Why filters stopped being enough
Email rules are deterministic. If the sender is this, apply that label. They work perfectly for the cases you anticipated and not at all for anything else — which, in most inboxes, is the majority of the volume.
The failure is easy to see. A client emails asking to move a deadline. There is no rule for that, because you did not know it was coming. It lands in the same undifferentiated pile as a newsletter and a receipt, and it is on you to spot it. Meanwhile you accumulate dozens of filters that each handle one narrow case and quietly rot as senders change.
An AI model reads the message instead of matching a pattern. It can tell that a deadline request is a request, that a thread has become urgent, and that a message from an unknown address is a genuine sales enquiry rather than spam. That is the entire difference, and it is why the approach handles the messy remainder that rules never reached.
What triage actually decides
A workable triage setup sorts incoming mail into a small number of buckets. More than about five and the classification quality drops, because the boundaries blur.
- Needs a reply from me. A person is waiting on a human answer. This is the only bucket that should reliably reach your attention.
- Urgent. Time-sensitive enough to interrupt for — an outage, a same-day deadline, a client escalation.
- Informational. Worth knowing, no action required. Notifications, confirmations, CC threads.
- Noise. Newsletters, marketing, automated reports. Archive and move on.
- Needs action, not a reply. An invoice to pay, a form to complete. Often the bucket people forget, and the one that causes the most trouble when it is missing.
The model can also do things filters simply cannot: summarise a forty-message thread into three lines, extract the deadline buried in paragraph four, or flag that a polite message has an unhappy tone.
Setting it up
1 Install the desktop app
Install the free OpenClaw Desktop App on macOS or Windows. It runs an AI agent locally on your machine, which is what makes the private version of this setup possible.
2 Choose where the model runs — this is the important decision
Email is among the most sensitive data you own. Two options, with a real difference between them:
- Local model (free, private). Install Ollama and run the model on your own hardware. No email content leaves the machine, and there is no cost. Classification is a well-suited task for smaller models — it is pattern recognition on short text, not deep reasoning — so the capability gap matters less here than almost anywhere else.
- Cloud model (paid, sharper). Claude, GPT or Gemini via an API key. Better at nuance and at drafting replies that sound like you. Your email content is transmitted to that provider for processing.
If your inbox carries client work, legal, medical or financial correspondence, choose local. It is the one configuration where the privacy question is answered structurally rather than by a policy document.
3 Connect your email
Connect your mail account to the agent. The pre-built Gmail inbox-zero triage skill covers the common setup; other skills in the catalog handle labelling and reply drafting.
Grant the narrowest permissions that do the job. Read and label access is enough to start. Do not grant delete or send until you have watched the classifications for a while.
4 Define your categories in your own words
Describe what each bucket means for you. Generic definitions produce generic results. Compare:
Urgent: anything time-sensitivewith:
Urgent: production outages, anything from a client
mentioning a deadline in the next 48 hours, or any
message where someone is waiting on me to unblock them.
Not urgent: internal FYIs, anything I am only CC'd on.The second gets dramatically better results. Specifying what does not belong in a category is as valuable as specifying what does.
5 Run it on a schedule
Continuous processing is rarely what you want — it recreates the interruption problem you were trying to solve. Schedule triage a few times a day and have the agent send you a short digest of what needs attention, straight to WhatsApp or Telegram if you prefer to stay out of the inbox entirely.
Start in label-only mode
This is the part people skip and then regret.
For the first week, give the agent permission to label and nothing else. No archiving, no moving, no sending. Every message stays exactly where it was, wearing a label that records what the model thought.
At the end of the week you will have a week of visible decisions to audit. You will find that one of your category definitions was ambiguous, that a particular sender is consistently misread, and that one bucket you thought you needed is never used. Fix those before granting more authority.
Escalate slowly: label only → archive obvious noise → draft replies for review → send automatically for a narrow, well-tested category. Most people stop at drafting, and that is a perfectly good place to stop — the value is in not starting from a blank page, not in removing yourself from the loop.
Drafting replies
Classification saves attention. Drafting saves the most time, and the pattern that works is drafting into your drafts folder rather than sending.
Two things improve the output sharply. First, give the model examples of how you actually write — a handful of your own previous replies teaches tone far better than an adjective like "professional". Second, tell it what to do when it is unsure: a draft that says "I do not have enough context to answer this" is far more useful than a confident, wrong reply that you nearly sent.
Be realistic about the limits. A local model produces serviceable drafts for routine correspondence and struggles with anything requiring negotiation, nuance or relationship history. Anything sensitive should be written by you.
Free versus paid
Commercial email-AI products — the assistant features built into mail clients and standalone triage tools — typically charge somewhere between $10 and $30 per user per month. They are polished, require almost no setup, and process your email on their servers.
The self-hosted route costs nothing, requires 15 to 30 minutes of setup plus tuning, and keeps everything on your machine. It is also more flexible: because you are describing categories in plain language rather than choosing from a vendor's fixed list, you can build triage rules that no product would have shipped.
The deciding factor is usually not price. It is whether you are comfortable with a third party processing your inbox. For a personal account full of newsletters, probably fine. For a professional inbox under a confidentiality obligation, the local setup is the one that survives scrutiny.
Where it goes wrong
Too many categories
Twelve buckets sound thorough and classify badly, because the boundaries between them blur and the model guesses. Three to five is the range that works.
Vague definitions
"Important" means nothing to a model. Name the senders, the projects, the time windows. Specificity is the single biggest lever on quality.
Trusting it too early
Granting archive permission on day one is how an important message gets buried. Label-only for a week is cheap insurance.
Automating a broken process
If you receive 200 emails a day because you are CC'd on everything, triage treats a symptom. Unsubscribing and fixing the CC habit removes more volume than any amount of automation.
What to expect
Realistically: classification that is accurate most of the time and occasionally, confidently wrong. Drafts that are a decent starting point and rarely sendable as-is. A meaningful reduction in the time spent deciding what to look at, which is where most inbox time actually goes.
What it will not do is achieve inbox zero on your behalf. Messages that need a real decision still need you — triage just makes sure you see them first rather than finding them on Friday underneath four hundred newsletters.
Frequently asked questions
What is AI email triage?
AI email triage uses a language model to read the content of each incoming message and classify it — urgent, needs a reply, informational, or noise — then act on that classification by labelling, sorting, archiving or drafting a response. The difference from a traditional filter is that no rule has to be written in advance. A filter matching sender or subject only catches what you predicted; a model reading the message understands intent, so it correctly handles a request for a deadline extension whether it comes from a known contact or a stranger.
Is there a free AI email triage tool?
Yes. Running the OpenClaw Desktop App with a local model through Ollama gives you AI email triage at no cost — no subscription, no per-message fee, and no email content leaving your machine. Most commercial email-AI products charge a monthly per-seat fee, typically in the range of $10 to $30. The trade-off with the free route is setup time and a somewhat less capable model.
Is it safe to give an AI access to my email?
It depends entirely on where the model runs. With a cloud service, the content of your emails is sent to that provider's servers to be processed, which is a genuine consideration for legal, medical, financial or client correspondence. With a local model, the email never leaves your computer — there is no API call to make. If your inbox contains anything confidential, the local setup is the one that actually resolves the concern rather than relocating it.
Can AI reply to emails for me automatically?
It can, and for most people it should not — at least not at first. The reliable pattern is to have the AI draft a reply and leave it unsent for you to review. Drafting removes the hard part, which is starting from a blank page, while keeping a human decision before anything reaches another person. Fully automatic sending is reasonable only for narrow, high-volume, low-stakes categories where you have watched it behave correctly for weeks.
How is this different from Gmail filters and Priority Inbox?
Gmail filters are deterministic rules matching sender, subject or keywords, and they are excellent at exactly the cases you thought to define. AI triage reads and interprets, so it handles unanticipated messages, understands tone and urgency, and can summarise rather than just sort. Gmail's own Priority Inbox does apply machine learning, but you cannot direct it — you cannot tell it that anything mentioning a specific project matters this month. The two work well together: keep filters for the obvious deterministic cases and let AI handle everything that falls through.
How long does AI email triage take to set up?
Roughly 15 to 30 minutes for a basic setup: install the desktop app, choose a model, connect your email, and define your categories. Budget more time for tuning. The first week is worth running in label-only mode so you can see the decisions it makes without it moving or archiving anything, and adjust your category definitions before granting it more freedom.
What happens if the AI misclassifies something important?
This is the risk that matters, and the reason to stage the rollout. In label-only mode a mistake costs nothing — a wrong label on a message still sitting in your inbox. Only after watching it perform accurately should you allow archiving, and even then it is sensible to exclude specific senders from automatic handling entirely. Never let an unreviewed model archive mail from your accountant, your lawyer or your largest client.