Solution · Email and client requests
AI email triage: the agent sorts, you approve
An AI agent reads every incoming message, labels it, pulls out what matters (deadline, amount, attachment) and drafts a reply in the client’s language. Your team checks it, then sends.
- Drafts approved by a person
- Replies in French, German, English or Italian
- Swiss hosting available
What an email triage agent does
In a small business, the shared inbox is often the first bottleneck. Client questions, supplier invoices, documents, reminders and advertising all land in the same place. Someone has to read everything, decide who handles it and answer the questions that come back every week.
A triage agent does that first pass. It reads each email, recognises the type of request, picks out the useful details and forwards the message to the right person with a short summary. For routine questions, it drafts a reply based on your templates and documents.
Nothing is sent without your say-so. You decide which categories may get an automatic reply, such as an acknowledgement, and which always go through a person. The agent works inside the email you already use, so your team keeps its habits.
The problem
Where the time goes today
Everything lands in one place
One info@ address receives client requests, invoices, job applications and advertising. Someone sorts it by hand, several times a day.
The same questions, again and again
Opening hours, documents to send, the status of a file: the answer already exists somewhere, but it has to be found and rewritten every time.
Messages slip through
When someone is away, urgent emails wait. A deadline or a complaint can go unnoticed for days.
Interactive demo
Try the agent on an inbox
Pick a fictional company and run the agent: it labels every message, extracts the key details and drafts replies. You can also paste an email of your own.
Email triage agent
Reads, sorts, summarises and drafts your replies. You keep the final word.
Inbox
Select an email, or run the agent.
Try it with your own email
Everything runs in your browser: nothing is sent or stored. Simplified demo: answers are prepared in advance or computed by simple rules in your browser. In production, a language model suited to your data does this work, and someone on your team signs off.
I want this for my businessWhat’s happening, step by step
-
Emails arrive Input
The agent watches your inbox, like a colleague opening the post the moment it lands.
-
The AI understands AI
A language model reads the message in its own language and recognises what it is: an invoice, something urgent, a new client…
-
It pulls out what matters AI
Deadline, amount, file number, attachments: everything important is picked out and summed up in one line.
-
To the right person Code
Rules agreed with you send each request to the right person. No AI here: clear rules.
-
A draft is ready AI
For routine requests, the agent writes a reply in the client’s language, based on your templates.
-
You approve Human
Nothing goes out without your say-so: you read, adjust if needed, and send.
Run the demo: each step lights up here.
ModelApertus 70B · Swiss open source
Data processedA server in Switzerland or yours
Picture a colleague who reads every email the second it arrives, understands what it’s about, passes it to the right person and drafts a reply. It never sleeps, never loses anything, and always waits for your go-ahead.
Words to know
- Language model
- The “brain” that reads and writes text: GPT, Claude, Gemini, Mistral, Apertus…
- AI agent
- A program that uses a language model to complete a multi-step task, with your tools.
- Human approval
- The moment a person reads and decides. At NotHuman, it’s the default.
Why it’s safe
- Nothing goes out without your say-so: the agent prepares drafts, someone on your team approves them.
- Who gets what follows rules agreed with you, written in code: the AI does not decide.
- When in doubt (an answer in the wrong format, low confidence), the email goes to a person.
- You choose where the AI runs and where your data is processed.
The agent’s pipeline
- Input Mailbox Microsoft 365 (Graph API), Gmail or IMAP. A new email triggers the pipeline.
- Code Pre-processing Clean-up, language detection, masking of sensitive data where needed.
- AI Classification and extraction The model returns schema-validated JSON: category, priority, fields, summary.
- Code Routing rules Your business rules: who gets what, urgency thresholds, office hours.
- AI Drafting A draft in the sender’s language, from your templates and knowledge base.
- Human Approval Draft placed in Outlook or Gmail, or in an approval queue.
AI reads and writes. Code checks. Humans decide.
Stack
Guardrails
- Structured output: if the model’s answer doesn’t match the schema, the email goes to a person.
- Hidden instructions (prompt injection): the agent can label and draft, never send on its own.
- Confidence threshold: below 0.8, the message is flagged for a person.
- Full audit log: every decision is recorded and searchable.
What the agent produces (JSON)
{
"email_id": "f2",
"language": "de",
"category": "urgent_tax_office",
"priority": "high",
"summary": "Reminder from the tax office, due 5 November. The client thinks he has already paid.",
"fields": {
"client": "Thomas Keller",
"deadline": "2026-11-05",
"action": "verify_payment"
},
"route_to": "marc@alvelia-fiduciaire.example",
"draft": {
"language": "de",
"status": "awaiting_approval"
},
"confidence": 0.94
}Choose the engine
How we set it up
- Day 1 Free audit 30 minutes to look at your volumes, your categories and what must stay human.
- Weeks 1–2 Test on your emails We calibrate categories and rules on a sample of your messages and measure the quality.
- Weeks 2–4 Go-live The agent is connected to your email, with your approval rules.
- Every month Care Monitoring, adjustments and a monthly report.
What we need from you
- Access to the mailbox, read-only to start
- Your categories and who handles what
- A few template replies you already use
What we set up
An agent built around the way you work
Your categories, your reply templates, your approval rules. Not a generic tool.
Labels and priorities
Each message gets a category (client, supplier, document, complaint…) and an urgency level. Sensitive messages are flagged.
Key details extracted
Deadlines, amounts, file numbers and attachments are picked out and added to the summary, or sent to your tracking sheet.
Draft replies
For routine requests, the agent drafts a reply in the client’s language from your templates. You read, adjust and send.
Routed to the right person
Requests are assigned according to your rules (topic, client, language), with a three-line summary.
Acknowledgements
If you want, an acknowledgement goes out straight away, with a realistic response time.
Tracking and figures
A monthly report shows volumes per category, response times and the questions that come up most.
How we work
How it goes
-
Free audit
In 30 minutes we look at your volumes, your categories and what must stay in human hands.
-
Test on your emails
We test the triage on a sample of your messages and measure the quality before anything goes live.
-
Go-live
The agent is connected to your email with your approval rules. We aim for 2 to 4 weeks after the audit.
-
Monthly care
We monitor triage quality, adjust the rules and send you a report every month.
Our commitments
Your email stays under your control
A human approves
By default, no email is sent without approval. You choose what may go out automatically.
You choose where your data is processed
Leading models or, for sensitive data, open-source models such as Apertus, the open Swiss language model, hosted in Switzerland.
Built with Swiss data protection in mind
A data processing agreement, only the data the task needs, logged access. We also help you inform your clients.
Frequently asked questions
Does the AI send emails without me knowing?
No. By default, the agent prepares drafts and a person approves them. You can allow automatic sending for simple cases, such as an acknowledgement, and switch it off at any time.
Which email systems does it work with?
We connect Microsoft 365 (Outlook), Google Workspace (Gmail) or a standard IMAP mailbox. Your team finds the labels and drafts in the email they already use.
My emails contain client data. Is this compatible with Swiss data protection law?
It can be, if the processing is set up properly: a data processing agreement, only the data needed, a provider chosen with care, and clients informed. For sensitive data, we can use an open-source model hosted in Switzerland. See our guide ChatGPT at work and data protection.
What if the AI picks the wrong category?
It happens, especially at first. That is why we test on your own emails first and flag uncertain messages for a human check. Every correction is used to improve the rules.
How long does it take to set up?
We aim for a first version in service 2 to 4 weeks after the audit, depending on the number of categories and tools to connect.
How much does it cost?
The price depends on your email volume, the number of categories and the tools to connect. The audit is free; you then get a clear quote with a set-up cost and a monthly fee. Request a quote.
NotHuman
How many hours does your team spend sorting email?
In a free 30-minute audit, we estimate the time you could save and propose a concrete plan.