Taming the Flood: Using AI to build an email system that learns how you work

Email is the number one grump in most people's working lives. It is a flood, and it is unstoppable. Every solution that has been tried over the years, from inbox zero methodologies to priority inbox features, eventually collapses under the weight of volume. Gmail has rules, and they help, but they are very blunt tools. You can filter by sender or by keyword, but you cannot easily say “this person is important but only when they are writing about this topic” or “I subscribe to 70 newsletters and I would like them sorted by subject area, not left in a heap.”

That is the problem we set out to solve, and the result is a system that runs silently every fifteen minutes, classifies everything that comes in, and leaves each executive with only the emails they actually need to see.

The system works on four tiers.

• Tier 1 is action: emails from VIPs, partner organisations, universities, and anything containing signals that suggest a response is needed. These stay in your inbox.

• Tier 2 is reading: newsletters and industry content, automatically sorted into 12 categories, such as AI and Automation, Business, Economics, Entrepreneurship, Learning, Marketing, and more. These are archived out of your inbox but labelled and waiting for when you want to catch up.

• Tier 3 is archive: notifications, booking confirmations, calendar updates, service alerts. These are filed away immediately, and you never see them unless you specifically look for them.

• Tier 0 is uncertain: the sender is unknown, or the system lacks sufficient information to classify them. That’s just our system. You can set up anything you want.

It is the uncertain tier that makes the learning work. When the system encounters a sender it has not seen before, the email remains in your inbox with a clear label indicating that it does not know what to do with it. You look at it, decide whether it is a VIP, a newsletter, or a notification, and classify it once. We use Google Workspace, so for us, that means changing the tag from T0 to the appropriate alternative. From that point on, the system recognises that classification and applies it to every future email from that sender, retroactively updating any previous messages it had stored. You teach it once, and it learns permanently. It does not ask again, it does not nag. It just works for you, in a way that everybody understands. You change the folder (in Gmail, the tag) and the system does the rest.

The newsletter sorting is worth dwelling on because newsletters are a significant source of clutter. I have over 70 newsletter subscriptions, and they arrive at all hours of the day. Some are straightforward to classify by sender alone: a technology review always falls under Tools, and a macroeconomics digest always falls under Economics. But some writers cover multiple topics. An AI researcher might write about pedagogy one week and large language models the next. For those, the classifier reads the subject line and body, examines weighted keywords, and assigns the correct category for that particular edition. It gets this right consistently because the rules are calibrated against what we have actually observed over several months of use.

The entire process runs in the background via a scheduled task on the office server, every fifteen minutes, with no visible window, no pop-ups, and nothing the user ever sees. On an average day, 25 to 35 notifications are auto-archived per user run, newsletters are sorted into their categories, and the inbox is left with items that genuinely need attention. The estimated time saving (let alone the stress) is an hour per day per user.

There are 93 Gmail filter rules (which we also set up with AI by analysing our previous patterns) that handle the obvious and instantaneous cases: known newsletter addresses, system notifications, and no-reply senders. Those fire immediately with no delay. The Python classifier, which runs every fifteen minutes, catches everything else: the edge cases, the ambiguous senders, and the content-based newsletter sorting. The two layers work together so that nothing falls through the cracks and nothing is processed twice.

Each executive has their own configuration. One person might want to disable the uncertain tier entirely, so unclassified emails stay unlabelled. Another might want a different set of VIPs or different newsletter categories. One of our directors has an additional function: an AI summariser that produces a digest of his newsletters each morning, with enough context to decide which one or two are worth reading in full and which can be safely skipped. The configurations are centralised on the server and can be updated at any time, so changes made once apply everywhere.

One point worth making on data protection: the classification runs entirely on a local machine. No email content, no subject lines, no sender addresses are sent to any external AI service. The Gmail API connection is standard authorised access to the user’s own account, and the classification logic is rule-based Python running locally. When the system reads a newsletter’s body to determine its category, that processing happens in memory on the machine and is not transmitted anywhere. For a tool that touches every email you receive, that matters.

The business case for this was straightforward. We were all spending too much time and too much cortisol managing email, and not enough time on productive work. The outcome we wanted was an inbox that, by the time you sit down in the morning, had already been sorted into what needed your attention, what was there for your interest, and what could safely be ignored. We reviewed the plan, built it, and within a few weeks, it had become one of those things nobody notices because it just works. Which, for a tool like this, is exactly the point.

A technical note on what it took

The distinction that matters is between human time and computer time. My role was to define what I wanted: the tier structure, who counts as a VIP, what the newsletter categories should be, and how the learning mechanism should work. Claude Code handled the building: the classification logic, configuration files, Gmail integration, and scheduled tasks. I reviewed the output, identified the issues, and they were corrected. The five-step discipline from this series is in practice.

The system was working within a couple of sessions. Since then, we have refined it for edge cases and rolled it out to other executives, each with their own configuration. But the core was right quickly, because the brief was clear. The total investment of my time, across all of it, was a few hours spread over a few weeks. The AI did the rest.

About The Series

This article is part of a series on practical AI applications at ChangeSchool.

Article 1: Getting Real with AI

For the strategic view of leading AI in organisations, see Viren Lall’s companion series. - https://www.changeschool.org/news

About the Author

Neil Marshall is Development Director of ChangeSchool, an EFMD award-winning executive education delivery partner.

ChangeSchool is an SME without a dedicated coding resource. Every application described in this series was built using AI coding tools. They are all vibe coding projects, and all within the reach of a startup or microbusiness willing to spend some hours and give it a try. If you have any questions about what this was or how we did it, or any suggestions for better ways forward, reach out to him on neilm@changeschool.org

Technical: Our process is the same for every project. Once the business need and objective are defined, we research the approach using Google Gemini Deep Research and Perplexity Deep Research, then give that research to Claude Code to build it. None of the technical terms below were things we knew beforehand; they emerged from that research process. It takes minutes to get up to speed. This project was built in Python using Claude Code. The classifier uses the Gmail API (developers.google.com/workspace/gmail/api) with OAuth 2.0 authentication and rule-based classification, driven by JSON configuration files. No machine learning models are needed for the core classification; it runs entirely on local logic. The newsletter summariser uses Google’s Gemini 3 Flash model (a very cost-effective option) via the Google Genai Python library.