If you sell through conversations, you've probably had this fear. If a machine writes for you, your messages will end up sounding like everyone else's, and whatever made you different will dissolve into generic politeness.
The fear is real, but the cause is usually misread. What erases your voice is not the technology. It's the way you ask the machine to write. A message that sounds off almost always comes from a vague instruction, not from a limit of the model.
This article covers what to give an AI so its drafts sound like you, what to keep out of the system, and how to check the result. It doesn't cover whether you should delegate in the first place. We looked at that when comparing what a human setter and an AI setter each do best.
What an AI actually copies when it imitates a voice
It helps to take the mystery out of this, because once you understand it, you use it differently.
What a model picks up from a set of messages is surface patterns, and there are more of them than you'd expect:
- Sentence length, and how much it varies from one message to the next.
- Punctuation: how often you use ellipses, exclamation points, capital letters, or no punctuation at all.
- How you open. This is often the single most recognizable thing about how a person writes.
- Vocabulary: the words you use for your work, your offer, your clients.
- What you never say. The phrases you avoid define a voice at least as much as the ones you use.
None of this is personality in the usual sense. It's statistics on form. That explains two things at once: why the result can be eerily close to how you write, and why it will never know what you actually think about a specific client.
What to give it: your messages, not a description of yourself
The most common mistake is describing your style instead of showing it. "Warm but direct, professional without being cold" describes almost every salesperson alive, so it produces a message that sounds like almost every salesperson alive.
The useful material is your real messages. Around thirty real exchanges is usually enough for the patterns above to show up. Thirty raw messages beat three that you polished for the occasion. The conversations where you were rushed count as much as the ones where you took your time, because your real voice includes your ordinary days.
Lean toward exchanges that went somewhere. Not because they're better written, but because they contain the phrasing that works on your audience. A sales voice isn't only an aesthetic. It's also whatever gets people to reply.
Two extra things are worth spelling out, because no set of messages will reveal them on its own:
- A list of things you never say. Maybe you never write "hey there," never use "circle back," never put an emoji in a first message.
- Your fallback lines for when you don't know. This prevents the most visible flaw of generated messages: answering with full confidence a question that has no answer yet.
Here is the kind of fallback line worth writing down once, in your own words:
Good question, and I don't want to guess. Let me check and I'll get back to you tomorrow.
What not to give it
There's a clear line here, and holding it protects you as much as your leads.
What you know about someone outside the conversation stays out. Something a client told you on a call, something you heard from a mutual friend, something they asked you to keep quiet: none of it belongs in the written history, and none of it should feed automated drafting. The risk isn't theoretical. An AI that has a piece of information will use it to personalize, and a lead who sees a detail they shared somewhere else resurface in a DM knows exactly what happened.
Commitments with numbers deserve the same caution. A timeline, a price, a guarantee: these are things you either keep or don't, and they should never appear in a message you haven't read. The same goes for anything touching health or sensitive personal situations, where loose wording can do real harm.
Your clients' conversations with each other stay out too, if you run a group or community. Those messages aren't yours.
The test that tells you if it works
Fancy evaluations don't help much here. A simple test is enough.
Take five generated messages, mix them with five you actually wrote, and give the ten to someone who knows you well. Don't tell them what you're looking for. Just ask them to point out the ones that don't sound like you.
- They flag one or two: you're in an acceptable range, and their explanation tells you what to fix.
- They flag all five: the problem is almost always the material you provided. Too few messages, or messages that were edited too much.
- They flag none: double-check that they understood the task, because that result is rare.
This beats any automated score for one reason. It measures the only thing that matters, how a human reader perceives the message, and it often reveals habits you didn't know were part of your writing.
What drifts over time, and why
A setup that works in January can sound off by September, and the machine isn't what changed.
Your offer evolves, your vocabulary follows, and phrasing learned from old exchanges ends up describing a business you don't quite run anymore. Your audience changes too. What worked on your first clients doesn't necessarily work on today's.
There's a second mechanism, and it's sneakier. If you approve generated messages without ever editing them, your own writing habits slowly align with them. The material you'd provide today already contains part of what the machine wrote. The voice starts feeding on itself and gets thinner.
The fix is modest. Actually edit a share of the drafts, not just approve them. And refresh the material two or three times a year from recent conversations. If you're following up old conversations, those threads are a good place to check whether your voice from last year still matches the one you use now.
What your leads should know
The transparency question comes up as soon as an AI drafts anything, and there are now rules around it. If you sell to people in Europe, the EU AI Act sets disclosure duties when a person interacts with an AI system, spelled out in its article on transparency obligations. Other jurisdictions are moving in the same direction, and platforms have their own policies on automated messaging.
In practice it comes down to one distinction. A system that talks to a person on its own is squarely what these disclosure duties target. A tool that prepares a draft, which a human reads, edits and sends under their own name, is in a different position: the message that goes out is the message of a person who approved it.
That's one reason human approval before sending isn't a cosmetic safeguard for us. It changes the nature of what gets sent. We wrote a full piece on why a human in the loop matters if you want the longer version.
The part you can't hand off
Some things no setup will produce, and it's better to know that from day one.
Judging whether to push is yours. An AI sees that a conversation has been open for six days. It doesn't know this person told you they're going through a hard time, or that their silence probably has nothing to do with your offer. That's also why a lead left on seen deserves a human read before anyone writes back.
Deciding to turn a client down is yours too, and over time it's often the most profitable decision you make. No system trained to move conversations forward will suggest, on its own, that you end one.
What you promise carries your name. An AI can write the sentence. It can't keep it, and the relationship you're building rests on that difference. Delegating the writing is reasonable. Delegating the commitment is much less so, and knowing where the line sits between the two is the real work.
FAQ
What is Ficiel?
Ficiel is an AI inbox for coaches and setters: it surfaces old Instagram, WhatsApp and LinkedIn conversations worth following up, drafts the message, and nothing is sent without your approval. Follow-up drafts are built from your own messages, in your tone, and you always approve before anything goes out.
How many messages do I need to give an AI so it writes like me?
Around thirty real exchanges is usually enough for the patterns that make up a voice to show. Authenticity matters more than volume: thirty raw messages beat ten rewritten ones, because rewritten messages describe how you'd like to write, not how you actually do.
Will my leads notice it's AI?
What people pick up on is mostly two things: a reply that ignores what was just said, and a sameness of tone no real person has. A message drafted from your own exchanges and read by you before it's sent has neither problem.
Do I have to disclose that I use AI?
It depends on what the system does. An AI that converses with someone on its own falls under disclosure rules like the ones in the EU AI Act. A tool that drafts a message you read and send yourself puts that message under your responsibility, which is a different situation. Rules vary by country, so check the ones that apply to where your leads are.
Can I use my clients' conversations to train the tool?
Your own messages are yours, and they're the normal material for this. What other people wrote to you belongs to a private exchange. And anything shared outside the written conversation, on a call or through someone else, should never go into the system.
What if the drafts still don't sound like me?
The problem is almost always the material, not the settings: too few messages, edited messages, or a set of conversations from a period when you were selling something else. Pull thirty recent conversations that went somewhere, don't clean them up, and run the friend test again.
