AI Office Email Operations: From 200 Unread to Zero

AI Office Email Operations: From 200 Unread to Zero

The first time a 22-person logistics company in Hessen showed us their operations director’s inbox, the unread count was 847. By Friday it was 1,200. Nothing was getting lost exactly — but nothing was getting handled either. Important supplier threads sat two clicks below a newsletter from an industry association. A customs broker’s deadline warning was filed in the same mental bucket as a Slack notification mirror.

This is the failure mode Email Operations exists to fix. Not “AI for email” in the marketing sense — the actual deployed module that runs in production inside the AI Office bundle we ship to DACH SMEs. The module is €299 per month, deploys in under a week, and replaces a meaningful slice of what a human executive assistant does during a 4-hour morning window.

Here is how it works, what it does well, what it doesn’t, and what it costs compared to the alternatives.

The Architecture: Three Loops, Not One

Most AI email tools on the market — Superhuman Mail, Fyxer, SaneBox Pro — run a single loop. They read your inbox, classify messages, and either draft a reply or file the message. One pass. One model. One decision boundary.

Email Operations runs three loops, each with a different latency budget and a different failure tolerance:

1. Triage loop (every 60 seconds). A lightweight classifier reads inbound mail, scores it on five dimensions (sender importance, topic, deadline signal, request type, sentiment), and assigns one of four buckets: respond, review, archive, flag. This loop runs on a smaller, faster model — the cost of a wrong classification is low because the human reviews the digest.
2. Draft loop (every 15 minutes). For everything in the respond bucket, a larger model drafts a reply in the configured voice. The draft sits in a queue, never auto-sent. A human reviews once or twice a day, edits what needs editing, hits send on the rest.
3. Follow-up loop (daily at 07:00 local). A separate agent scans sent mail, outgoing threads, and calendar for promises made or commitments due. If a thread has gone silent for 5+ days with an open question, the loop drafts a nudge. If a deadline is in the next 48 hours, it surfaces in the morning digest.

The three-loop structure matters because the failure modes are different. A bad triage decision is recoverable in 60 seconds. A bad auto-sent reply is not recoverable at all. A missed follow-up is recoverable for a few days, then it isn’t. The architecture matches the cost of being wrong to the latency of the loop.

What The Triage Classifier Actually Does

The classifier is the part most teams underestimate. It is not “sort personal from work.” It is not “move newsletters to a folder.” The five-dimension score produces a single decision per message: what bucket does it go in, and how urgently.

In practice, the configuration looks like this for a 20-person Mittelstand company:

• Respond: Direct questions from known contacts, requests with explicit deadlines, customer escalations, supplier confirmations requiring acknowledgment
• Review: Threaded conversations with mixed participants, anything from a new contact that requires a judgment call, financial notifications (invoices, payment receipts, tax authority correspondence)
• Archive: Newsletters, automated notifications, internal CCs where the user is not the primary recipient, calendar invitations the user didn’t accept
• Flag: Anything from a regulator, legal counsel, BaFin correspondence, anything matching a custom keyword list (the user’s company name plus common typo-squats is a common default)

The default configuration is built from anonymized patterns across our deployments. DACH SMEs have a different sender-importance hierarchy than a US startup — vendor relationship weight, Innungskrankenkasse communications, IHK notifications, and Steuerberater threads all carry different priors than they would for a US-based team. The classifier is fine-tuned for the DACH context, not Boston.

The output lands in a daily digest, delivered at 07:15 local time. Not a notification storm — one email, one Telegram message, one place. The digest lists the flag items first (because they are the ones that actually matter), then respond items with their drafts attached, then review items with a one-line summary, then archive count.

The Draft Loop: Voice Is The Hard Part

Triaging is mechanical. Drafting replies is where most AI email tools break. They either sound like a robot wrote them (because a robot did), or they sound like a slightly drunk version of the human (because the voice prompt is one line in a config file).

The Email Operations module spends two of its five deployment days on voice training. Not the model — the voice. The agent reads the last 200 sent messages from the user, clusters them by recipient type, and produces a voice profile broken into four registers:

• External formal (clients, regulators, suppliers): more formal German constructions, full sentence endings, conservative punctuation, no contractions
• External informal (known industry contacts): direct, short, occasional anglicism
• Internal (colleagues, direct reports): terse, can be telegraphic
• Family/personal (auto-detected by recipient domain, never trained on): bypassed entirely, flagged for human review

The voice profile lives as a system prompt that the draft loop reads on every generation. When the user edits a draft and sends it, the agent compares the edit to the draft and adjusts the voice profile incrementally. After a month of use, the agent drafts 60-70% of replies that go out unchanged. The 30-40% that get edited teach the model what it got wrong.

A common pattern: the agent’s first draft is too long. The human edits it down. The agent notices that respond items from a particular contact type tend to be edited to under 80 words. The voice profile updates. By week six, the drafts come in at the right length.

What It Doesn’t Do

Email Operations is not a personal assistant. It is not a CRM. It is not a calendar tool. Three things the module explicitly does not handle:

• Meeting scheduling from email threads — this is the Calendar module. The Email module will surface a scheduling request as a flag with a “schedule needed” note, but it will not propose times or check availability. That is a separate agent.
• Lead qualification and pipeline updates — the CRM module reads the same inbox and updates deal stages, but the Email module stays out of pipeline decisions. Mixing the two creates a situation where the agent’s email behavior changes when a deal moves to “closed-won,” which is the wrong incentive structure.
• Reading attachments in detail — the agent triages the email and notes attachments, but does not OCR contracts, parse invoices, or summarize PDFs in line. Document Operations handles that. Mixing it into the email loop means the wrong model gets used for the wrong job.

The discipline of keeping modules narrow is what makes the AI Office bundle defensible. The Full AI Office tier (€1,199/month) is not “do everything in one agent.” It is five agents, each with a clear boundary, plus a coordinator that routes between them.

The Cost Comparison

The honest comparison is not “AI vs human executive assistant.” It is closer to four options, each with a different cost structure:

Option	Monthly cost (DACH SME)	What you actually get
In-house EA (full-time, part-time, or shared)	€2,800–€4,500 + 25% overhead	Calendar, email, travel, occasional personal tasks, knowledge that walks out the door
Virtual assistant (offshore, hourly)	€800–€2,000	4-6 hours/day of email triage and scheduling, no context retention between sessions
SaaS AI email tool (Superhuman + SaneBox + ChatGPT)	€80–€150 per user/month, plus integration time	Email and calendar features, no orchestration, no voice training, no DACH context
AI Office Email Operations	€299/month	Voice-trained drafts, DACH-tuned classifier, three-loop architecture, integrates with Calendar and CRM modules

The math that matters: a DACH operations director spending 90 minutes a day on email triage is costing the company roughly €1,400/month in loaded salary. Email Operations at €299/month pays for itself if it saves 20 minutes a day. The median we see across deployments is 70-90 minutes a day saved, after a 2-week calibration period.

The cost is not the only consideration. The Virtual Assistant option looks cheaper on paper but breaks down when you add context — a VA does not learn your voice, does not know that “the Wolfsburg project” is a specific deal, does not remember that you already told supplier X the delivery would slip. The AI Office module has access to the same inbox, the same sent mail history, and the same calendar. Context is the moat, not the cost.

What Deployment Looks Like

A typical deployment takes 4-6 working days:

• Day 1: API access to the user’s mailbox via OAuth, read-only first, then read-write on a sandbox account. Voice training begins — 200 sent messages sampled and analyzed.
• Day 2-3: Classifier configuration. The agent reads 30 days of inbox history and surfaces a proposed configuration. The user reviews and adjusts. DACH-specific senders and keywords are added.
• Day 4: Voice profile finalized. Test drafts sent to a review address. The user marks 20 drafts as good, 20 as bad. The agent adjusts.
• Day 5: Three loops activated in observation mode. The agent triages, drafts, and surfaces follow-ups, but does not push to the digest yet. The user reviews everything in a dashboard.
• Day 6: Digest delivery goes live at 07:15 local. Daily review with the user for the first week to tune thresholds.

By day 14, the median user is reviewing drafts in 15 minutes a day and spending another 10 minutes on the morning digest. The previous 90-minute morning email block drops to 25 minutes.

The Takeaway

Email Operations is the most-deployed module in the AI Office bundle for a reason. Email is the highest-frequency operational task in any DACH SME, it has a measurable time cost, and the architecture for an AI to do it well is well-understood. The three-loop structure (triage, draft, follow-up) keeps the failure modes bounded, the DACH-tuned classifier handles the sender-importance hierarchy that US-trained tools miss, and the voice training means drafts get sent more often than not.

The €299/month is not the interesting part. The interesting part is what you do with the 70 minutes a day your operations director gets back. Most of our clients redirect that time to customer-facing work that was previously getting squeezed out — the kind of work that actually moves pipeline.

AI Office Email Operations deploys in under a week. The bundle starts at €299/month per module, and the full AI Office (all five modules plus a dedicated Hermes instance) is €1,199/month.