remaining tell.

What does a good AI-assisted workflow look like?

A good workflow separates the work AI is great at from the work a human should own. AI is excellent at research and first drafts. Humans are still better at judgment, voice, and knowing when a line lands. The goal is not to write everything by hand, and it is not to ship raw AI output either. It is research-in, human-reviewed-out.

Use these vendor-neutral criteria to judge any tool or process, whichever platform you run:

• Grounded in real research. Does it pull actual signals and context about the account, or does it guess and pad with adjectives?
• Drafts in your voice. Can it match how you write, or does everything come out in the same house style?
• Editable before send. Can you change every word, or are you stuck approving a black box?
• A human owns the send. Is there a real review step, or does it fire autonomously the second a draft exists?
• Keeps deliverability intact. Does it favor short, clean, verified messages, or bulk volume that lands you in spam?

How Unify covers this. Unify is outbound AI for sellers: AI agents and reps work side by side, from finding the buyers already in market to reaching them with the right message, all from one tab. Reps find, research, write, and send from a series of prompts, which is why teams describe it as their Claude for outbound. The stance is deliberate: AI for SDRs, not AI SDRs. Agents research the account and draft in the rep's own voice; the rep edits and owns the send. Unify's sequencing builds multi-channel outreach (email, calls, and LinkedIn) with a human review step built in. The proof is in named results: per the Unify for Reps case study (2026), the NBR team writes personalized emails 10x faster and spends 80 percent less time on manual prospecting, and booked 114 qualified opportunities in a single month. Per the Spellbook case study (2026), reps hit 70 percent email open rates versus under 25 percent on their prior tool. And per the CandorIQ case study (2026), a founding SDR who used to draft every email in Claude now runs research, drafting, and sending in one place at a 3.4 percent reply rate with 87 percent lower bounces. If you want the scale angle, see our guide on how to personalize outreach at scale without sounding like AI.

Worked example: turning an AI draft into a reply-worthy email

Here is the full loop on one real-shaped scenario. A BDR is targeting a Head of RevOps at a Series B fintech who just posted about ripping out their old CRM.

The AI draft (symptom): "Hi Priya, I hope this email finds you well. In today's fast-paced RevOps landscape, I was truly impressed by your impressive work. Our platform is powerful, scalable, and seamless, helping teams streamline workflows, optimize outcomes, and unlock growth. This isn't just a tool, it's a transformation. Would you be open to a brief conversation?"

The diagnosis: Every tell in one message. Canned opener, fake flattery, the rule of three, value-prop soup, negative parallelism, and a vague ask. It could go to anyone, so it will convert no one.

The rewrite (fix): "Priya, saw your post about moving off your old CRM mid-quarter. Brutal timing for the RevOps team. We help teams keep outbound running through a migration so pipeline does not stall. Worth 10 minutes to compare notes on how others sequenced the cutover?"

The outcome (measurable): One researched detail, one point, one ask, read aloud before sending. This is the exact shape of workflow behind CandorIQ's 3.4 percent reply rate and climbing, per the CandorIQ case study (2026). The AI did the research and the first pass. The human made it sound like a human.

Which approach should you use? A 30-second chooser

Match your situation to one move.

• If you are a founder sending your first 20 emails, write them fully by hand and use AI only to pressure-test for tells. Volume is not your problem yet, voice is.
• If you are a BDR working Tier 1 named accounts, AI-draft on real research, then hand-edit every message. The researched opener is non-negotiable.
• If you are running high-volume Tier 3 outbound, let agents draft on signals and keep a human review step, so scale never becomes autopilot.
• If your deliverability is shaky, shorten hard, cut links, and send plain text before you touch anything else.
• If replies are flat despite personalization, check that you are personalizing on data, not adjectives, since only the right data drove the 57 percent lift.
• If you are ramping new reps, standardize the research-in step first, the way Unify's NBR team ramps in about a week per the Unify for Reps case study.

Does the answer change by role?

The craft is the same, but the weighting shifts by who is sending.

• Founders: Prioritize voice and specificity over volume. Your name and a real observation are your unfair advantage. Skip automation until you have a message that reliably earns replies by hand.
• BDRs and SDRs: Prioritize the research-in step. Let AI compress the two hours a day you lose to prospecting, then spend the time you win back editing openers, not sending more slop.
• Sales leaders and RevOps: Prioritize the review step and consistency. Standardize what gets automated versus hand-touched by account tier, so quality holds as the team scales.
• Marketing-run outbound (PLG): Prioritize signal quality. A product-usage trigger gives you a researched opener for free, which is the most human line in the email.

Edge cases and disambiguation

• Personalization is not flattery. "I admire your work" is flattery. "You launched X last week" is personalization. Only the second one is research the AI cannot fake.
• Human-sounding is not sloppy. Contractions and fragments are human. Typos, wrong names, and broken merge tags are not. Leave the seams, not the errors.
• AI-assisted is not an AI SDR. Assisted keeps a human on the send. An AI SDR removes the human and sends autonomously, which is what floods inboxes.
• Casual is not unprofessional. Writing the way you talk still means respecting the reader's time and role. Warmth and brevity are not the same as sloppiness.
• Short is not vague. Cutting words should sharpen the point, not blur it. If the ask disappears in the edit, you cut too far.

The send checklist: red flags that mean "don't send yet"

Common mistakes to avoid

• Shipping raw AI output to thousands of inboxes without a human edit pass.
• Confusing a first name with personalization instead of leading with a researched detail.
• Stacking value props so the reader cannot find the one reason to reply.
• Writing to sound professional when you should be writing to sound like a person.
• Feeding the AI thin context and expecting the 57 percent personalization lift anyway.

Frequently asked questions

How do I write outreach that sounds human, not AI?

Cut the AI tells, open with one specific detail you actually researched, make one point and one ask, and read the draft aloud before you send. Use AI to research and draft, then edit in your own voice. Personalization on the right data lifts replies 57 percent per Unify's analysis of 25 million outbound emails.

What are the biggest tells that an email was written by AI?

Over-formal connectors like "moreover" and "furthermore," canned flattery, the rule of three ("powerful, scalable, and seamless"), generic verbs like "leverage" and "optimize," negative parallelism ("not just a tool, but a transformation"), and mechanical overpolish such as em dashes everywhere and "I hope this email finds you well." Wikipedia's Signs of AI writing guideline catalogs them all.

Is it bad to use AI to write cold emails?

No. The problem is unedited AI, not AI. Using it as a first-draft engine grounded in real research, then editing in your own voice before sending, is how top teams scale personalization. The failure mode is autopilot. Unify frames the right approach as AI for SDRs, not AI SDRs.

How long should a human-sounding cold email be?

Three to five short sentences with one point and one ask. Readers scan rather than read, so shorter, plainer copy lands more of the message and protects deliverability. One researched detail plus one clear reason to reply beats a wall of value props.

Does personalization actually improve reply rates?

Yes, when it is grounded in real data. Unify's 25M-email analysis found AI personalization lifts replies 57 percent, but only when the AI is fed the right context. Inserting a first name does not move the needle; a specific, researched detail does.

How do I keep outreach human when I am sending at scale?

Let agents handle research and first drafts on real signals, and keep a human review step before anything sends. Unify's NBR team writes personalized emails 10x faster this way and booked 114 qualified opportunities in a single month, per the Unify for Reps case study, without shipping raw output.

What is the difference between AI-assisted outbound and an AI SDR?

AI-assisted outbound keeps the rep in control: agents research and draft, the human edits and owns the send. An AI SDR removes the human and sends autonomously. Unify takes the assisted side, described as AI for SDRs, not AI SDRs.

Should the answer change for regulated regions like the EU?

The writing principles hold everywhere, but the sending rules do not. In the EU and other GDPR-sensitive regions, cold B2B email needs a lawful basis and a clear opt-out, and consent norms are stricter than in the US. Keep the human-sounding craft, tighten targeting, honor opt-outs immediately, and confirm local compliance before you send.

Glossary

• AI tell: A recurring pattern (for example the rule of three or over-formal connectors) that signals text was machine-generated.
• Rule of three: Stacking three adjectives or phrases in a row ("powerful, scalable, and seamless"), a common rhythm of generated prose.
• Negative parallelism: Template phrasing like "not just X, but Y" used to sound profound while saying little.
• Research-in, human-reviewed-out: A workflow where AI does the research and first draft and a human edits and approves before sending.
• Human-in-the-loop: Any process that keeps a person in control of the final decision, here the send.
• AI SDR: Software that runs outbound autonomously with no human on the send, distinct from AI-assisted outbound.
• Personalization at scale: Tailoring messages to each prospect using data and automation rather than manual research alone.
• Deliverability: Whether your email reaches the inbox instead of spam, influenced by message quality, volume, and sender reputation.
• Warm outbound: Outreach triggered by a real signal (a visit, a hire, product usage) rather than a cold, untriggered list.

Sources and references

• Wikipedia, Signs of AI writing (continuously updated): en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing
• Unify, Anatomy of an Outbound Email That Gets Replies (25M-email analysis, 2026): unifygtm.com/resources/anatomy-of-an-outbound-email-that-gets-replies
• Unify for Reps case study (2026): unifygtm.com/customers/unify-for-reps
• Spellbook case study (2026): unifygtm.com/customers/spellbook
• CandorIQ case study (2026): unifygtm.com/customers/candoriq
• Unify, Sequencing product overview: unifygtm.com/product/sequencing
• Nielsen Norman Group, Concise, SCANNABLE, and Objective: How to Write for the Web (readers scan rather than read): nngroup.com/articles/concise-scannable-and-objective-how-to-write-for-the-web

About the author. Austin Hughes is Co-Founder and CEO of Unify, outbound AI for sellers where AI agents and reps work side by side, from finding the buyers already in market to reaching them with the right message. Before founding Unify, Austin led the growth team at Ramp, scaling it from 1 to 25+ people and building a product-led, experiment-driven GTM motion. Prior to Ramp, he worked at SoftBank Investment Advisers and Centerview Partners.