Triage, summarise, draft — and never send

Three jobs, three risk levels

"AI for email" is three separate things, and they need separating because their failure modes are nothing alike.

Classifying — deciding what each message is. Wrong answers are cheap for most categories and expensive for a few.

Summarising — telling you what a thread says. Wrong answers are quiet, because omission leaves no trace.

Drafting and sending — writing the reply. Wrong answers leave the building with your name on them.

Almost all the safe value is in the first two.

Label, never move

The single most important rule in inbox automation has nothing to do with AI, and applies with more force once AI is involved:

A rule that moves a message out of sight is a rule that can hide something.

Apply labels or flags, leave everything in the inbox, and let the labels drive your reading order. A misclassified message you can still see costs you three seconds. A misclassified message filed into a folder you check monthly can cost a customer, a deadline or a regulatory response.

The same applies to anything that marks messages as read, or that archives on your behalf. Classification is advice. Advice does not get to move your post.

Tune for the expensive mistake

Categories are not symmetric. Missing a promotional email costs nothing; missing a complaint that becomes a formal grievance costs a great deal.

So for the categories where being wrong is expensive, deliberately over-flag:

Label as NEEDS ATTENTION anything that could plausibly be a complaint, a legal notice, a regulatory communication, a safeguarding concern or a deadline. When uncertain, label it. Over-labelling is preferred.

You will read more messages than strictly necessary. That is the correct trade, and it is a decision you make once rather than a judgement the tool makes for you on a Tuesday.

The daily digest, not the constant assistant

The pattern that survives contact with real work is a single pass, once a day:

Below are the subject lines and first 200 words of everything that arrived since yesterday. Produce three lists. One: messages needing a decision from me, each with the decision being asked for in one line. Two: messages where somebody is waiting on me and a date has passed. Three: everything else, one line each. Do not draft replies. Quote the sender and subject exactly so I can find them.

Then you work from the digest with the originals one click away. This is better than continuous triage for a reason worth naming: it produces one artefact you read attentively, rather than forty small suggestions you approve while thinking about something else — which is the condition in which automation bias does its worst work.

Summarising a thread

The useful demand is not "summarise this":

From this thread: what was decided, who agreed to do what by when, what question is still open, and what was asked of me and never answered. If any of these is not present, say so explicitly rather than omitting the heading.

The last clause is doing the real work. The characteristic failure of a summary is the missing sentence, and a heading that must appear even when empty is the only cheap defence against it.

It never sends

An automatic out-of-office reply is bounded: it says the same thing to everybody and commits to nothing. A generative auto-reply is not bounded. It can agree to a deadline, accept terms, confirm a price, or apologise for something in a way that matters later.

In many jurisdictions an exchange of emails can form or vary a contract, and in most organisations a message from your address is treated as your statement. The rule is therefore absolute and simple: drafting is automatic, sending is yours. Read it cold, in full, and press the button yourself.

Somebody else's personal data

Your inbox is full of other people's information — clients, patients, applicants, colleagues discussing colleagues. Connecting a mail account to an outside service means sending that material to a third party, and the people it concerns did not agree to it and cannot be asked practically.

That is a decision for your organisation, not for you individually, and the block on confidentiality later in this course covers how to make it. If you are reading this and thinking "I already connected mine", that is worth raising now rather than at an incident.

Do the boring rules first

A large share of what people want from AI triage is achieved by ordinary deterministic filters that have existed for twenty years and are free: sender-based rules, mailing-list headers, keyword flags in Thunderbird, Gmail or Outlook.

Set those up first. They are exact, they never invent, they cost nothing, and they handle the high-volume, low-judgement traffic that makes up most of an inbox. Then apply the model to what is left, which is the part that actually needed judgement — and which is now small enough to read.

For the local option, a small script over IMAP plus a model running in Ollama will classify a day's mail on a laptop, with nothing sent anywhere.