Email automation with AI: how it works

You don't have to automate everything at once. There are three levels, and each one builds on the previous.

1. AI helps you write and summarize

An AI assistant such as ChatGPT, Claude, Copilot in Outlook or Gemini in Gmail summarizes a long email thread, drafts a first version of a reply or rewrites your text in a different tone. You still do everything yourself, just faster. This works per person, without integrations. Which tools are out there, you can read in AI tools for business.

2. AI sorts and routes a shared inbox

AI reads every email in info@ or service@, gives it a label (order, question, complaint, invoice) and puts it with the right person or in the right folder. What's urgent goes to the top. Nobody has to go through the whole inbox any more to see what's for whom.

3. AI processes the email from start to finish

An AI agent reads the email, extracts the data, looks up the customer and the order in your ERP or CRM and prepares a reply or a draft order. An employee approves with one click, or adjusts something first. This is where the big time saving is, because the looking up and retyping disappear as well.

A logical order: level 1 for everyone, and level 3 for the one or two email flows that take up a lot of time.

An email flow is a good candidate if a lot of it comes in, if the same kinds of email keep coming back and if you can check the data against a system. Think of:

• Orders by email or as a PDF attachment, which are now retyped into the ERP
• Requests and quote requests that need to go into the CRM or your own system
• Status questions such as 'where is my order?' or 'has my request arrived?'
• Invoices and supplier documents that need to be read in and booked
• Appointments to make, move and confirm
• Recurring questions about delivery times, terms or returns

Less suitable are emails where the conversation itself is the point: an escalation with an angry customer, a sensitive conflict, a negotiation that goes differently every time. There AI can at most summarize or draft a first version. The human writes and decides.

Not sure? Test the email flow with the seven questions in which processes to automate.

Behind an automated inbox there's no magic, just a series of fixed steps. In plain language:

1. 01The mailbox is connected. Through the Microsoft 365 or Google Workspace API, a workflow reads new email in a shared inbox, with only the permissions it needs.
2. 02AI recognizes the kind of email. An order, a status question, a complaint, an invoice or something else. If the AI is unsure, the email goes to a human.
3. 03The data is extracted. Customer name, order number, items, quantities, requested date. Also from a PDF attachment.
4. 04The AI looks things up in your systems. Does the customer exist? What's the status of the order? Is the price right? For that you connect AI to your ERP or CRM, see connecting AI to your systems.
5. 05A reply or entry is prepared. A draft email in your tone, based on your own terms, product information and earlier replies. That's called RAG: the AI answers from your knowledge, not from its memory. Or a draft order in the ERP.
6. 06A human approves. The employee sees the email, the draft and the data on one screen and sends it with one click, or adjusts something first. More on that in human in the loop.
7. 07Everything is logged. What came in, what the AI did, who approved it. That way you see what goes well and where the instructions can improve.

We usually build this as a workflow in n8n, with AI at the points where something needs to be read, written or assessed. n8n can run on your own server, so the email stays within your own environment. How n8n compares to other tools, you can read in n8n, Make or Zapier.

Three NinA clients that have email and documents processed by AI:

• Van Berkel Professionals received more than 100 requests a day by email and form. These were retyped into their own system by hand. Now the incoming flow is read, filtered and placed in their own systems, and a human approves every request. That saves 192 hours a month.
• An enterprise customer service team receives 6,000 customer questions a month. An agent prepares the answer before the ticket opens, with the right order details. The share of questions with a prepared answer went from 30% to 80%. An employee always checks and sends.
• EasyToys received product information in supplier PDFs and emails, which was retyped and translated per language. AI now reads them, writes the product copy in six languages and prepares everything in the ERP. That saves 132 hours and € 6,600 a month.

The pattern is always the same: AI does the reading, looking up and preparing, the human keeps the final say.

Don't start with the technology, start with the time. In our APO method you give an email flow three numbers: how long the work takes each time, with how many people, and how often a month.

A worked example: a task of forty minutes, with two people, eighty times a month. That's 1,280 hours a year. At € 45 an hour, this task alone costs € 57,600 a year. This is a worked example, not a client figure.

AI never takes all of that over. Exceptions, checks and maintenance remain human work. So calculate with the share AI can really take over, and look at the expected scenario, not the best case.

You can run the numbers yourself with the AI savings calculator. How to turn that into a full business case, you can read in the business case for AI.

Email is full of personal data: names, addresses, phone numbers and sometimes sensitive information. So arrange this up front:

• Business accounts, not free consumer versions. Business versions from the large providers don't train on customer data by default.
• A data processing agreement with every party that processes the email.
• Data in the EU where possible, or on your own servers.
• Least privilege. The AI only reads the inboxes it needs and only prepares what's needed. Sending goes through a human.
• A log of what the AI did and who approved what.
• Agreements on retention. How long are emails and drafts kept, and who has access?

More on the GDPR in using ChatGPT safely and using AI safely.

When it goes wrong in practice, it's rarely the model. It's the setup:

• Sending automatically without a check. A wrong answer sounds just as convincing as a right one, see AI hallucinations. Start with drafts that a human sends, and only give more room when the numbers show it goes well.
• No route for exceptions. Every inbox has emails that don't fit anywhere. Decide up front what counts as an exception (the AI is unsure, a complaint, an amount above a threshold) and who it goes to.
• No owner. Products, prices and terms change. If nobody is responsible for the instructions and the knowledge the AI answers from, quality slowly declines. Appoint someone who watches the numbers and updates the instructions.

More on what can go wrong in why AI projects fail.

We start with the AI Quick Scan: one day on site with you, with the people who handle the inbox. We map which emails come in, how much time they take and which ones lend themselves to AI. Within a week you have a business case per opportunity and a roadmap. The Quick Scan costs € 6,000.

After that we build within the AI Partnership, from € 4,500 a month, for a minimum of twelve months. We build AI agents and workflows connected to your mailbox and systems. In the first quarter, the first process goes live. We only start if the partnership pays back at least 2x in the first year.

What goes live, we keep running with a maintenance contract, including AI usage: Bronze € 245 a month (on your own servers), Silver € 445 or Gold € 895 a month. All amounts exclude VAT.