---
title: "How to Make Your AI DMs Sound Like You (Not a Bot)"
slug: make-ai-dms-sound-like-you
author: "Leonardo Maldonado"
category: "AI Setter Operations"
articleType: how-to-guide
tags: ["make AI sound like you","AI DMs that sound human","train AI on your voice"]
publishedAt: 2026-08-03T09:00:00.000Z
updatedAt: 2026-08-13T09:00:00.000Z
canonical: https://setluca.com/blog/make-ai-dms-sound-like-you
---# How to Make Your AI DMs Sound Like You (Not a Bot)

> To automate Instagram DMs without sounding like a bot, pull 30-50 of your real sent replies and tag the patterns in them: length, openers, punctuation, emoji, how you say no. Feed that fingerprint to the AI, kill the named tells below, and keep every draft behind a human review queue.

## Key takeaways

- To automate Instagram DMs without sounding like a bot, extract a voice fingerprint from 30-50 of your own sent replies instead of describing your tone in a prompt.
- Ten named tells cause most robotic drafts, from the concierge greeting to the apology wrap around your price.
- Test with a blind read: shuffle ten drafts with ten real messages and see if someone who knows you can sort them.
- Generic costs money. McKinsey found 76% of consumers get frustrated when interactions aren't personalized.
- Copy your voice, not your worst habits. Typos, over-apologizing, and four-message essays stay out.

---

How do you automate Instagram DMs without sounding like a bot? You stop guessing at "your voice" and go measure it. Your last 50 sent DMs already contain the answer: how long your messages run, whether you capitalize, which two emoji you actually use, what you type when you turn someone down. That measured pattern is a voice fingerprint, and it's what the AI copies. Then you remove the habits that read as machine-written, and you keep a human on the approve button. Below: how to extract the fingerprint, ten before/after rewrites, how to test the result, and what happens when the voice drifts.

## Why do AI DMs sound like a bot in the first place?

Because the model was handed an adjective instead of evidence. Told to be "friendly and professional," it writes whole, polite sentences, which is how precisely nobody texts a lead at 11pm on a Sunday.

People notice the pattern even when they can't name it, and they're worse at naming it than they think. In a Twilio study reported by Customer Experience Dive, 90% of consumers failed to pick out AI-generated voice clips, while three-quarters believed they could spot AI text. So your leads are confident and wrong at the same time, which is the worst combination for you: a DM that feels slightly off makes them suspicious without giving them a reason they can check.

**Chart:** Horizontal bar chart. Consumers who believe they can spot AI in text: 75 percent. Consumers who failed to correctly identify AI-generated voice: 90 percent. (Source: Twilio via Customer Experience Dive, 2025.)

## How do you build a voice fingerprint from your own past DMs?

A voice fingerprint is a short sheet of measured facts about how you write, pulled from messages you already sent. Describing yourself doesn't work. "Casual but direct" produces the same draft for every coach on earth, because every coach on earth writes that. Counting works, because the counts are different for everybody. Nobody else has your ratio of questions to statements.

Scroll back through your lead threads and copy your own replies only, 30 to 50 of them, into one document. Take a spread: first replies, two or three price conversations, a follow-up to someone who went quiet, one where you turned somebody down, one where you were short because you were busy. Don't tidy them. Then count. Ten minutes with a highlighter gives you this:

| Signal | What to count | Example answer from a real sheet |
| --- | --- | --- |
| Opening move | Greeting, name, or straight into the point | 8 of 10 start mid-thought, no greeting |
| Message length | Median words, and messages per turn | Median 11 words, usually 2 back to back |
| Capitalization | Share starting lowercase | 68% lowercase |
| Punctuation | Full stops, ellipses, dashes, line breaks | Almost no full stops; line breaks instead |
| Emoji | Which ones, how often, where | Only 🙂 and 🙌, one per eight messages |
| Question habit | Share ending on a question | 74% end on a question |
| Signature phrases | Words you repeat that nobody else uses the same way | "gotcha", "no stress", "here's what i'd do" |
| How you say no | Your exact wording when you decline | "honestly not a fit for you yet, here's why" |
| How you name a price | Order of number, context, question | Bare number, then a question. No apology |
| Never-say list | Words that never appear once | "reach out", "circle back", "investment" |

Copy that table, replace the right column with your own counts, and paste it in as the AI's voice rules. Ten measured lines beat two paragraphs of adjectives. The [AI setter for coaches](/ai-setter) guide covers the pipeline this sits inside.

## Which ten tells make AI copy read as generic?

Most robotic drafts come from ten specific habits, and each one has a clean rewrite. The fix is never more polish. It's the small untidy moves a real person makes when they're typing with one thumb. Below is the same line twice: generic default first, coach's voice second.

**1. The concierge greeting.** Nobody opens a DM like a hotel front desk.

> Hello! Thank you so much for reaching out about my coaching program.

> hey -- saw your comment on the reel

**2. The restatement opener.** Repeating the lead's message back burns your first line.

> I understand that you're interested in learning more about one-on-one coaching.

> yeah, 1:1 is the one that'd fit

**3. The brochure dump.** The whole offer arrives before anyone asked for it.

> My program is a 12-week transformation covering training, nutrition and accountability, with weekly check-ins and 24/7 support.

> quick q before i explain it -- are you training already or starting from zero?

**4. The office vocabulary.** Email words wandered into a text thread.

> Feel free to reach out at your convenience and we can circle back.

> just shout when you know

**5. The enthusiasm tax.** Exclamation marks stacked up to fake warmth.

> Great question! Absolutely! I'd be delighted to help you with that!

> good q. yeah i can help with that

**6. The permission ask.** Asking to ask, instead of asking.

> Would you like to schedule a call to discuss this further?

> want me to send a couple of times for thurs?

**7. The apology wrap.** Hedging before the number tells them the number is a problem.

> I completely understand that budget is a consideration. The investment for the program is $1,800.

> it's $1,800 for the three months. does that land anywhere near what you had in mind?

**8. The balanced sentence.** Every clause weighted the same, forever.

> You'll build real strength, gain lasting confidence, and finally see the results you deserve.

> mostly you'd stop guessing what to do on training days

**9. The grammar tell.** Semicolons in a thread nobody proofreads.

> That is a common concern; however, most clients find the schedule quite manageable.

> most people worry about that. it's 3 sessions a week, 45 mins

**10. The copy-paste opener.** The same first line sent to forty people.

> Hey! Thanks for the follow -- what are your goals right now?

> hey, the deload question -- how many weeks in are you?

Tell 3 is the expensive one. Explaining more is how you lose, in DMs as much as on calls -- across 326,000 recorded calls, Gong found the deals that closed averaged 57% rep talk time against 62% on the ones that didn't. The brochure dump is that gap, compressed into one message. Our post on [Instagram DM openers that get replies](/blog/instagram-dm-openers-that-get-replies) covers the first-line problem in more detail.

## What shouldn't you copy from your own DMs?

Your samples contain your voice and your bad days in the same file. Four things stay out.

- **Real typos.** Deliberate lowercase reads as casual. "teh" and "recieve" read as careless, and at scale they read as broken. Keep the loose grammar, drop the misspellings.
- **Over-apologizing.** If half your samples open with "so sorry for the slow reply," the AI will apologize for a reply that took nine seconds.
- **Four-message essays.** Most coaches have a few threads where they wrote a wall. Leave those out unless walls are genuinely your normal.
- **Loose claims.** Any sample where you promised a result, a timeline, or a body-composition number teaches the model to promise it again. Put "guarantee" on the never-say list.

Strip client names and health details too. Training samples come from lead conversations, not from people who told you something private.

## A worked example: Priya, a mindset coach

Priya is a composite, built from patterns coaches describe rather than one person. She's in Lisbon, sells a three-month 1:1 container at EUR 1,800, and gets 30 to 50 DMs a week off Reels about burnout. Her fingerprint, pulled from 42 sent replies: median 11 words, 68% start lowercase, only 🙂 and 🙌, roughly three-quarters of her messages end on a question, and she names her price as a bare number with no cushion.

Below is the draft before the fingerprint went in, against the one after. The thread is illustrative, and Priya approved every message before it sent.

> **Lead:** hey, saw your reel about the sunday dread thing. that's literally me every week

Before, with tone described as "warm and professional":

> **Draft:** Hello! Thank you so much for reaching out. I'm so glad the reel resonated with you! Sunday anxiety is something many of my clients struggle with. My three-month coaching container is designed to help you build lasting mental resilience. Would you like to schedule a call to discuss further?

After, with the fingerprint sheet in place:

> **Draft (approved):** oh the sunday one gets everyone >is it the week ahead you're dreading, or the job specifically? different problem depending

> **Lead:** the job i think. i've been saying i'll leave for two years lol

> **Draft (approved):** two years is a long time to be almost-leaving >what would have to be true for you to actually go?

The second version asks before it pitches, and it runs at Priya's real length. That's the whole difference, and it's worth what it's worth: **81%** of people ignore messages that aren't relevant to them, per Attentive's consumer research. A templated DM is an irrelevant message with your name on it. Priya still reads every draft before it sends. What changed is that she now approves most of them instead of rewriting them.

## How do you test whether it actually sounds like you?

Run three tests before you trust it, and don't skip to the last one.

**The blind read.** Take ten AI drafts and ten messages you genuinely sent. Shuffle them, strip the labels, and hand the list to someone who knows how you text: a long-term client, your partner, an old setter. Ask them to mark which ten are the AI. Chance is ten right. If they get fifteen or more, your fingerprint is missing something, and their reasons for each guess tell you what.

**The "would I have sent this" gate.** For each draft in the queue, ask whether you'd have hit send on that exact text, not whether the text is good. Change one word and it's fine. Change three and the voice hasn't landed. Log which three, because that's your next rule.

**The read-aloud test.** Say the draft out loud at your normal texting speed. If you stumble on a clause, you didn't write that clause and neither would you.

Then decide what to do with the result:

- **Fifteen or more correct, or you're rewriting most drafts:** keep review on for everything and add ten samples from the message types you keep rewriting.
- **Eleven or twelve correct, and one small tweak per draft:** keep review on, and start feeding your repeated edits back as rules.
- **Ten or eleven correct across two runs a month apart:** auto-send the low-stakes replies, a first hello or a scheduling nudge, and keep review on for price, objections, and anything emotional.
- **Any run where a draft made a claim you wouldn't make:** stop, fix the never-say list, re-test before anything else.

## Why should a human approve before send?

The review queue is what makes the message yours, which is why auto-send is off by default on every Luca plan. A draft you approved is a reply you sent, and that's a different thing legally and commercially from a bot posting in your name while you sleep.

It does two other jobs. Every edit you make is a correction the model learns from, so the queue is where the voice actually tightens -- you're not just checking work, you're teaching. And it keeps a person between a lead's message and your account, which is the shape platform pacing rules are built around anyway. [Luca's pricing](/pricing) includes the queue at every tier.

## How does the voice drift, and how do you pull it back?

Voice drift is real, and it's usually your doing rather than the model's. Four causes, in the order they turn up.

- **Approval autopilot.** After a few good weeks you stop reading and start tapping. Small generic phrasings slip through, and each approved draft confirms them.
- **Your own voice moved.** New offer, new price, new positioning. The samples describe the coach you were in March.
- **Channel bleed.** Samples pulled from WhatsApp or email carry a different register into Instagram. Keep a sheet per channel; the [multi-channel DM playbook](/blog/multi-channel-dm-playbook) covers where the registers diverge.
- **Edge-case sprawl.** The model meets message types your samples never covered, and fills the gap with default politeness.

Pull it back with a monthly audit: read twenty approved messages in a row, cold, and mark every line you wouldn't have written. More than four marked means refresh your samples. Re-run the blind read once a quarter and after any offer or price change, and keep a note of every edit you make twice. A repeated edit is a missing rule.

## Edge cases worth deciding now

- **The lead who's also a coach.** Peers read DMs professionally and clock a template instantly. Flag accounts that follow other coaching pages heavily and handle those yourself.
- **A lead writing in another language.** Your fingerprint is in English. Translated drafts lose the rhythm and gain formality, so route non-English threads to review with no exceptions.
- **The lead who references something only you'd know.** A call you had, a workshop they attended, a DM from six months ago. The model doesn't hold that memory and shouldn't improvise around it.
- **A voice note.** A tidy text reply to thirty seconds of audio reads as a machine. Answer with your own voice note or one short human line.
- **Two brands, one person.** A coaching account and a business account need two fingerprints. One sheet across both produces a blur that fits neither.

## Troubleshooting: what's actually broken

| Symptom | Likely cause | Fix |
| --- | --- | --- |
| Drafts sound like your captions, not your DMs | The model learned from public posts, written for strangers | Retrain on sent DMs only, and delete the caption source |
| Every draft opens the same way | Not enough opener variety in the samples | Add ten first-reply samples and note "rotate openers" as a rule |
| The voice is right but the drafts feel cold | Samples skew toward your busy, clipped replies | Add three or four samples from unhurried threads |
| Leads reply "is this a bot?" | Timing, not wording: a same-second answer at 3am | Hold sends to a human-plausible gap; see [is Instagram DM automation safe](/blog/is-instagram-dm-automation-safe) |
| The voice was fine, now it isn't | Approval autopilot, or your offer changed | Run the monthly audit and refresh samples against the current offer |
| Drafts are good until money comes up | No price samples in the set | Add three real price conversations, one where you held the number |

## Which mistakes keep the drafts sounding generic?

1. **Describing your tone instead of measuring it.** "Friendly, casual, professional" is the prompt every coach writes, and it produces the same DM for all of them.
2. **Cleaning the samples first.** Fixing capitalization and grammar before you paste them in trains a polished bot. The mess is the signal.
3. **Sampling only your best replies.** Twelve carefully written messages don't show the model what you sound like on a Tuesday with eight threads open.
4. **Approving on autopilot in week one.** The early edits carry the most information, and blind approval costs you the whole training signal.
5. **Treating a robotic draft as harmless.** Zendesk's 2025 CX Trends report found 63% of consumers will switch to a competitor after a single bad experience, up 9% year over year. One templated reply to a warm lead is a lost lead.

## Where should AI stop sounding like you?

AI can't fake a shared memory or a moment that genuinely matters to someone, and it shouldn't try. If a lead mentions a call you had, a loss they're carrying, or something they've never told anyone, the model wasn't there and doesn't know. A smooth reply lands worse than no reply.

The failure has a shape you'll recognize the second you see it. A lead writes "my dad passed in March and I've let everything go," and a perfectly-voiced draft asks what their training split looks like. The voice was right. The judgment wasn't there at all. And the lead is gone.

The recovery ladder runs in four steps. Route grief, health disclosures, and money trouble straight to you, before the model drafts. Keep auto-send off on any thread where a lead has mentioned something personal. If a tone-deaf message did send, reply yourself inside the hour and name it plainly. Don't explain the software; people forgive a human correction and won't forgive a bot defending itself.

Above that line the split is clean. The AI drafts, qualifies, and follows up in your measured voice. You take the threads where being a person is the whole point, which is also how you learn [how to close clients in the DMs](/blog/how-to-close-clients-in-dms) instead of outsourcing the part that closes. [What an AI DM setter is](/blog/what-is-an-ai-dm-setter) and [how to sell in the DMs without being salesy](/blog/how-to-sell-in-the-dms-without-being-salesy) are the two to read next.


## FAQ

### How do I automate Instagram DMs without sounding like a bot?

Pull 30-50 of your real sent replies and count the patterns: median length, opener style, capitalization, emoji, how you name a price. Feed those measured rules to the AI, remove the named tells like the concierge greeting and the brochure dump, and keep every draft behind a human review queue.

### How many message samples does the AI need to learn my voice?

Thirty to fifty sent replies is the working range for a fingerprint you can count patterns in. Fewer than thirty and your medians aren't stable. Cover a spread: first replies, at least three price conversations, a follow-up to a ghost, and one where you turned somebody down.

### What are the biggest tells that a DM was written by AI?

The concierge greeting, the restatement opener, the brochure dump, office vocabulary like "circle back," stacked exclamation marks, and an apology wrapped around your price. Over-explaining is the costly one: Gong's analysis of 326,000 calls found won deals average 57% rep talk time against 62% for lost ones.

### How do I test whether my AI DMs actually sound like me?

Run a blind read. Shuffle ten AI drafts with ten messages you genuinely sent, strip the labels, and ask someone who knows how you text to pick the AI ones. Ten right is chance. Fifteen or more means the fingerprint is still missing something they can name.

### Should I train the AI to copy my typos?

Copy the loose grammar, not the misspellings. Lowercase starts, missing full stops, and line breaks instead of commas all read as human. Actual typos read as careless at scale. Leave out over-apologizing, four-message essays, and any sample where you promised a result you can't guarantee.

### Won't leads get upset if they find out it's AI?

Not if you're honest and the reply is useful. You never deny AI involvement, since that breaks Meta's rules. Zendesk's 2025 CX Trends report found 64% of consumers are more likely to trust AI that shows warmth, and a draft you approved in your own voice rarely raises the question at all.

### Why does my AI voice drift back to sounding generic?

Four usual causes: you started approving without reading, your offer or price changed and the samples didn't, samples from another channel bled in, or the model met message types your set never covered. Run a twenty-draft audit monthly and refresh the samples when the offer moves.

### What's the one thing AI still can't do in DMs?

Handle genuine emotion or shared history. It doesn't know about the call you had last week or the loss a lead mentioned in passing. Route grief, health disclosures, and money trouble to yourself, and let the AI carry qualifying and follow-up where speed matters more than nuance.


## Sources

1. [Customer Experience Dive -- "Consumers can't tell the difference between humans and AI" (Twilio, 2025)](https://www.customerexperiencedive.com/news/consumers-cant-tell-the-difference-between-humans-and-ai/806709/)
2. [Attentive -- "New Global Study Reveals Consumers Demand More Personalization in Marketing" (2025 Consumer Trends Report, CITE Research)](https://www.attentive.com/press-releases/new-global-study-reveals-consumers-demand-more-personalization-in-marketing-81-ignore-irrelevant-messages-while-personalized-experiences-drive-loyalty-and-sales)
3. [McKinsey -- "The value of getting personalization right -- or wrong -- is multiplying" (Next in Personalization, 2021)](https://www.mckinsey.com/capabilities/growth-marketing-and-sales/our-insights/the-value-of-getting-personalization-right-or-wrong-is-multiplying)
4. [Gong -- "Talk-to-Listen Ratio" (analysis of 326,000 sales calls)](https://www.gong.io/blog/talk-to-listen-conversion-ratio)
5. [Zendesk -- "CX Trends 2025"](https://www.zendesk.com/newsroom/articles/2025-cx-trends-report/)

---

Published by SetLuca, the company behind Luca.