Lyron
AI

AI Email Assistant for Shared Inboxes

The assistant reads every incoming message, pulls the facts out of your systems and puts a finished reply draft next to it. Whether it sends that draft itself is decided by a threshold you set per category – anything below it stays a draft.

Context

Writing the reply is not the bottleneck

Most replies in a shared inbox are three sentences long. The time does not go into those three sentences, it goes into everything before them: which order is this about, what does the warehouse say, has the invoice gone out, and is anyone actually responsible or will nobody read this today? Someone working an info@ inbox opens three windows for an answer they could recite from memory.

That is exactly where automation gets difficult. A language model writes those three sentences effortlessly – and it writes them just as happily when it never found the order number. It does not fail loudly, it keeps composing. A fluent text is no evidence that the answer is correct. A delivery date wrongly promised in your name, sitting in your customer's inbox, costs more than the twenty minutes it saved.

The real work therefore sits in the question of when a draft may leave without a human, not in writing it. We derive the assistant's confidence from things that can be checked: was the sender matched to exactly one record, does every statement in the draft come from a connected source, does the message contain a single request? That produces a confidence score – and only a threshold you set per category decides whether it is sent.

Use cases

Which emails are worth it

We start with the category that occurs most often and has the clearest source. Every further category later uses the same mechanics of draft, confidence score and threshold.

Most common starting point

Order and delivery status

The most common case: where is my delivery? The answer sits in the ERP, not in anyone's head – the assistant fetches it and writes it down.

Order numberDelivery dateTracking numberPart shipment

Sending documents again

Invoice copy, delivery note, data sheet, certificate. The assistant finds the file belonging to the record and attaches it – or says that it does not exist.

Invoice copyDelivery noteData sheetCertificate

Confirming and moving appointments

The requested slot is checked against the calendar and two alternatives are added. Whether it books or only proposes depends on the category.

Slot requestAlternativesCancellationCalendar check

Standard answers from your knowledge base

Warranty period, payment terms, spare parts, directions: answers that are documented somewhere and still get typed out fresh every day.

WarrantyPayment termsSpare partsDirections

Sorting what people have to handle

Complaints, cancellations, press, job applications: no draft, but category, urgency and owner are settled before anyone opens the message.

ComplaintCancellationEscalationAssignment

Internal inboxes

IT support, HR, purchase requests: the same mechanics turned inward, with far less risk – a good place to start.

IT supportHRPurchase request
Example

One morning in a shared inbox

Five messages as the assistant prepared them: the mail on the left, the draft and its source in the middle, the confidence score against the category threshold on the right.

service@ · Tuesday morning 5 of 34 messages
  1. M. Bergmann · order 4127 When is my delivery arriving? Order status

    Your consignment left our warehouse on 24 July and is due for delivery on 29 July. The tracking number is 003401…

    ERP · delivery note 4127-1
    94threshold 90
    sent
  2. Accounts, Ritter Ltd Please resend invoice INV-20871 Document request

    Attached is invoice INV-20871 of 8 July as a PDF. The outstanding amount is due on 7 August.

    Document store · INV-20871.pdf
    96threshold 90
    sent
  3. S. Karadeniz · service contract Can we swap Monday for Thursday? Reschedule

    Thursday works for us at 9:00 or 14:30. Let us know which suits and we will move the appointment.

    Calendar · engineer North, week 31
    83threshold 85
    waiting for approval
  4. P. Ostermann · order 3988 Unit failed again after two weeks Complaint

    Draft ready: receipt acknowledged, warranty runs to 03/2028, engineer visit proposed.

    ERP · order 3988, warranty active
    91threshold: always
    a person decides
  5. Purchasing, Halbach & Sons What does servicing twelve units cost? Pricing and discount

    No draft. Prices are negotiated by sales, not by the assistant – the message goes to the North team with a summary.

    Source not approved
    never automatic
    forwarded
goes out automatically waits for a person no draft, forwarded line in the bar = threshold

Row four is the instructive one: a high confidence score and still no send. Complaints are a blocked category, so the score is irrelevant there.

Row three sits two points below the threshold and waits, even though the draft looks usable. The calendar had reported only one of the two requested slots unambiguously, and that gap feeds straight into the confidence score.

How it works

From an incoming message to an approved reply

  • Connect the inbox and fence it in

    The assistant gets access to exactly one shared inbox, never to personal ones. Folders, labels or the recipient address decide which messages it sees at all.

  • Identify the request and match the sender

    Subject and body produce a category, an urgency and the references: customer, order or invoice number. Without an unambiguous match to a record there is no draft, only a question back.

  • Fetch facts instead of composing them

    Only what comes from a connected source may enter the draft – ERP, CRM, document store, knowledge base. Every statement carries the record it came from, visible internally.

  • Score the confidence and apply the threshold

    Matching, source coverage and how clear-cut the request is produce a number. Above the category threshold the reply goes out; below it, or in a blocked category, it stays a draft with the reason attached.

  • Send, write the record, adjust

    The reply leaves from the shared inbox and the record lands in your CRM or ticket system. Every human correction is logged – that is what you use each month to decide which threshold moves.

Impact

What changes day to day

Today

  • Everyone reads the same inbox, nobody owns it
  • One status answer means opening three systems
  • Urgent messages sit between newsletters and invoices
  • Reply quality depends on who happens to have time
  • On Monday the inbox is full before anyone starts

With the assistant

  • Every message arrives with category, owner and deadline
  • The draft already carries the facts from your systems
  • Routine answers are out before anyone has read them
  • Uncertain cases sit as drafts, not as errors in sent items
  • What went out is logged with its source and its score
Limits

Where the assistant stops

Four things this product does not solve. We raise them before the quote, not afterwards:

  • Below a certain volume it does not pay off. If your shared inbox gets twenty messages a day and nearly every one is a one-off, you will spend more time on rules and checking than answering ever cost. It carries where a handful of categories cover most of the traffic. If that is not your situation we say so in the intro call – and you keep the €2,490.
  • The confidence score measures evidence, not truth. It says how well an answer is covered by sources, not whether the record is right. If the ERP holds the wrong delivery date, the assistant will state it with complete composure and no threshold in the world will notice. That is why we start with categories where a mistake is embarrassing rather than expensive.
  • What is written down nowhere cannot be answered. If the delivery date lives in a spreadsheet on someone's desktop or in the scheduler's head, there is no draft, only a question back. An inbox assistant makes the gaps in your master data visible within days. Useful, but rarely comfortable.
  • Attachments and catch-all messages stay manual. A message with three separate requests, a chain forwarded twelve times, a photo of the broken part: the assistant deliberately does not answer these, it only files them. Reading attachments properly is a project of its own, not part of the entry price.
Systems

Fits your inbox

Microsoft 365OutlookGmailSharePointHubSpotDynamics 365ERP systemsn8n
Scope and price

Scope and price

The entry price covers one shared inbox with up to five categories and one source of facts. What moves the price beyond that, we say before the quote.

from €2,490 one-off
  • One shared inbox connected, for reading and sending
  • Up to five categories, each with its own threshold and tone
  • One source of facts for the drafts (ERP, CRM or knowledge base)
  • Drafts placed in the inbox, with the source noted internally
  • An approval route with a named owner per category
  • A log of every decision: category, score, sent or approved
  • Documentation, handover session and 30 days of support

What increases the price

  • More than five categories, or several inboxes and sites
  • Reading attachments rather than only the message body
  • Replies in several languages, each reviewed in its own right
  • Write actions: raising tickets, booking slots, starting returns
  • Connecting an industry or legacy system with no usable interface

Several inboxes, reading attachments and write actions in your operational systems typically land in the range of our AI Agent Pilot package from €4,900. We quote the binding fixed price after the intro call.

All prices excl. VAT · operation and further development optionally via a support package

Included

What you get

  • A working assistant on your inbox

    Connected, tested on real messages and signed off in writing

  • Documented categories and thresholds

    What each category may do and from which score – changeable without asking us

  • Decision log

    Per message: category, confidence score, source, sent or approved

  • Handover and a first correction round

    How your team approves, corrects and when a threshold should be moved

Questions & answers

Frequently asked questions about the inbox assistant

Only where you explicitly allow it, and only above the threshold you set for that category. Everything else sits in the inbox as a finished draft waiting for one click. In practice we start with every category in draft mode and switch the first one live once you have spent a few weeks seeing what the assistant would have written.
Not from the language model alone, because a model's self-assessment is worth little. The score is built from things that can be checked: whether the sender was matched to exactly one record, whether every statement in the draft comes from a connected source, whether the message holds a single request, and how clearly the category was recognised. Which criterion carries how much weight is agreed with you and written down.
No. We connect one shared inbox, not the personal mailboxes of your staff. Within that inbox access can be narrowed further, for example to folders, labels, a recipient address or a time range. What is read, which fields leave the system and how long anything is retained is documented before go-live.
The EU AI Act requires transparency wherever people deal with an AI system, and a reply released by a person before sending is assessed differently from a fully automatic one. Regardless of the legal reading we recommend a short note in the signature as soon as a category answers automatically. It costs nothing and saves awkward questions. The binding assessment for your case belongs with your legal advisers.
No, those are separate products. A personal chat subscription includes no API access and is not an account you leave running unattended on a company inbox. We set up a dedicated API or cloud account with a data processing agreement, in an EU region and with your data excluded from the provider's training. Which provider that is we decide with you, weighing privacy, cost and the quality of its German.
Then something about the category is wrong, and that should be visible. Every correction is logged next to what the assistant had written and what actually went out. If a category is corrected regularly we raise its threshold, improve its source or take it out of automatic mode again. The assistant does not quietly learn from your corrections, and that is deliberate: otherwise nobody could explain why it answers differently this week than last.

Which emails do you answer for the hundredth time?

In the free intro call we take one real morning out of your shared inbox and sort it into categories. Afterwards you know how much of it could be answered automatically at all – and whether that adds up for you.

Book a free intro call
Practical guide

Where AI email assistant creates value in everyday work

The assistant sorts email, detects intent, researches approved knowledge and prepares traceable replies or next steps.

Three concrete operating scenarios to compare with your own process.
01

Prioritise the inbox

Intent, customer, urgency and responsible team are detected from content and context.

02

Prepare replies

Approved information is combined with sources into a reviewable draft.

03

Update business records

Tickets, CRM or tasks receive the relevant data without manual re-entry.

A strong fit when …

Recurring requests can be answered from approved knowledge while uncertain cases are handed to people with clear context.

  • You handle recurring emails using repeatable rules.
  • The intake, target system and accountable business role can be named clearly.
  • Exceptions are allowed to remain visible and move to people deliberately.
Transparent potential estimate

Estimate time savings with your own volume

The calculator uses 5 minutes today and 2 minutes after automation as fixed example assumptions. It does not replace process analysis.

Illustrative estimate based on the visible assumptions — not a guarantee.

55Hours per month
660Hours per year
Additional measures after launch Handling time Escalation rate Source coverage
Frequently asked questions

What decision-makers should know before starting

How does AI email assistant work in practice?
A message arrives in an approved shared inbox. The workflow then validates the required data, runs approved steps and routes exceptions to the responsible person with context.
Which systems can be connected?
Typical integrations include Microsoft 365, Gmail, CRM, SharePoint, Ticketsystem. The decisive factors are a stable interface and clearly defined ownership of each data field, not a specific tool.
Which tasks deliberately stay with the team?
Sensitive, legally relevant or insufficiently supported answers remain drafts with sources for staff review.
How is the automation introduced?
We start with a narrow knowledge and task scope, test real examples and expand only after documented approval. A tightly scoped first process typically takes 3–6 weeks; scope, interfaces and approvals determine the actual plan.
How can the benefit be measured?
Before implementation we record volume and current handling time. After launch we also compare Handling time, Escalation rate, Source coverage. The calculator on this page is a transparent estimate, not a promise.
Content reviewed on 26 July 2026 About Lyron AI