When it reads your messages

Of all embedded AI, the features that read your correspondence deserve the most attention. They operate on everything, continuously, and what they touch is frequently sensitive and rarely only yours.

Summaries of threads

The “here is what this thread says” box at the top of a long email chain. Genuinely useful — and it drops exactly the things E4·L3 warned about.

Conditions. “We can do Thursday if the figures arrive by Tuesday” becomes “They agreed to Thursday.”

Who said what. In a thread with several people, attribution is where these most often blur — and in a disagreement it is the only thing that matters.

The actual ask. Buried in paragraph four of a long message, which is exactly where summarisation deprioritises it.

Tone. Somebody being carefully polite about being extremely unhappy reads as neutral.

Use the summary as a map, as E4·L3 taught — to decide which messages to read properly. Never as the basis for a reply, and never for anything with a commitment in it.

Suggested replies

The one-tap responses. Convenient, and they carry a specific risk that is about you rather than the technology.

They are optimised for plausibility and speed. That combination encourages replying without reading properly — you glance, the suggestion looks right, you tap.

The failure is not that the reply is wrong English. It is that you have agreed to something. “Sounds good!” on a message whose second paragraph committed you to a date, a cost, or a position.

Rule: never use a suggested reply on a message you have not read to the end. That sounds obvious and it is precisely the behaviour the feature is designed to make unnecessary.

For anything with a decision, a number or a person’s feelings in it, write your own — E4·L2’s rule about bad news, apologies and disagreement applies at least as strongly here.

Priority sorting

Deciding which messages you see first, and which land in a low-priority pile.

The failure mode is invisible by construction: you cannot notice the message you were not shown. Unlike a bad summary, there is no output to inspect.

Two habits worth keeping: look through the deprioritised folder occasionally, and be aware that anything unusual — an unfamiliar sender, an atypical format, a first approach from someone who matters — is exactly what a sorted-by-typicality system handles worst.

The privacy that is not yours

The part most people never consider, and the most important in this lesson.

Your correspondents wrote to you. They did not choose your email provider, your assistant, or the summarisation feature you enabled. They have no way to know their message is being processed and no way to object.

That matters most when:

Someone wrote something confidential on the understanding it was going to a person.
The message contains third-party personal data — a colleague’s health, a client’s details, someone’s address.
You are in a role where correspondence is privileged or protected — legal, medical, journalistic, HR, safeguarding, pastoral.

For those roles specifically, an always-on summariser reading everything is a professional question rather than a preference, and your organisation’s rules almost certainly have something to say about it.

E4·L8’s paste test framed this as a decision about what you paste. Embedded features remove the pasting step — the content goes automatically — which makes it a settings decision made once, on behalf of everyone who writes to you.

Try it now (5 minutes)

Find a thread summary in your email. Read the whole thread, then re-read the summary.

Note what it dropped — most people find a condition or an attribution. Then decide, deliberately, whether you want that feature on given who writes to you.

Check your understanding

1. Email thread summaries most often lose: The subject line Conditions, attribution and the actual ask The names of participants

2. The real risk of suggested replies is: The English is often wrong They encourage replying without reading, so you agree to things They are too formal

3. Whose privacy does an always-on email summariser involve? Only yours, since it is your account Also your correspondents', who never chose it and cannot object Nobody's, if the provider does not train on it

Recap

Treat thread summaries as maps, never as the basis for a reply, because conditions, attribution and the ask are what they drop. Never send a suggested reply on a message you have not read to the end. Remember that priority sorting fails invisibly, since you cannot notice what you were not shown. And recognise that these features process other people’s messages — a decision made once, on behalf of everyone who writes to you.