Automating email processing with AI is as tempting as it is scary. Tempting, because an overflowing shared inbox costs hours every week. Scary, because a wrong answer sent to a customer costs far more than the time saved.
There is a middle ground, and it's the one we recommend: AI sorts, qualifies and drafts. A human approves and sends.
Why not an AI that replies on its own?
Language models are very good at understanding an email and writing a plausible reply. Plausible doesn't mean right: an invented delivery date, a commercial promise nobody approved, the wrong tone with an unhappy customer. In a customer exchange, a single mistake can damage the relationship.
Keeping a human in the loop isn't a lack of ambition. It's what lets you roll out quickly, without waiting for the tool to be perfect, and then decide case by case what can go out automatically.
What AI does very well
Classify each request as it arrives
Quote, after-sales, billing, complaint, job application: AI reads the email and its attachments and files it in the right category, without relying on exact keywords.
Spot urgency
A customer threatening to leave, a blocked delivery, a close deadline: these requests jump to the front of the queue instead of waiting their turn.
Extract the useful information
Order number, customer reference, amount, date: AI extracts them and flags what's missing, so the missing document can be requested straight away.
Draft a reply
Based on your template answers and in your tone, with the details of the case. The person handling it only needs to read, adjust if needed, and send.
How to build the triage, step by step
- Analyse a real sample. A month of requests is usually enough to see the types, volumes and who handles what.
- Define categories and urgency rules with the people who answer today, not on their behalf.
- Gather your template answers. They give the AI its base and keep your tone.
- Run the tool in parallel. For a few weeks it classifies and suggests while the team carries on as before. You compare and adjust.
- Switch over once the classification is reliable, with a shared tracker: who is handling what, and since when.
Questions to ask before you start
- Where does the data go? Customer emails contain personal data. Know which service processes the text, where, and with what guarantees. Keep third-party services to the minimum.
- What can go out without approval? Nothing at first. Later, perhaps, an acknowledgement or a purely factual answer.
- How does the team correct the tool? Adding a category or changing a template answer should be possible without a contractor.
When it's not the right project
If you get a handful of requests a week, an inbox rule is enough. If every request calls for a fully bespoke answer, AI can help classify but won't save much time. And if you're looking for a public chatbot on your website, that's a different project.
This is what we build on our inbound requests projects. And if you're weighing several projects, our method for choosing what to automate first will help you decide.
Frequently asked questions
A plausible reply is not necessarily a correct one: an invented delivery date, an unapproved commercial promise, a badly judged tone. A single mistake can damage a customer relationship, hence the principle: AI sorts, qualifies and prepares, and a human validates and sends.
Classify each request as soon as it arrives, spot urgency, extract useful information such as the order number or customer reference, and prepare a reply from your template answers, in your company's tone.
Analyse a real sample of requests, define the categories and urgency rules with the people who reply today, then gather the template answers. The tool then runs in parallel for a few weeks while we compare and adjust, and you switch over once the sorting is reliable.
Related service · Inbound requests
Let's look at your shared inbox in 30 minutes.
A free, no-commitment call: we tell you what AI can sort and draft for you, and honestly if now is not the right time.