AI email assistant: what it can and cannot do for your inbox in 2026

Every mail client now has an AI button. The interesting question is not whether it can write an email, it obviously can, but which parts of inbox work it should own, and which parts you should never hand over. Here is a clear map, what to check before you give one your mailbox, and what to demand from any assistant you keep.

What can an AI email assistant do well?

• First drafts. Given the thread, a good assistant writes a reply in your voice that needs editing, not rewriting. The saved time is the blank page.
• Triage. Newsletters, receipts and notifications are easy to classify; threads that need a human reply are easy to surface.
• Follow-ups on a schedule. The mechanical persistence that humans are worst at, with stopping rules that machines are best at.
• Summaries. A forty-message thread in five lines, when you have been away.

Those four cover most of the hours. Knowledge workers spend roughly a quarter of the week on email, and the time goes to drafting, deciding and chasing rather than reading, which is why an assistant that only summarises saves less than one that drafts. The research on where the time goes is in how often you should check email.

What does it do badly?

• Judgement about whether to reply at all. The most valuable inbox decision is the one to ignore something, and the assistant cannot know your priorities.
• Tone under pressure. Apologies, bad news and negotiation need a human hand on the draft; the assistant's default is smoother than the situation deserves.
• Relationships. An assistant can maintain a thread; it cannot know that this person hates being chased.

The pattern is the same one that decides whether AI can replace a personal trainer: the machine is excellent at the parts with a clear rule and a fast feedback loop, and poor at the parts where the rule is "it depends on the person". Email has both, in roughly equal measure.

PeakLook does this from your own data. Drafts the reply. Keeps the follow-up going. Join the free beta →

Will an AI email assistant send email without me?

It should not, with one deliberate exception. Drafts stay drafts until you send them. The exception is a follow-up sequence you explicitly scheduled: those steps go out on the schedule you set, and only until the person replies. That is automation you asked for, with a stopping rule, not an assistant deciding on its own. The design of that stopping rule matters more than the copy, and the sequence version is spelled out in email sequence examples that work.

Any product that sends unscheduled replies on your behalf has crossed from assistant to impersonation. The right model is an approval gate: the machine proposes, you approve, and anything that fires automatically fires because you set it to. The general case is argued in why approval gates are the safety feature.

Is it safe to give an AI assistant access to your inbox?

Read the permissions. For Microsoft 365, an assistant asks for specific Microsoft Graph scopes such as Mail.Read and Mail.Send, and you can see exactly which at the consent screen. The questions to ask of any assistant:

1. Which scopes does it request? Read and send are reasonable for a drafter. Full mailbox delegation, contacts and calendar write access need a reason.
2. Where is the mail processed? On the vendor's servers, on your own infrastructure, or sent per request to a model provider of your choosing with your key.
3. What is stored? The least possible, stated plainly. Your model key should travel with each request and never sit in their database.
4. Can you tell what the machine wrote? A visible line between machine text and your text is a safety feature, not a cosmetic one.
5. Can it send twice? A scheduled sender needs a lock; two devices or a retry should never produce a duplicate.

Which mailboxes do AI email assistants support?

Fewer than the marketing suggests. Most dedicated assistants started on Gmail, because Google's API is the easier one to build on; Outlook and Microsoft 365 support came later and in some products is still partial. Built-in assistants only work in their own suite: Gemini in Gmail, Copilot in Outlook. The practical consequence is that the first filter on any shortlist is the mailbox, not the feature list.

How to use one without sounding like a machine

The giveaway of a machine-written reply is not the grammar; it is the evenness. Every sentence the same length, every paragraph wrapped in a courtesy, nothing left out. Three habits fix it.

• Give it your own sent mail as the voice sample, not a style description. "Friendly but professional" produces the house style of every assistant; forty of your real replies produce yours.
• Edit by deleting. Cut the opening pleasantry and the closing summary. What remains is usually the reply you would have written.
• Keep the awkward sentence. When the draft smooths over a disagreement or a delay, put the plain version back. The assistant's instinct is to soften; yours should be to say the thing.

The test is whether the recipient could tell. If a reply reads like you on a good day, it is yours; if it reads like a help centre, send it back to the draft.

What to demand from one

Your voice, not a house style. Your keys, sent per request and never stored, if it uses a model provider on your behalf. A visible line between what the machine wrote and what you wrote. Send controls that cannot fire twice. And an off switch per contact, because the person who hates being chased exists in every address book. A product that meets all five is rare; the comparison of AI email assistants says which ones meet which.

How PeakLook draws the line

PeakLook drafts, triages, sequences and summarises; you read, decide and send. Anything the AI wrote or will send is gold in the interface so you always know who wrote what. Your Claude, OpenAI or Gemini key travels with each request and is never stored, and a shared send lock with two ledgers means a follow-up cannot go out twice. It works with Microsoft 365 and Outlook accounts through Microsoft Graph today, as an Outlook add-in and as standalone phone, watch and desktop apps; Gmail is on the roadmap.

Hen Eytan · Founder, UPeak

Builds the UPeak apps and writes up the research behind them.