Thinking Thinking

Answers

What does an AI email assistant actually do?

It sorts incoming mail, drafts replies and summarises threads, and it is bad at knowing what matters and at acting without a human.

By Thinking Desk 3 min read

Key takeaways

  • Sorting incoming mail into categories is the most reliable capability and saves the most time, because most inboxes are mostly mail that needs no action.
  • Drafting works for short routine replies and across languages, but anything with a number, a date or a commitment has to be checked, not just read.
  • Retrieval and summarising have improved most, and Shortwave’s search over a full Gmail archive is the strongest example of it.
  • The models cannot know what matters to you, because importance depends on what you are trying to achieve, which the mail does not contain.
  • Autonomous features fail on exactly the cases that are expensive, so enable them for one narrow category and review for a month before widening.
On this page
  1. Short answer
  2. The three things they really do
  3. The two things they do badly
  4. What this means for a buying decision

Short answer

An AI email assistant does three things reliably: it sorts incoming mail into categories, it writes a draft reply you then edit, and it summarises or searches across messages you would otherwise have to open. It does two things badly: it cannot tell what actually matters to you, and it cannot be trusted to act on its own without a human reading the result first.

The three things they really do

1. Sorting. Every tool in this class classifies incoming mail into buckets: needs a reply, newsletter, receipt, notification, probably spam. This works, and it is the part that saves the most time, because most inboxes are 80 percent mail that needs no action. Shortwave, Fyxer, Spark and BoxBee all do it, with different names for the same mechanism.

2. Drafting. Given a thread, the model produces a reply in roughly your tone and the right language. For short, routine, factual answers it is genuinely faster than typing, especially across languages. Fyxer goes furthest here and writes drafts before you ask for them. The output still needs reading, and for anything with a number, a date or a commitment in it, reading means checking.

3. Retrieval and summarising. A long thread becomes three sentences, and a question about something from eighteen months ago returns an answer instead of a list of search hits. This is the capability that has improved most, and Shortwave's search over a full Gmail archive is the best example of it on the market.

Weekly, Thursday morning, Central European Time.

The Thinking briefing

One email a week: what changed in European business software, which prices moved, and which comparison we published. No sales mail.

We store your address to send the briefing. One click unsubscribes you, and we never pass it on.

The two things they do badly

1. Knowing what is important. The model infers priority from surface features: who sent it, how urgent the wording sounds, whether you have replied to that sender before. It has no access to the thing that actually determines importance, which is what you are currently trying to achieve. So a quiet one-line mail from the person who decides your contract renewal gets filed as low priority, and a loud automated alert gets flagged. The failure mode that costs you money is a false negative, and no vendor in this category publishes a false negative rate.

2. Acting alone. Autonomous agents, auto-replies and calendar automation all work until the first case that does not match the pattern, and in email the exceptions are the expensive part: a cancellation, a price query, a complaint, a legal notice. An agent that archives or answers these without a human is not saving time, it is deferring a problem. Treat every autonomous feature as something to switch on for one narrow category, for example newsletters or delivery notifications, and review for a month before widening it.

What this means for a buying decision

If the three things above describe your problem, the category works and the choice comes down to price and where your mail is processed, which is what our ranking of AI email clients measures. If your problem is that you keep missing the one mail that mattered, no AI in this list fixes that, and a rule-based or permission-based inbox may serve you better.

One practical point that is easy to miss. Every capability above requires the tool to read the message body, which means the body goes somewhere. Before you enable drafting or an agent, check who that somewhere is and under which law it operates, using our GDPR checklist for AI email. The sorting feature is often the one with the best time-saved to risk ratio, and it is also the one you can keep while turning the rest off.

Questions and answers

Can an AI email assistant answer mail for me?
It can draft, and for short routine replies the draft is often good enough to send after reading. Letting it send without review is a different thing, and it fails on the cases that matter: cancellations, price queries, complaints and legal notices. Use autonomous sending for one narrow category at a time.
Will an AI assistant stop me missing important email?
Not reliably. Priority is inferred from who sent the mail, how urgent the wording is and your past replies, none of which tell the model what you are currently trying to achieve. A quiet one-line mail from an important sender is the classic miss, and no vendor publishes a false negative rate.
Does an AI email assistant read my messages?
Yes, every capability requires it. Sorting, drafting and summarising all need the message body, so the body is processed somewhere. Which company processes it, in which country and under which contract varies enormously between vendors, and that is the thing to check before you enable anything.

Sources

Every price and fact above was checked against these pages on the date shown.

  1. [1] Shortwave security and privacy accessed 2026-11-11
  2. [2] Fyxer pricing and features accessed 2026-11-11
  3. [3] Spark privacy policy, AI processing accessed 2026-11-11
  4. [4] BoxBee features accessed 2026-11-11
  5. [5] Missive AI documentation accessed 2026-11-11
Topics emailaiinbox

Thinking is published by TTA Technologies in Luxembourg, which also builds LUXPOS, Salonify, BoxBee and DepotGest. Read the full disclosure policy