--- title: "How to Train AI to Reply in Your Voice: 5 Steps" slug: how-to-train-ai-to-reply-in-your-voice author: "Leonardo Maldonado" category: "AI Setter Operations" articleType: how-to-guide tags: ["train AI on your voice","AI that sounds like you"] publishedAt: 2026-09-11T09:00:00.000Z updatedAt: 2026-09-11T09:00:00.000Z canonical: https://setluca.com/blog/how-to-train-ai-to-reply-in-your-voice ---# How to Train AI to Reply in Your Voice: 5 Steps > To train AI to reply in your voice, pick 25-30 of your real DM replies across eight message types, write them up as a short voice spec, paste in eight to ten example exchanges, then correct every draft in Luca's review queue. Promote each repeated edit into the spec. Over two weeks the drafts stop needing fixes. ## Key takeaways - Training AI on your voice starts with 25-30 real DM replies chosen for coverage, not a generic prompt or a personality slider. - The voice spec is a written artifact: length, openers, punctuation, emoji, how you say no, how you handle price, and a never-say list. - Eight to ten worked exchanges beat another paragraph of rules. Google DeepMind's 2024 many-shot research found gains kept coming as examples were added. - The review queue is where most of the learning happens. Any edit you make twice becomes a spec line. - One honest limit: AI can't invent a shared memory it was never told about. Those threads stay yours. --- How do you train AI to reply in your voice? You give it evidence of how you actually talk, then keep correcting it until the drafts match. There's no magic prompt. Four things do the work: real DMs chosen for coverage, a written voice spec, a handful of worked examples, and the edits you make when a draft misses. Cleaning up robotic tells afterwards is a separate job. Below: which DMs to pick, the spec template, eight before/after pairs, and the loop that tightens it. ## What does training AI to reply in your voice actually mean? It means the drafts read like something you'd type, so a lead would swear you wrote them. Same rhythm, same slang, the same way you name a price. A generic reply reads as a generic offer, and people notice: 88% of consumers say they're more likely to buy when the experience feels personal in the moment, while only 44% of brands manage it (Twilio, 2025). Training gives you three things you can actually hold: a set of chosen messages, a short spec, and a bank of examples. Measuring your own patterns is the other half, and [make AI DMs sound like you](/blog/make-ai-dms-sound-like-you) covers the counting. ## Step 1: Which past DMs should you train on? Pick 25-30 of your own sent replies from the last 90 days, chosen for coverage rather than quality. Nearly half of sales pros already use AI to write outreach (HubSpot, 2024), so having a model isn't the edge. What you feed it is. Coverage means one real example of every situation the AI will meet. Most coaches paste in thirty first replies, then wonder why the price drafts are bad. | Message type | How many | Why it earns a slot | | --- | --- | --- | | First reply to a new lead | 5-6 | Your most-used draft | | A qualifying question | 4 | Probing before pitching | | Naming your price | 3 | Highest-stakes draft you have | | Handling an objection | 3 | "too expensive," "let me think" | | Follow-up to a quiet lead | 3 | A lighter register | | Booking or rescheduling | 2-3 | Short, logistical, easy to fumble | | Turning someone down | 2 | Almost nobody has these | | An ordinary busy-day reply | 3 | Your baseline, not your best | Take your own messages only, unedited, one thread per lead. Then take five things out, however often they show up in your real DMs: - **Any promise.** One sample saying "you'll be down two stone by June" teaches the model to say it to everyone. - **Invented urgency.** "Two spots left" when there weren't. - **Anything private.** Client names, health details, anything said to you in confidence. - **Prices you no longer charge.** A stale number surfaces at the worst moment. - **Real misspellings.** Loose grammar reads as human, but "recieve" just looks careless. ## Step 2: What goes into your voice spec? The spec is a short plain-text file the model reads before every draft. Ten things carry almost everything that makes a DM read as yours: message length and how many you send at once, greeting habit, capitalization, punctuation, emoji, whether you end on a question, how you say no, how you name a price, your stock phrases, and a never-say list. Answer one line for each, in words you'd actually use. From memory is fine, since a wrong answer you fix in week one costs you nothing. Keep it under 40 lines, because rules that argue with each other produce stiffer drafts. ```abap VOICE SPEC -- [your name] Updated: [date] I sell [one 12-week 1:1 program at GBP 900] to [lifters who've stalled]. LENGTH [9-14 words a message. Two sends back to back is normal.] OPENERS [No greeting. Start on the thing they just said.] [First name only when I'm being warm.] STYLE [Lowercase starts. Line breaks instead of commas.] [Emoji: 💪 or 🙂 only, one in ten, never first line.] [End on a question unless I'm confirming a time.] PRICE [Number bare, no cushion, then one question about fit.] NO [Say it plainly and give the reason.] NEVER SAY [reach out - circle back - investment - journey - at your convenience] Never promise a result, a timeline, or a number on the scale. Never invent a discount, a deadline, or a spot count. HAND TO ME Grief, illness, money trouble, or a call we already had. ``` Keep old copies, so a bad week is one file away from undone, and keep the offer facts in that top line so a price change is a one-line edit. ## Step 3: Why do worked examples beat more rules? Rules tell the model what to do. Examples show it. Paste in eight to ten whole exchanges (the lead's message and your real reply), labelled by type, pulled from the set you gathered in step one. That's enough for a DM setter, because you only really send a handful of kinds of message. More does keep helping: Google DeepMind's 2024 paper on many-shot in-context learning found the output kept improving as examples went from a handful into the hundreds, though every extra one costs something to run. For DMs the gains flatten long before the cost does. ```abap EXAMPLE 3 -- naming the price LEAD: ok so what's the damage 😅 ME: GBP 900 for the 12 weeks ME: does that land anywhere near what you had in mind? ``` Rotate the bank: when you approve a draft that beats its own example, swap it in. ## A worked example: Marcus, and eight before/after pairs Marcus is a composite, built from patterns coaches describe rather than one person, and his numbers are illustrative. He's in Manchester, sells a 12-week 1:1 program at GBP 900, and gets 35 to 45 DMs a week off Reels about training plateaus. His corpus: 27 replies across the eight types, a 22-line spec, nine examples. Week one he rewrote 22 of about 30 drafts, and three edits kept repeating: the model opened with the lead's first name, wrote "Great question!" before answering, and explained the program before the number. Those became three spec lines, and by week three he was editing six drafts in thirty. Same message twice below. **1. First reply off a Reel comment** > "Hi there! Thanks so much for your interest in my coaching program." > "right, the plateau thing -- how long have you been stuck at the same numbers?" **2. Qualifying before pitching** > "Could you share more about your routine and your goals?" > "how many days a week are you lifting right now?" **3. Naming the price** > "The investment for my 12-week transformation programme is GBP 900." > "GBP 900 for the 12 weeks" > >"does that land anywhere near what you had in mind?" **4. "It's too expensive"** > "I completely understand budget concerns! Many clients find the value outweighs the cost." > "fair. what were you expecting it to be?" **5. "Let me think about it"** > "Of course, take all the time you need! Let me know if you have questions." > "sure. what's the bit you're not sure about?" **6. Following up after three quiet days** > "Just following up to see if you're still interested!" > "hey -- you still thinking about it or has life taken over? either's fine" **7. Turning someone down** > "Thank you for your interest! Unfortunately this isn't the right fit right now." > "honestly not the right time for you -- six weeks of just showing up would do more" **8. A question you can't answer** > "That's a great question! While I'm not a medical professional, I'd suggest your doctor." > "that one's a doctor question, not mine. worth asking before we start anything" Pairs 4 and 5 map to [objection handling in DMs](/blog/objection-handling-in-dms) and [how to handle "let me think about it"](/blog/how-to-handle-let-me-think-about-it), pair 3 to [the price conversation in DMs](/blog/the-price-conversation-in-dms). ## Step 4: How do you correct drafts in the review queue? Turn the model loose on real DMs, but keep every draft behind the review queue. People want a hand on the wheel wherever AI is involved, and 61% of Americans said they want more control over how it's used in their lives, up six points in a year (Pew Research Center, 2025). The queue is where you keep that hand. In Luca, auto-send is off by default: each reply lands in a queue where you approve, tweak, or rewrite before anything sends. The corrections are the training data, so log them as you go. The third column is the point. | What you changed | Why | Where it belongs | | --- | --- | --- | | Cut "Great question!" | Not something you'd type | Never-say list | | Removed the name from the opener | You only use names warmly | Opener rule | | Moved the price above the explanation | You answer first | Price rule | | Split four lines into two sends | Your real rhythm | Length rule | Any edit you make twice in a week becomes a spec line. Any edit that survives the new rule becomes an example instead, because a rule the model can't apply needs showing, not stating. **Chart:** How sales professionals use AI: 43% use AI at work, 47% use it to write outreach, and 22% use it to qualify leads, per HubSpot's 2024 State of AI in Sales report. (Source: HubSpot State of AI in Sales, 2024.) ## Step 5: How do you run the iteration loop and catch drift? The loop is draft, compare, correct, retrain, on a schedule. Without the schedule, training quietly turns into editing every message by hand forever, which is the thing you bought the tool to stop doing. | When | What you do | Roughly | | --- | --- | --- | | Every draft, weeks 1-2 | Edit in the queue, log what repeats | As you go | | End of week 1 | Promote repeated edits into spec lines | 15 min | | End of week 2 | Swap better approved drafts into the bank | 15 min | | End of week 4 | Run the blind read | 20 min | | Monthly after | Cold-read 20 approved drafts, mark what's off | 10 min | Drift then arrives on a trigger rather than a timer. Four of them force a retrain: a price change (fix the spec line and every price example that day), a second channel (run a separate spec; the [multi-channel DM playbook](/blog/multi-channel-dm-playbook) covers why), a one-season corpus, and any tool upgrade. ## How do you know the training worked? 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 out of twenty is chance; fifteen or more means the spec is missing something their reasons will name. Then branch on the result: - **Wrong in one message type only.** Add four or five samples of that type. - **The same wrong move everywhere.** A missing spec rule, not a missing sample. - **The rule exists and it keeps breaking it.** Replace the rule with an example. - **Still rewriting most drafts, or one-word edits on half.** Review stays on for everything. - **Two clean weeks plus a blind read at 10-12.** Auto-send first replies and scheduling nudges only; price and objections stay in the queue. - **A draft made a claim you wouldn't make.** Stop until the never-say list is fixed. ## Which edge cases should you decide now? - **You don't have 25 sendable DMs yet.** Write the spec from memory, start with four or five examples, and keep review on until the corpus fills. Don't pad with invented messages. A fabricated sample trains a fabricated voice. - **Half your samples were written by a setter.** Build from one person only. A mixed set produces a third voice neither of you recognises. - **Your offer changed halfway through.** Keep the older samples for rhythm, delete every one that names the old price or length. Voice survives a change of offer. Facts don't. - **You genuinely write long.** Set the length rule to your real median, not a "short DMs" rule you read somewhere. ## Why isn't the training loop improving your drafts? | Symptom | Likely cause | Fix | | --- | --- | --- | | Drafts got worse after new samples | Last batch was off-voice, or someone else's | Roll back a spec version, add samples one type at a time | | First replies land, objections don't | Coverage gap in the corpus | Add three or four objection samples, including one you lost | | The spec grows, drafts stiffen | Too many rules, and some of them clash | Cap the spec at 40 lines, move the fine detail into examples | | Voice is right, facts are wrong | Offer details scattered through the rules | Keep facts in the header line, update the day they change | | Drafts read like your best writing | You only fed polished samples | Add three ordinary busy-day replies | ## Which mistakes stall AI voice training? 1. **Writing the spec before collecting the samples.** Adjectives first gets you "friendly and professional," the prompt every coach writes and the reason every draft reads the same. 2. **Rewriting drafts outside the tool.** Typing your reply straight into Instagram feels faster and teaches the model nothing. The edit has to happen where it gets recorded. 3. **Feeding only the threads that closed.** The follow-up that got ignored and the lead you turned down teach range a win-only corpus can't. 4. **Treating training as setup.** A spec written once in September describes a coach who no longer exists by March. ## One honest limit Training gets the voice right, but it can't invent what the model was never told. If a lead brings up a joke from your last call, or something they said on their intake form, the AI won't know unless it's there in the thread. It won't make up a memory, and you wouldn't want it to. The failure has a shape worth naming. A lead writes "I've barely trained since my mum got ill," and a perfectly-voiced draft asks how many days a week they're lifting. The voice was right, the judgment was missing, and that thread is over. Fixing it takes three steps. Send grief, illness, and money trouble to yourself before the model drafts anything; that's the last block in the spec above. Keep auto-send off on any thread where a lead has told you something personal. And if a tin-eared message did go out, reply yourself within the hour and own it plainly, without explaining the software. People forgive a person correcting themselves. ## Should you tell leads an AI helped write your replies? Yes. If a lead asks, you say so, and you never deny it. Disclosing an automated experience is the honest default, and in places like California it's the law. Gartner's 2026 survey of 3,566 customers found 87% say it's essential that companies using generative AI offer a route to a human. The review queue means that route was never removed. Saying it costs you nothing, and hiding it costs you the thread the moment someone works it out. Reply only to people who messaged you first and you stay on the right side of the platforms too, which [is Instagram DM automation safe](/blog/is-instagram-dm-automation-safe) covers in full. For the wider setup: the complete guide to AI setters for coaches, [what an AI DM setter is](/blog/what-is-an-ai-dm-setter), and [how to automate Instagram DMs with AI](/blog/how-to-automate-instagram-dms-with-ai). Every [Luca plan](/pricing) includes the review queue. ## FAQ ### How many DM samples do I need to train the AI on my voice? Around 25-30 real replies, chosen for coverage rather than quality. Take five or six first replies, four qualifying questions, three price conversations, three objections, three follow-ups to quiet leads, and two where you turned someone down. Push to 40 or 50 if you also want to count patterns into a fingerprint sheet. ### How long does it take before the AI sounds like me? Usually a week or two of real DMs. Early drafts need heavy editing while the model learns from your corrections. As you approve and tweak, the edits shrink. There's no fixed number. It depends on how much you correct and how consistent your own voice is across samples. ### What should a voice spec include? Ten things: message length, opener habit, capitalization, punctuation, emoji use, whether you end on questions, how you say no, how you name a price, your signature phrases, and a never-say list. Keep it under 40 lines. Google DeepMind's 2024 many-shot research found examples keep improving output, so add eight to ten real exchanges alongside the rules. ### Do I need to keep correcting drafts forever? No. Correcting is heaviest at the start, when every edit teaches the model. Once drafts land consistently, you approve with small tweaks or clean. You can then auto-send low-stakes replies and keep the review queue on for price and objections, so you only weigh in where it matters. ### Is training AI on my voice different from making DMs sound less robotic? Yes. Making DMs sound less robotic is fixing tells after the fact: stiff openers, over-explaining, perfect grammar. Training is how you build the voice in the first place, with samples, a spec, examples, and corrections. Do the training well and there are fewer tells to fix. ### Will leads know an AI is drafting my replies? Tell them if they ask, and never deny AI involvement. Gartner's 2026 survey of 3,566 customers found 87% say it's essential to have an option to reach a human when a company uses generative AI. A review queue keeps that option open, and a draft you approved is a message you sent. ### What kind of DMs make the best training samples? Real ones you actually sent, across every situation the AI will meet. Include the ones you'd rather skip: a price you held, a lead you turned down, an ordinary busy-day reply. Leave the typos and lowercase in. The polished version of you is not the version your leads recognise. ### Can I train the AI without turning on auto-send? Yes, and you should at first. Keep auto-send off, let every draft land in the review queue, and correct each one. That edit-and-approve loop is where the model learns. Plenty of coaches leave price and objection threads in review for good, and that's a fine place to stop. ## Sources 1. [HubSpot -- State of AI in Sales 2024](https://blog.hubspot.com/sales/state-of-ai-sales) 2. [Twilio -- 2025 State of Customer Engagement](https://www.twilio.com/en-us/press/releases/socer-2025) 3. [Agarwal et al. (Google DeepMind) -- "Many-Shot In-Context Learning" (2024)](https://arxiv.org/abs/2404.11018) 4. [Gartner -- "87% of Customers Say Companies Using GenAI for Customer Service Must Provide Access to a Human Agent" (2026)](https://www.gartner.com/en/newsroom/press-releases/2026-08-04-gartner-survey-finds-87-percent-of-customers-say-companies-using-genai-for-customer-service-must-provide-access-to-a-human-agent0) 5. [Pew Research Center -- "AI in Americans' Lives: Awareness, Experiences and Attitudes" (September 2025)](https://www.pewresearch.org/science/2025/09/17/ai-in-americans-lives-awareness-experiences-and-attitudes/) --- Published by SetLuca, the company behind Luca.