How AI Learns to Write Emails in Your Voice (2026)
· The Agentys Team
How AI voice modeling learns your writing style from 90 days of sent mail — and why generic AI drafts sound wrong. Explainer on per-contact tone modeling, with limitations.
Generic AI drafts sound like a polished stranger wrote them. The technique that fixes this — building a per-contact voice model from 90 days of your sent mail — is what separates useful AI email tools from the ones that make you rewrite everything anyway.
Why Writing Email Is a Bigger Problem Than You Think
Before getting into how AI voice modeling works, it is worth being clear about the size of the problem it is solving. Knowledge workers spend roughly 28% of their workweek reading and responding to email (McKinsey Global Institute) — about 11 hours out of a 40-hour week, every week, without pause. Email is the dominant channel for professional communication, and composing replies is the single activity inside that channel that consumes the most time and cognitive energy. Writing is slow, especially when you care about tone.
The interruption cost compounds the writing cost. Research on attention has found that it takes about 20 minutes to return to full focus after an interruption. Email is an interruption machine: the average professional checks their inbox dozens of times per day, each check carrying that refocus penalty even when the email itself takes 30 seconds to read. The cumulative effect is not just lost time — it is the degradation of the kind of sustained attention that complex work requires. When you add the cost of writing replies to the cost of the interruptions that triggered those replies, email consumes a disproportionate share of the focus professionals have available each day.
This is the context in which AI email drafting has to be evaluated. The promise is real: if an AI can write your replies for you, those 11 weekly hours are recoverable. The problem is that the first generation of AI tools wrote replies that did not sound like you — and a draft you have to rewrite from scratch saves very little time.
Why Generic AI Drafts Sound Like a Polished Stranger
Open ChatGPT, paste in an email you received, and ask it to draft a reply. The result will be grammatically correct, logically organized, and completely anonymous. It will greet the sender with a phrase you would never use. It will close with a sign-off you abandoned years ago. It will match the formality of a job application letter regardless of whether the original email was a quick Slack-style note from a colleague you have worked with for six years. The problem is not that the AI is unintelligent. The problem is that it has no data about you. ChatGPT, Gmail's Gemini compose feature, and most first-generation AI writing tools operate from a blank slate on every prompt. They know the conventions of professional email in general, extracted from the billions of documents they were trained on. They know nothing about the specific human being who is about to click send.
Professional communication is not uniform. Most people write differently to their manager than to their direct reports, differently to a long-term client than to a cold prospect, differently on a Friday afternoon than on a Monday morning. These variations are not random — they reflect relationships, context, and history. When you write to a colleague you have collaborated with closely for two years, you skip the pleasantries, use shorthand you have developed together, and land the point in two sentences. When you write to a new enterprise buyer, you slow down, layer in context, and match the formal register of their own previous correspondence. A generic AI cannot reproduce either of those versions because it does not know either relationship exists. The output is always a compromise average — something that would not embarrass you if sent to a stranger, but that would feel oddly stiff or oddly breezy depending on the actual recipient.
This is why so many professionals who try AI email tools keep rewriting the drafts. The AI did not fail because the grammar was wrong. It failed because the voice was wrong, and voice is the hardest thing to fix after the fact. Rewriting tone takes almost as long as writing from scratch.
What 'Learning Your Voice from 90 Days of Sent Mail' Actually Models
When Agentys says it learns your voice from 90 days of sent mail, that phrase covers a specific set of measurements. The system does not simply scan your emails and store a vague impression of your "style." It extracts discrete, measurable signals from each message and maps them to the sender-recipient pair. Understanding what those signals are makes the approach more concrete — and makes it easier to evaluate whether any AI email tool is genuinely doing this work or just claiming to.
The first layer is greeting and sign-off conventions. These are the most consistent stylistic markers in professional email. Most people have patterns they apply reliably — first name only for close contacts, title plus last name for formal relationships, no greeting at all for rapid back-and-forth chains. Sign-offs follow the same logic: "Thanks" for internal messages, "Best regards" for external clients, nothing for quick internal replies. A voice model that has seen your last 90 days of correspondence can extract these conventions per contact with high confidence, because they are consistent enough to surface as clear patterns with enough data.
The second layer is formality and sentence structure. This is where most generic AI tools fail. Formality is not binary — it sits on a spectrum, and it shifts depending on the relationship, the topic, and the communication channel's established norms. Short declarative sentences signal speed and familiarity. Longer, subordinate-clause-heavy constructions signal care, distance, or seniority. An AI that has read 50 of your emails to a specific vendor can determine that you write to them at a formality level of roughly 7 out of 10: professional but not stiff, with an occasional first name and no formal title. It can then calibrate drafts to that level automatically, without you selecting anything from a dropdown.
The third layer is topic-specific vocabulary. Every professional has a lexicon that is specific to their industry, their company, and their relationships — the shorthand and technical terms they use without defining, because the people they email already share that context. A SaaS consultant and a corporate lawyer each have their own set of these. The AI should extract them as markers of register: if you use a term without definition five or more times across your history, it belongs to your standard professional vocabulary and should appear in drafts naturally rather than being avoided or explained.
The fourth layer — and the one that makes the per-contact model genuinely powerful — is relationship history. An email to someone you have corresponded with 200 times carries an entirely different weight than an email to someone you are contacting for the third time. The AI can read that relationship density directly from the sent folder. Dense correspondence history licenses a shorter, more direct tone. Sparse or new correspondence calls for more context, more courtesy signals, more explicit framing. This is something humans do automatically; a model that can do it programmatically is doing something meaningfully different from a generic drafting tool.
How Agentys Applies Per-Contact Voice Modeling at $16.99/mo
Agentys is an AI email assistant priced from $16.99/month that connects to your inbox via secure OAuth and processes your mail automatically. The voice modeling pipeline runs during onboarding and then continuously in the background. Here is how it works in practice.
When you first connect your account, Agentys reads your last 90 days of sent mail. For most professionals that is somewhere between 500 and 2,000 emails — enough to establish stable patterns for frequent contacts and reasonable baselines for infrequent ones. The analysis extracts the four signal layers described above: greeting and sign-off conventions per contact, formality and sentence structure per relationship, domain vocabulary from your overall writing, and relationship density from correspondence frequency. This initial analysis runs within a few hours. You do not configure anything. There is no survey asking about your communication style. The model reads your actual behavior, not your self-description of it.
After the first automatic processing cycle, your inbox has been triaged and draft replies are waiting. The drafts reflect the per-contact voice model: the one for your CEO uses your formal register, the one for your closest colleague uses your shorthand and informal sign-off, the one for a new external contact uses the middle register you typically apply to first or second interactions. Your job is to review and approve — most drafts can be sent after a read-through of 15 to 20 seconds. When you edit a draft before sending, the system logs the correction and refines the relevant parameters in the model. Over the first three to four weeks, accuracy climbs as the model accumulates more correction data. By week four, most users find that the large majority of drafts can be sent with little or no editing.
Disclosure: Agentys is the publisher of this article. The $16.99/month pricing above reflects the Starter plan as of May 2026; check [agentys.io](https://agentys.io) for current rates.
Where the Technology Still Has Limits
Per-contact voice modeling is a genuine advance over generic AI drafting, but it is not a replacement for your judgment. There are categories of email where even a well-trained model should not be trusted to produce a send-ready draft — and where Agentys, by design, flags the message for manual attention rather than generating a draft at all.
The most important category is emotionally sensitive communication. A client expressing frustration about a missed deadline. A team member sharing difficult personal news. A partner raising concerns about the future of a relationship. These messages require genuine empathy, careful word choice, and a reading of emotional subtext that no current AI reliably provides. A model can produce grammatically appropriate language, but appropriate is not the same as right. In sensitive situations, the cost of getting the tone wrong is high enough that the draft should be a starting point only — if it is generated at all. Many practitioners find it faster and less risky to write these from scratch.
The second category is novel or high-stakes positions. Contract negotiations, legal matters, significant pricing discussions, and strategic decisions carry implications that extend beyond the immediate email thread. A voice model trained on your routine correspondence has no way to understand the stakes of a particular message or to reason about its downstream consequences. These require your full attention and should not be approximated from historical patterns.
The third, more mundane limit is new relationships. The model needs history to be accurate. For a contact you are emailing for the first or second time, there is not enough data to establish a reliable per-contact profile. Agentys defaults to a conservative register for these cases — which is usually adequate but may not capture nuance you want to project in a first impression.
These limitations are worth naming clearly because they define the correct mental model for using any AI email tool. The right frame is not "the AI writes my emails." It is "the AI handles the 80% of emails that are routine so I can give full attention to the 20% that are not." That division is where the genuine productivity gain lives.
The reason generic AI email drafts fail is not technical — it is that they have no data about the specific person writing them. Per-contact voice modeling changes that by extracting measurable signals from your sent history: greeting conventions, formality calibration, domain vocabulary, and relationship density. The output is drafts that your recipients read as you, not as a language model. Agentys applies this at the inbox level automatically, so your review takes minutes rather than an hour. One caveat worth keeping: for sensitive, novel, or emotionally complex messages, the draft is a starting point, not a finished product. Treat it that way, and the tool earns its place in your workflow.