Office & managementNewVerified Real deployment at a named or precisely described company, with a measured result from a primary source.

AI sorts the shared customer inbox and routes each email (order, quote, after-sales, invoice) to the right person (Safran Landing Systems, Dossche Mills)

Added 2026-10-29 · Last updated 2026-10-29

Manufacturers in this case: Dossche Mills · Safran Landing Systems · all manufacturers using AI

Role

Customer service and sales administration (shared mailboxes such as orders@ or sales@) · Typical users: customer service or sales administration team, inside sales, office manager, after-sales coordinator

The problem

Orders, quote requests, delivery questions, complaints, invoices and spam all arrive in the same shared mailbox. Someone has to read each message, decide what it is and forward it, often first thing in the morning, so a message that arrives in the afternoon waits until the next day and urgent orders get lost. Safran Landing Systems, a maker of landing gear and braking systems for aircraft, receives 15,000 to 20,000 emails a month across five countries: requests for quotes, purchase orders, delivery requests and technical questions. Dossche Mills, a Belgian family-owned flour miller founded in 1875, received orders and requests in several multilingual shared inboxes across Europe, triaged them by hand across Outlook, SAP and SharePoint, and saw the workload grow as experienced employees left without being replaced. In a 45-person French components maker, the contact@ and orders@ addresses received 200 to 300 emails a day; a sales assistant spent on average 3 h 15 every morning sorting them, and in the quarter before the project 9 orders were handled late or lost, after which two customers went to competitors.

How AI helps

The AI reads each incoming email in full (subject, body, sender, attachments), puts it in a category (order, quote request, after-sales, accounting, supplier, other), pulls out the key details (customer, product reference, quantity, urgency) and sends it to the right team or person. Urgent orders are flagged at once instead of waiting for the morning sort. When the AI is not confident, or the message mixes several requests, it goes to a person to decide. In the French SME, product references are checked against the real catalogue and order data goes into the ERP only as a draft; at Dossche Mills the system also prepares draft replies and order entries in SAP; at Safran Landing Systems it works in English, French and Chinese for five sites.

Step by step

Count a normal week of the shared mailbox: emails per day, the mix by type, time spent sorting, and how long a customer waits for a first reply.
Write down 5 to 8 categories and, for each, who should receive it and what counts as urgent; this is the rulebook the AI follows.
Test on 50 past emails: compare the AI's category and routing with what your team actually did, and fix the category definitions where they disagree.
Go live with a human check: the AI routes only when it is confident and sends the rest to one person; orders reach the ERP only as drafts.
Every week, review misrouted emails and corrections, and adjust the rules.
After two to three months, measure again: sorting time, first-reply time, lost or late orders.

Tools

Tungsten TotalAgility (AI email classification and routing), delivered by DS Informatique: used at Safran Landing Systems
Tekst (email classification, routing, draft replies and SAP order automation): used at Dossche Mills
Custom AI agent built by Vantacrew on Claude models, connected to the shared mailbox and the ERP: used at the French components maker
Other tools to evaluate: AI classification in Microsoft Power Automate (AI Builder) or in your help desk or CRM, shared-inbox tools such as Emailgistics
To test the idea without new software: 30 to 50 anonymized emails and any LLM (ChatGPT, Claude, Copilot, Gemini) with the prompt below

In the AI tools index: Tekst · Tungsten TotalAgility · Vantacrew · all AI tools used by manufacturers

Result

Safran Landing Systems (aircraft landing and braking systems), Tungsten TotalAgility delivered by the integrator DS Informatique: 70% reduction in manual email processing, more than 20,000 emails a month handled across five sites, emails classified automatically in English, French and Chinese; described as Safran's first successful AI pilot for sales administration (Tungsten case study, a vendor claim). Dossche Mills (flour and bakery ingredients, Belgium), Tekst: 'high accuracy' on email classification, order volume maintained with a reduced workforce, order processing time per case 'meaningfully' cut; no figures published (Tekst customer story). French components maker (45 people, not named), AI agent built by Vantacrew, measured six months after going live at the end of 2025: daily sorting time from 3 h 15 to 53 minutes (-73%), first-reply time from 26 hours to 4 h 30, lost orders from 9 to 1 per quarter, 89% of emails classified without human help, 2.3% routing errors, ERP re-keying down 64%; cost €9,400 setup plus €410 a month, payback 4.2 months by the vendor's calculation; the assistant moved to sales follow-up and nobody was laid off (Vantacrew case study, a vendor claim).

As reported by the source; most figures are vendor or self-reported and not independently verified.

Human review: cautions

The Safran and Dossche Mills results are published by the software vendors, and Dossche Mills gives no figures. The French SME is not named and its figures come from the firm that built and sells the agent.
A misrouted order is worse than a slow one: keep a human check on low-confidence emails, high-value orders and messages that mix several requests.
Orders extracted by AI should reach the ERP as drafts; a person confirms quantities, prices and dates before production or shipping.
Customer emails contain personal and commercial data: check where the AI service processes and stores them, and what your contracts and privacy rules allow.

Try it yourself

You are a customer service coordinator at a manufacturer.

I will paste: (1) a list of our teams and what each one handles [for example: orders, quotes, after-sales, accounting, purchasing]; (2) what counts as urgent for us; (3) 30 to 50 anonymized emails from our shared mailbox [sender company, subject, body, attachment names].

Do this:
1. Propose 5 to 8 email categories with a one-line definition each and the team that should receive it.
2. For each email, give a table row: email number, category, team, urgent yes/no, customer, product references and quantities exactly as written, confidence (high / medium / low), and one line explaining why.
3. List the emails that mix several requests or that you could not classify with high confidence, and say what a person should check.
4. Draft a short acknowledgment reply for each category, to be sent only after a person approves it.
5. List the rules you would add to reduce low-confidence cases.

Do not compute totals or percentages yourself; tell me which spreadsheet formulas to use. Use only the data I paste in; if something is missing, write 'not found' or list it as a question instead of guessing. Do not send, order, post or change anything in any system: produce drafts for my review and approval.

Want more like this? Join the newsletter

Related cases

All 12 Office & management use cases →