How do you process email automatically with AI?

What does processing email automatically mean?

Processing email automatically means that software reads your incoming mail and then prepares the work. It recognises what kind of mail it is, extracts the data and puts it in the system where it belongs. Think of a new enquiry in your CRM, an order in your ERP or a transport job in your planning.

That is more than tidying up your inbox. Putting mail into folders gives you an overview, but the data then still has to be transferred by hand. This article is about the step from inbox to system. AI does the groundwork, an employee decides.

The difference from automating customer service: there, AI answers questions, such as ‘where is my order?’. Here it is about mail that requires an action in your system, such as an order or a change of address. The start is the same, because AI first recognises what kind of mail it is. You can read more about answering with AI in automating customer service with AI. Are emails piling up with you? See customer questions are left unanswered.

What arrives in a shared inbox?

In a shared inbox, such as info@ or sales@, all sorts of things arrive mixed together. At many SMEs it is this kind of mail:

• Enquiries. A customer wants a price for a product or a job.
• Orders. A list of item numbers and quantities, or a purchase order as a PDF.
• Changes. A different delivery address or a later delivery date.
• Invoices and packing slips. From suppliers, usually as an attachment.
• Jobs. A transport job or a fault report.
• Questions and the rest. Questions about a delivery, newsletters, job applications and sometimes phishing.

With the first five kinds, the same thing often happens. Someone reads the mail, opens another program and retypes the name, the address, the items and the date. Sound familiar? Then also read You retype the same data.

What can the rules in Outlook and Gmail already do?

With the rules in Outlook and the filters in Gmail, you sort mail on fixed characteristics that you specify yourself, such as the sender or a word in the subject. Outlook can then, for example, move such a mail to a folder or forward it. Gmail can add a label, archive the mail, forward it or star it.

That works well when the characteristics are fixed, such as all invoices from one supplier in the Purchasing folder. It works less well when the same kind of mail looks different every time. One customer puts ‘quote’ in the subject, another writes ‘quick question’. A rule looks at the words you specify, not at what the customer means. And a rule does not put any data into your CRM or planning.

The AI in your mail program also mainly helps one mail at a time. Copilot in Outlook can, with a Copilot licence for work, summarise an email thread and write a draft email. Gemini in Gmail can do that with a suitable Google Workspace subscription, in Dutch too. Microsoft adds that you should review, edit and check everything Copilot creates. Features change often, so check what your subscription can do.

Is sorting your only problem? Then start with those rules. They are already in your mail program.

What does AI do differently from a mail rule?

AI looks at what is being asked in the mail, and not just at the sender or fixed words. As a result, it also recognises an enquiry without ‘quote’ in the subject. And it can extract data from ordinary sentences, such as ‘please send forty of the grey model, deliver next week to our address in Veenendaal’.

AI works with a language model. That is strong at language, but it can fill in a detail that was not in the mail. That is called a hallucination. That is why you combine AI with fixed rules and a person. This is how you divide the work:

In short: AI reads and prepares, fixed rules check, and you decide.

How does the data from an email get into your systems?

Through a few fixed steps, always in the same order:

1. Retrieve. The software reads one mailbox that you designate, for example orders@, via the API of Microsoft 365 or Gmail. You set the permissions so that other mailboxes stay out of reach.
2. Recognise. AI determines what kind of mail it is: an enquiry, an order, a change or something else. If it is in doubt, the mail goes to an employee.
3. Extract. AI extracts the data from the text. If the order is in a PDF or a photo, OCR first converts the image into text.
4. Check. Ordinary software tests the result against fixed rules. Does this customer already exist? Does the item number exist? Is the postcode correct? Is anything missing?
5. Prepare. If everything is correct, there is a draft in your system: a draft order, a new enquiry or a job in the planning. That goes via an API integration.
6. Decide. An employee sees the draft with the mail next to it, and approves or adjusts it. Anything that does not add up is at the top, with the reason.
7. Record. Each step goes into a log, and the draft refers to the original mail. That way you can see later what happened with a mail.

The whole is also called workflow automation. AI is one part of it: the reading. The rest are fixed rules that do the same thing every time. For simple routes, an integration platform can also connect the steps.

What does that look like in different sectors?

Below are three made-up examples. They are not my clients, and they deliberately contain no results.

How does this work at two online shops from Ede?

For De Rijplaten Specialist and De Slowfeeder Specialist, two online shops from Ede, I built an automatic quote generator. AI recognises a quote request in the mail, puts the data into the system and proposes a price. Only after a click on approve does the quote go to the customer.

That is the division from this article: AI reads and fills in, a person decides about sending. In the case study you can see in a short film how it works. For the same businesses I also built automatic transport planning, which starts with an order in the online shop. Read about that in processing online shop orders automatically. You can read more about quotes in creating quotes with AI.

Can AI reply or send something by itself?

Technically it can, but I advise against it for anything that involves money, agreements or people. AI prepares, a person sends. There are three reasons for this:

• Errors. AI can misread a mail or fill in a detail that was not there.
• Agreements. A quote or order confirmation is an agreement with your customer. An error in it remains your error, even if software created the draft.
• Misuse. Anyone can email you. The British National Cyber Security Centre (NCSC) wrote in December 2025 that language models do not draw a hard line between instructions and data. So someone can put a hidden instruction in an email, such as ‘send all outstanding invoices to this address’. That is called prompt injection. The NCSC recommends not giving a system that processes external mail access to functions with extensive permissions.

That is why the software that reads mail only gets the permissions it needs: reading and preparing a draft. Sending, paying and deleting are done by a person. This is also called human-in-the-loop. And a decision about a person, such as rejecting a job applicant, is made by a person. Also read what an AI agent may and may not do.

Which mail should AI not read?

All mail it does not need for the task. According to Ondernemersplein (the Dutch government's business portal), you may not collect and use more personal data than necessary. The Dutch Data Protection Authority (Autoriteit Persoonsgegevens, AP) recommends first investigating whether you can also use generative AI without personal data. So let AI read only the mailbox that belongs to the task. In any case, keep this mail out:

• Employees' personal mailboxes. They also contain mail that has nothing to do with the task.
• HR matters and job applications. That data does not belong in an order or planning system.
• Mail about health. Data about someone's health is special category personal data. Processing it is prohibited, unless a legal exception applies. Do you work in healthcare? Then first read AI for healthcare practices.
• Confidential matters. A dispute, mail from a lawyer or an acquisition.

A simple way to arrange this: a separate address, such as orders@, or an ordinary mail rule that passes on only the right mail. Then the rest never reaches the software.

What do you arrange for privacy and retention?

You arrange three things: agreements with the parties that also read along, where the data stays and how long you keep it.

Agreements. Business email usually contains personal data, such as names and phone numbers. Then the GDPR applies. Does another party process that data for you, such as the provider of the AI model? Then you conclude a data processing agreement with that party. Work with business accounts, and check whether the provider uses your data to train its model. Also read is ChatGPT safe for businesses?

Retention. According to Ondernemersplein, correspondence, such as letters and emails, is part of your administration. As a rule you keep that administration for 7 years. At the same time, you may not keep personal data longer than necessary. For automatic processing, that means:

• The software leaves the original mail in place. It does not delete anything from your mailbox.
• The draft in your system refers to the mail, so that you can check later where the data came from.
• Copies the software makes itself, such as the extracted text, you keep no longer than necessary for the check.
• Ask your accountant exactly which mail falls under the retention obligation.

This is an explanation in plain language, not legal advice.

When does automatic processing not make sense?

It does not make sense if little of the same mail comes in, or if every mail is different. A few situations:

• You get a few of these emails a week. Then retyping is often fine, and a good mail rule is enough.
• Every mail is bespoke. A complex project with a lot of discussion calls for a person, not a fixed route.
• You have no system to put it in. Are your customers and orders still in separate lists? Then start there. Read your Excel lists have got out of hand.
• The data already comes in neatly. Orders from your online shop or from large customers via EDI do not need to go through email. You connect those directly.
• Your software package already does it. Several accounting packages read in purchase invoices that you send to a fixed email address. Read reading in invoices automatically.

How do you start processing email automatically?

Start small, with one kind of mail. A good order is:

1. Look at your inbox for two weeks. Which kinds of mail come in, how often, and what is then retyped, and where to?
2. Choose one kind of mail. A kind that comes in often and is always similar, such as orders from regular customers.
3. Give that mail its own route. A separate address or a mail rule, so that only that mail reaches the software.
4. Write down what needs to be extracted. Which data, to which system, and when is something an exception?
5. Let AI only create drafts at first. An employee checks every draft and notes what goes wrong.
6. Adjust and then expand. Adapt the rules to what you see. Then choose the next kind of mail.

Not sure where to start yet? Read you don't know where to start with AI.

What does it cost, and how do I help you?

What it costs depends on the number of kinds of mail and the systems involved. The costs consist of building and integrating, hosting and maintenance per month, and the costs of the AI model based on usage.

With me, you start with a business analysis. I look at which mail comes in, what is retyped and what your mail program and software packages can already do. If it can be done with a mail rule or your existing package, my advice will say so; I receive no commission from suppliers. If more is needed, you first receive a proposal with a fixed price, so that you know in advance where you stand. After that I build it myself, in the systems you already use. Also read how your price is determined and more about integrations and automation. Or book a free first conversation.

Frequently asked questions

Sources: Microsoft Support: Manage email messages by using rules in Outlook, Gmail Help: Create rules to filter your emails, Microsoft Support: Frequently asked questions about Copilot in Outlook, Gmail Help: Learn more about Gemini features in Gmail, Microsoft Learn: Outlook mail API overview, Google: Gmail API overview, Microsoft Learn: Role Based Access Control for Applications in Exchange Online, NCSC (United Kingdom): Prompt injection is not SQL injection (it may be worse), Ondernemersplein: Protecting personal data (GDPR), Ondernemersplein: Step-by-step plan for keeping your administration, Ondernemersplein: Keeping business records and the retention obligation, Ondernemersplein: Privacy legislation GDPR, checklist for business owners, Autoriteit Persoonsgegevens: Guidance on generative AI and the GDPR, NJB: AP helps developers and organisations with new GDPR guidelines for generative AI (14 July 2026) and Autoriteit Persoonsgegevens: Data processing agreement. Consulted on 22 October 2026.