Most AI email tools have the same failure mode. They can produce a grammatically correct, professionally worded reply to almost any message. The problem is that the reply sounds like it was written by someone who has never met you.
Voice is the thing they miss. And it turns out voice is harder to fake than most people assume, because it is made up of dozens of small, consistent signals that recipients pick up on without consciously noticing them.
What voice actually is
When we talk about someone's email voice, we are not talking about whether they write in complete sentences. We are talking about a cluster of very specific behaviors that show up consistently across thousands of messages.
Greeting choices. "Hi" versus "Hey" versus first name only versus nothing at all. People make this same choice over and over, often without realizing it, and recipients notice when it changes.
Sign-off patterns. "Thanks" versus "Best" versus just the person's name at the bottom. Some people never use a sign-off at all for short replies. Some use different closings depending on the relationship. This is highly personal and surprisingly consistent.
Reply length and structure. Some people write in short, punchy bursts. Others explain everything in full. Some use numbered lists instinctively. Others never do. This is not about formality. It is about how a person's thinking shows up on the page.
Tone calibration by context. How much warmth comes through in a routine acknowledgment? How direct is this person when they need something? Most people shift register depending on who they are writing to, more casual with long-term clients, more formal with new contacts. An AI that applies one mode to everything gets this wrong in ways people notice.
Punctuation habits. The Oxford comma or not. Ellipses used casually or almost never. An exclamation point in every third message or used sparingly. These are tiny individually, but they stack up into a recognizable pattern.
Why generic AI training does not capture this
Large language models are trained on enormous amounts of text. They produce fluent, professional email without any trouble. But "fluent and professional" is a style in itself, and it is nobody's actual style.
The model does not know that you never use passive voice in client emails. It does not know that you always reference the person's name in the second paragraph. It does not know that your sign-off changes depending on whether the conversation has gone well. It just writes something that looks right, because it has been trained on what professional email tends to look like.
The result is plausible but generic. Recipients who know you well read it and something feels slightly off. They cannot always articulate what. They just know it does not quite sound like you. That small friction, repeated across hundreds of emails, starts to change how people feel about a relationship.
What prompt-tuning gets wrong
The most common workaround is asking users to describe their own voice. "Write in a casual but professional tone. Keep replies under 100 words. Always thank the person for reaching out."
This helps a little. But there are two problems with it.
First, most people are not good at describing how they write. They know their email when they read it back, but they cannot articulate the specific choices that make it theirs. Voice is mostly below the level of conscious awareness.
Second, rules described in plain language are interpreted generically. "Casual but professional" means something different to the model than it does to you, because the model has no reference point for your specific version of casual.
What works instead
The only approach that actually captures real voice is learning from real examples. Not rules written down by the user. Actual sent emails from the actual person.
Reading a sample of someone's sent mail reveals patterns they could not have articulated themselves. Most people do not consciously know how they start a message to a vendor versus a long-term client. They just do it a certain way, every time.
When you build a profile from those examples, you get something qualitatively different from a prompt-tuned assistant. You get replies that recipients read and think: yes, that sounds like them.
That is the bar. Not "sounds like a professional email." Not "sounds like a good AI assistant." Sounds like that specific person.
It turns out that bar is achievable. It just requires building from real data rather than generic instruction. The hard part is doing it in a way that does not require storing your emails permanently or exposing them to people who do not need to see them. But that is a different article.