AI
Can AI sort our inbox so nobody spends all day reading emails?
By Paul Meakin8 min read

Why does reading the inbox take someone all day?
Because reading is rarely the job. The job is working out what each message is, who should have it and whether it can wait. A shared inbox lands a supplier invoice, a customer complaint, a CV, three automated notifications and a question about an order all in the same list, and one person has to look at each to find out which is which.
Often that person is not making a difficult call. They are recognising a type of message and moving it. In many inboxes the difficult calls are buried under everything else, and a count of your own mail will show how many there are.
That split is what makes this worth automating. Sorting is repetitive and frequent, and it follows a pattern. Deciding what to do about a complaint does not.
What should plain rules do before AI touches anything?
The boring part first. Gmail's filters let you send email to a label, or archive, delete, star or automatically forward it, based on search criteria you set. Outlook has its own version. If a message can be identified from its sender or its subject line, a rule will handle it every time, for nothing, and you can see exactly why it did what it did.
Automated notifications, supplier statements and system alerts can often be caught this way. Clear them first and the pile may look quite different. Write down what is left, because that is the actual problem.
What remains is the mail where the wording varies and the meaning matters: a customer asking the same thing in a dozen different ways, or a supplier query with no useful subject line. Rules are poor at that. Reading it is what AI is for.
Where does AI reading actually work?
It works where the output is a label, a summary or a draft that someone can check at a glance. It works less well where the output is an action nobody looks at.
| Task | Who should do it | Why |
|---|---|---|
| Known senders and fixed subject lines | A plain rule | Predictable, free and easy to audit |
| Working out what a free text email is about | AI, with spot checks | The wording varies, and a wrong label can be caught and corrected, as long as someone is checking |
| Suggesting an order number or date for a record | AI suggests, and a validation step or a person writes it | Quick, but it needs a check against your own data |
| Summarising a long thread for the person who picks it up | AI | They still read the original before they act |
| Drafting a reply to a routine question | AI drafts, a person sends | The reply goes out under your name |
| Complaints, refunds, job applications, anything legal | A person, helped by the summary | Judgement, and possibly the law |
| Changing bank details or paying anything | A person, through your normal checks | An email should never be enough |
We covered what happens to attachments and paperwork once they arrive in how to keep track of documents customers send by email. This is the step before that.
What has to stay with a person?
Decisions that matter to someone. Filing a supplier statement is unlikely to be a significant decision. Rejecting a job applicant, or sorting applications into a pile nobody opens, could be a different matter.
UK data protection law now draws that line in Article 22A of the UK GDPR. It says a decision is based solely on automated processing if there is no meaningful human involvement in the taking of the decision, and it treats a decision as significant if it produces a legal effect for the person, or a similarly significant effect. The government's factsheet on the Data (Use and Access) Act explains that the reforms allow solely automated decisions in wider circumstances, as long as organisations implement measures to provide the relevant safeguards. It lists safeguards that include telling people about significant decisions made about them, letting them make representations and challenge those decisions, and letting them obtain human intervention.
So the law has loosened, but it has not gone away. And the word to watch is meaningful. A person clicking approve on everything the system suggests, without reading it, is an uncomfortable thing to rely on. Whether a process crosses the line is a question for your data protection adviser, and worth asking early.
There is a plainer point too. Email is full of personal data. If the text goes to an outside AI provider, that provider is likely to be acting for you, and Article 28 requires processing by a processor to be governed by a contract or other legal act. Read the terms of the tool before you connect a live inbox to it, not after.
Where do people get this wrong?
The expensive mistake is giving the AI too much power. Every email from a customer or supplier is text written outside your business, and the AI reads all of it the same way. The National Cyber Security Centre's piece on prompt injection puts it directly: under the hood of a large language model, no distinction is made between data and instructions. So an email can contain a line telling the system to do something, and the AI may try.
The same piece gets more specific. It argues that the risk cannot be fully removed and has to be reduced, and it says that if a system's security cannot tolerate the remaining risk, it may not be a good use case for this technology. Its email example is the one to remember: do not let a model processing emails from random external people have access to privileged tools.
In practice, that gives you a simple design rule. The part that reads the inbox can add a label, write a summary or prepare a draft. It cannot send, delete, forward outside the business, pay anyone or change a record that matters. A person, or a separate step with its own checks, does those. That limits the damage. It does not make the label trustworthy, because a crafted email can still get itself labelled wrongly, which is why you keep checking.
The second mistake is never measuring it. A model that sorts confidently is not necessarily sorting correctly, and you cannot take its confidence as proof. Nobody good judges a new starter on the interview. They check the work during probation, and this deserves the same.
The third is automating a mess. If nobody can say who owns supplier queries, labelling them faster only gets them to nobody sooner. Our post on where to start with AI makes the same point from further back.
How would you build this without betting the business on it?
In stages, with the system watching before it does anything.
Start by counting. Sort a normal week of the inbox by hand into types. You want to know what arrives, how much of it there is and who ends up dealing with it. That list becomes your labels.
Then set the plain rules and let them run for a while. You may be able to stop here, and that is fine.
Next, connect the AI so it only suggests. It labels each remaining email while the person carries on as normal, and you compare the two. Look hard at where they disagree, because those messages show you where your categories are unclear, as well as where the model is wrong.
When it agrees often enough for your comfort, let it label for real, with set categories going to a person, and anything that fits no category going there too. Keep a log of each email, the label it gave and any correction a person made, and sample it regularly. Drafts come after that, and only with a person pressing send.
We build this as AI powered solutions and, for the rules and routing around it, process automation.
When is this not the answer?
When there is not enough mail to justify it. If one person clears the inbox in a few minutes each morning, filters and a tidy label structure will do, and AI adds a cost and a risk for very little.
Nearly every email may need real judgement. A small specialist firm may find that reading each message is the work, and a good summary is the most it can usefully use.
The mail may be sensitive. Health details, legal matters or anything about children raise the stakes, and you should take data protection advice before sending that text to any outside tool.
Or email may be the wrong channel in the first place. If customers keep emailing for the same three things, a form that captures the details properly may remove the sorting altogether.
Where do you start?
Count a week of the inbox by type, then mark which types a rule could catch, which need reading and which need a decision.
If most of the pile turns out to be reading rather than deciding, tell us what's stuck and we'll map the quickest fix. We will tell you straight if filters will do. Time to talk yet?
Common questions
Can we let AI answer routine customer emails on its own?
We would not start there. Have it draft and let a person send, then look at how often the drafts go out unchanged. Even then, keep its power to act narrow, because every incoming email is text from outside your business.
Does using AI on our inbox mean we need to tell customers?
It may affect what your privacy notice should say, particularly if an outside provider receives the text. That is a question for your data protection adviser, with the tool's terms in front of them.
Will Gmail or Outlook do this without anything extra?
Their filters and rules handle anything you can identify from fixed details such as the sender or subject. Reading free text and working out what it means is what needs something more.
How do we know the sorting is right?
Sample it. Have a person regularly check a batch of labels against the original emails, and record what was wrong. If nobody is checking, you do not know.
Sources
Paul Meakin, Founder
Twenty years of fixing businesses from the inside, eighteen of them in recruitment from consultant to national operations, before building the automation, web applications and compliance systems Staxxd runs today.
More about Paul