Email still takes up to 28% of the knowledge worker’s workweek, according to the McKinsey Global Institute, and most tools that promise to fix it are autoresponders with better marketing. What we build is different: a system that learns from your past conversations, your documentation, and your reply habits, then drafts replies your team only has to review and send.
The language model is the least interesting part of that system. What decides whether a draft gets sent or rewritten is everything around it. Here is how the pieces fit together in production.
It Starts With Your Past Conversations
Before writing a word, a good system reads. During setup it goes through a few hundred of your recent threads and extracts two things.
The first is a style guide: how you open, how you sign off, how formal you are, how long your sentences run, how you deliver bad news. It lives in a plain document your team can read and correct. If your support team writes “Hi Tom” and never “Dear Mr. Jansen”, the drafts will too.
The second is a Q&A library. Recurring questions get paired with the answers your team actually gave, which usually comes out to 50 to 100 entries covering pricing, shipping, compatibility, returns, whatever your inbox is made of. You go through the list once, drop the outdated entries, and approve the rest.
A model without context writes like a press release. Fed your phrasing and your real past answers, it writes drafts that read like your team on a good day.
A Knowledge Base It Can Search, Tools It Can Call
That Q&A library becomes the core of the knowledge base, indexed with semantic search rather than keywords. “How much is shipping to Belgium?” and “wat kost verzending naar België?” land on the same entry. Product documentation, price lists, and policy pages get indexed alongside it.
Static knowledge has limits, though. No FAQ can answer “where is my order?” For that the system calls tools: an order lookup, a stock check, a calendar, a CRM record. On a real email, the two together look like this:
Inbound, 9:14 am. “Hi, I ordered an amplifier last Tuesday, order 84312. It still hasn’t arrived. Where is it?”
Draft, 9:14 am. “Hi Tom, sorry for the wait. Your order left our warehouse yesterday, and the carrier expects delivery tomorrow before 6 pm. Here is the tracking link. Let me know if nothing shows up.”
Behind that draft: one FAQ entry about delivery delays, one order lookup, phrasing from the style guide. In front of it: a human who clicks send.
The harder question is what happens when nothing matches. If neither the knowledge base nor a tool produces the answer, the email is flagged for a human instead of answered with a guess. The same honesty applies to attachments: it will never pretend to have read a document it could not open.
Audiovolt, a Dutch car audio retailer, runs on this setup for hundreds of product questions a day. Reviewing a grounded draft takes about 30 seconds; writing the same answer from scratch took 5 minutes. Mailflow, our AI auto-responder, is a simplified self-serve version of the same approach.
Templates Still Have a Job
Not every reply should be generated. Order confirmations, quotes, onboarding sequences, anything a lawyer or an accountant signed off on: that wording should be fixed, reviewed once, and sent identically every time.
The interesting cases are hybrids. A quote request comes in, the template provides the structure and the legal footer, and the model fills in the rest: the right products, the right discount tier, and one paragraph that actually responds to what the customer wrote.
Our rule of thumb: if someone approved the exact wording, make it a template. If a human would rephrase it for every customer, let the model draft it.
WingBuddy, a sales email assistant we built for a client, runs on that hybrid. Their reps write a rough draft, and the AI reshapes it with the company’s templates and the style of its top performers.
Want this on your own inbox? Our Email Automation package covers the whole chain: style and FAQ extraction, knowledge base, templates, and deployment on your mailboxes, with human validation built in. See what’s included
The Plumbing: Threads, Signatures, and Your Actual Mailbox
This is the part demos skip and production depends on.
Drafts have to appear inside the existing thread, with the history quoted, from your address. Your team reviews them in the inbox it already works in, not in a separate dashboard. Labels handle triage along the way: drafted and ready for review, needs a human, filtered out.
Signatures are managed and validated centrally, so every draft leaves formatted and on-brand.
None of this should depend on your provider either. The automation talks to a common mail interface, so the same workflows run on Gmail, on Outlook, or on any mailbox that speaks SMTP. Changing providers does not mean rebuilding the automation, and nobody migrates to a new platform.
Attachments go both ways: inbound invoices and forms are read and their data extracted, outgoing drafts can carry brochures, spec sheets, or the quote itself.
Humans Approve, and Everything Is Logged
Default mode is draft, not send. Every reply waits for an approve, an edit, or a reject, a principle we apply to every AI project we deliver.
Each run is also logged step by step: which filter matched, what the search retrieved, which tools were called, what the model wrote. When a draft is wrong, you can see exactly why, fix the FAQ entry or the template behind it, and the correction carries over to every future draft.
Auto-send has to be earned. Once a category has shown weeks of accurate drafts, you can let it go out on its own. Some clients do. Many keep a human on every send and still recover most of the time, because reviewing takes seconds where writing took minutes.
Built on Our Own Stack
The engine underneath is Vectoria, our in-house retrieval and workflow engine, self-hosted on EU infrastructure. Your past emails, product docs, and pricing stay in a knowledge base you control rather than in a third-party vector database. For clients in finance, healthcare, and legal, that matters as much as the hours saved.
Setup takes 2 to 4 weeks, extraction and review included, and our clients typically save 15 to 20 hours a week across a team.
Explore the Email Automation package: transparent pricing, one setup fee, three monthly tiers.
Want to see it answer a real email first? Book a free 30-minute discovery call, bring a handful of typical messages from your inbox, and we’ll walk through what the system would do with each one.