An AI email assistant writes in your tone by combining two separate systems: tone matching, which learns your vocabulary, sentence length, and sign-off style from emails you've already sent, and RAG (retrieval-augmented generation), which pulls the specific facts a reply needs from a knowledge base you control. Tone matching decides how a draft sounds; RAG decides what it says. A draft that gets both right reads like you wrote it yourself, in one sitting, with the right answer already in hand.
Tone matching: learning your voice from sent mail
Tone matching works from a sample of emails you've actually sent — not a style you describe in a settings form, but the writing you've already produced. The model looks at patterns: how long your sentences tend to run, whether you open with "Hi" or "Hey" or a name, how formal your sign-off is, whether you use exclamation points or keep things flat, how much you hedge versus state things directly.
None of this is copied verbatim from any single email — it's a profile built across many messages, so the assistant isn't quoting your past self, it's writing new sentences that follow the same patterns. The more sent mail it has learned from, the more consistently a draft will sound like you rather than like a generic "helpful assistant" tone that shows up in most AI writing tools.
RAG: retrieving facts instead of guessing them
Tone alone doesn't solve the harder problem: getting the content of the reply right. If a customer asks about your refund window or a specific product spec, a model with no grounding will produce a fluent, confident-sounding answer that may simply be wrong — this is the well-known failure mode often called hallucination. RAG fixes this by giving the model somewhere real to look before it writes anything.
In practice, RAG means your knowledge base — product docs, a pricing sheet, policies, an FAQ you've uploaded — gets split into small chunks and stored as embeddings, a numerical representation of meaning rather than exact wording. When a new email arrives, the assistant searches those embeddings for the chunks most relevant to the question being asked, and hands only those chunks to the drafting model as context. The draft is then written using that retrieved material, not from the model's general training data.
Why retrieval beats a model's memory
A general-purpose language model has no idea what your refund policy actually says today — it wasn't trained on your business. RAG closes that gap without retraining anything: you update your knowledge base whenever a policy or price changes, and the very next reply reflects it, because retrieval happens fresh at draft time rather than being baked into the model months earlier.
How the two systems work together on one draft
When a message needing a reply comes in, the assistant runs both processes before writing a single word: it retrieves the relevant knowledge-base chunks for what the email is asking, and it loads the tone profile built from your sent mail. The drafting step then combines them — the retrieved facts constrain what the reply can correctly say, and the tone profile shapes how it's phrased. You end up with an answer that's both accurate and recognizably yours.
- Tone matching: learned from your own sent emails, not a style description
- RAG: retrieves facts from your knowledge base instead of guessing
- Retrieval happens fresh at draft time, so updates to your knowledge base apply immediately
- Neither system sends anything — every draft still waits for your approval
What a knowledge base actually adds
If you skip the knowledge base entirely, the assistant can still match your tone — sent-mail learning doesn't depend on it — but it will draft replies from general knowledge alone whenever a question needs a specific fact about your business. Adding even a short set of documents (a pricing page, a policy doc, a product FAQ) is what lets the assistant answer correctly instead of just plausibly. The more current and specific the knowledge base, the fewer edits a draft needs before you send it.
Nothing here changes who's in control. Retrieval and tone matching only shape the draft that lands in your Gmail or Outlook drafts folder — a person still reads it, edits anything that's off, and decides whether to send.
DraftKite runs this exact combination — tone learned from your sent mail, facts retrieved from your own knowledge base — on every reply it drafts. Start the 7-day free trial, no card required, and see how close the first drafts already sound to something you'd write yourself.