In short
- The Emails and Support agent, currently in preparation, will sort each message, create the ticket, prepare a draft in the sender's language and pass it to the right person. A person will approve it before it is sent.
- It will stop and hand over to you on delicate cases: a serious complaint, a legal threat, a GDPR request, health data, an unsupported language.
- The most solid measure we have found: +15% of requests resolved per hour, in a study published in 2025 covering 5,172 support agents at a single company. At your company, we measure before we promise.
The problem: an inbox everyone reads and nobody looks after
The contact@ address receives everything. A question about a delivery time. A supplier's invoice. A request for a quote. A complaint. Advertising.
Everyone can read it, so nobody is really responsible for it. Three things go wrong:
- Replies are late. A simple message waits, buried between two longer ones.
- Messages go to the wrong place. The invoice lands with the salesperson, the quote request with the accountant.
- The same replies are written ten times. Opening hours, order tracking, documents to provide: everyone rewrites them in their own way.
Let's take a typical scenario, made up for the example. An SME with 40 employees near Lille has customers in France and in Flanders. Two people share the contact@ inbox on top of their own work. On Monday morning, they start by sorting instead of replying.
What the agent will do: sort, prioritise, create the ticket, suggest a draft
An AI agent is a program that reads, prepares and suggests, following rules written with you. This one is in preparation. For each new message, it will have four tasks:
- Label the message: enquiry, complaint, invoice, prospect (a potential customer) or junk.
- Set a priority, following the rules you have chosen.
- Create or update the ticket, that is, the tracking record for the request.
- Suggest a draft reply based on your knowledge base: your documents, your frequently asked questions. Then pass the case to the right person, with a sentence explaining why.
It will also chase a ticket left unanswered beyond the time you promised. At the audit, we check how to connect the agent to your ticketing tool and to your documents.
To read messages and prepare drafts, the agent will be able to connect to Microsoft 365 or Google Workspace through their official interfaces (APIs). An official interface, or API, is the door the vendor provides so that software can read and write in your mailbox, with your permission. Both allow a draft to be created without sending it. At Microsoft, sending is a separate operation (Microsoft Graph). At Google, the draft receives the “draft” label (Gmail API). The draft will be able to wait in your mailbox, or in a dedicated approval screen: we will make this choice with you.
Replying in the customer's language: French, Dutch, Spanish
The agent will write the draft in the language of the message received. The languages covered will be written in the agent's rulebook, from the start of the project.
This is useful in Belgium. The Belgian Constitution (Article 4) recognises four language regions: the French-speaking region, the Dutch-speaking region, the bilingual region of Brussels-Capital and the German-speaking region.
It is useful in Spain too. Castilian is the official language of the State. The other Spanish languages are also official in their autonomous communities, that is, their regions, according to each one's statute (Spanish Constitution, Article 3).
A Belgian customer can therefore write to you in German. A Spanish customer can write to you in a language other than Castilian. If the rulebook does not cover that language, the agent will not take a chance. It will stop and hand over. Our advice: only add a language if someone at your company can read the draft before it is sent.
What stays human
The agent will prepare, the team will approve. At the start, every message sent will go through a person. And some cases will never be entrusted to it.
Never a commitment without a person. A discount, a guaranteed deadline, a change to a contract: the agent will be able to prepare a draft, but never send it alone.
The stop cases. The agent will stop and hand over to someone on your team when it spots:
- a serious complaint or a legal threat;
- a request to exercise GDPR rights, for example access to one's data or its erasure. The GDPR requires a reply without undue delay and at the latest within one month of receiving the request. This period can be extended by two months where requests are complex or numerous. The person must then be informed, with the reasons, within the first month (Article 12, text published by the CNIL). The clock starts when the message is received: the request must quickly reach the person in charge;
- health data or other sensitive data, such as political opinions or trade union membership. The GDPR classes them as special categories (Article 9);
- a doubt about the sender's identity;
- a language that is not covered.
Personal data. The agent will only process what it needs for the request. What is kept, and for how long, will be written in its rulebook. Every access will be recorded in a log.
Gaining autonomy, one level at a time
Autonomy is set action by action, over four levels:
| Level | The agent | The team |
|---|---|---|
| 0, shadow | proposes, does not act | does the work; we compare |
| 1, systematic approval | prepares every reply | approves, corrects or rejects |
| 2, approval by exception | sends on its own the standard replies written in the rulebook | approves the rest, checks a sample |
| 3, supervised autonomy | acts alone | reads the report, receives the alerts |
This agent will start at level 0, during a “shadow” trial period of a few weeks, with a length set with you. It will then move to level 1. Level 2 is only planned for standard replies to written cases: acknowledgement of receipt, opening hours, order tracking. The agent will only move up a level as needed and with your written agreement. It will automatically drop back a level after a serious alert.
As a precaution, a reply sent by the agent on its own will tell the recipient that they are talking to an AI system. The EU AI Act provides for this, unless it is obvious (Article 50).
What a published study measured: +15% of requests resolved per hour
The most solid measure we have found comes from three researchers, Erik Brynjolfsson, Danielle Li and Lindsey Raymond. Their study “Generative AI at Work” appeared in 2025 in The Quarterly Journal of Economics, an economics journal (published article). A first version from 2023 gave slightly different figures: +14% and +34% (NBER working paper 31161). We cite the published version.
The authors followed 5,172 technical support agents at a business software vendor. This vendor is in the Fortune 500, the ranking of the largest US companies by revenue (Fortune). Its customers are small and medium-sized US businesses. The agents answer them by instant messaging (chat). A generative AI assistant was given to them, in successive waves. It reads the conversation and suggests a reply to the agent in real time, and the agent stays responsible for the exchange. Results:
- +15% of requests resolved per hour, on average;
- +30% for the least experienced and least skilled agents;
- for the most experienced and most skilled, a small gain in speed and a slight drop in quality;
- more polite customers, who ask less often to speak to a manager;
- fewer employees leaving, especially among newcomers.
The authors suggest an explanation: the tool could capture and spread the ways of working of the best-performing agents. With it, the least skilled agents start to write like the best. An equipped agent with two months' seniority does as well as an unequipped agent with more than six months at the company. This is the role your knowledge base will play: the person who starts on contact@ will build on the replies of those who know the company.
Let's be honest about the limits. The study covers a single company, technical support by chat for US customers, and a tool different from the one we are preparing. It does not measure the emails of a European SME. This figure gives an order of magnitude, not a promise. Your own measurement will be made during the shadow period.
The indicators to track
The supervision dashboard, in preparation, will track four figures, included in the monthly report:
- Time to first reply: the time between the message arriving and the first reply being sent.
- Draft approval rate: the share of drafts that your team approves.
- Misrouted tickets: cases sent to the wrong person, then reassigned.
- Human time saved: the sorting and writing time the team no longer has to spend.
The shadow period will give the starting point. After that, these figures will decide on a change of level, not a feeling. And if the agent hits a repeated error, it will open a ticket with us itself: we will be alerted, and you will follow the ticket in your future client area.
Going further
- The agents: the four typical agents and what each will prepare for you.
- The method: from the audit to going live, with your two decisions.
- Supervision: the dashboard, the alerts and the monthly report.
- Data and security: where your data is processed and where the agent can be installed.
- The audit: a review of your tools, your message volumes and the agents that would be useful for you. It costs €490 excl. VAT, deducted from the quote if you sign it within 30 days.