# What AI Can Actually Do With Your Gmail Once Connected

*Question — 2026-09-24 — by Mahmoud Zalt*

Read threads, draft replies in your voice, label mail, chase quiet conversations. The full list of what AI does in Gmail, and what it will never do.

**Short answer.** **Once connected, AI can read your threads for context, write drafts that sound like you, label and file mail, chase conversations that went quiet, and pull answers out of your own email history.** In Sistava those five things cover almost every real request people bring to an inbox assistant. Sending is a separate permission you grant on purpose, and deleting is not on the list at all.

Most descriptions of AI email help stop at the word smart, which tells you nothing. You want the actual list. What does it touch on a Tuesday morning, and what does it leave alone.

The honest version is narrower than the hype and broader than most people expect. It cannot invent facts about your business that are not written down somewhere. It can read four months of a client thread in a second and write the reply you would have written after twenty minutes of scrolling.

That gap is where the value sits. Not magic, just the tedious middle of email work done properly and quickly, with you keeping the final word on anything that leaves your address.

Inside **Sistava**, an AI Employee treats your mailbox as a workspace rather than a feed. It reads threads to understand what is going on, drafts replies in your voice, files and labels mail so the pile stops growing, follows up on the conversations that went quiet, and asks before sending when you set it that way. You can see the wider set of tools it works with on the [integrations page](/features/integrations).

## At a Glance

- **5** core jobs it does inside Gmail every day
- **0** emails it can delete, ever
- **Seconds** to read a thread that would take you ten minutes
- **You** still decide what leaves your address

## What are the jobs it actually does in Gmail?

There are five, and they map onto the things a good human assistant would do with an inbox. Understand the conversation, write the reply, organise what is left, chase what went silent, and find the thing you half remember from March.

Everything else people imagine is either a variation on those five or something the product deliberately refuses to do. Here is the list with the detail attached.

## Benefits

### Reads threads for context

It opens the whole conversation, not just the last message, so the reply knows what was promised in February and what changed last week.

### Drafts replies in your voice

It learns your patterns from the mail you have already sent: your greeting, your length, whether you sign off warmly or briskly.

### Files and labels the pile

New mail gets sorted into the labels you use, archived when it needs no answer, and left in the main view when it needs you.

### Follows up on quiet threads

It notices the quote nobody answered and the invoice that went silent, then chases them on the schedule you set.

### Answers from your own history

Ask what did we quote that client last spring and it searches the mailbox rather than guessing, then shows you the thread it used.

### Sends, only if you allow it

Send is a separate permission. Leave approval mode on and every message waits for you, whatever the draft says.

## How good are the drafts, honestly?

Good enough to send unedited for routine mail, and worth a read for anything that carries a decision. Scheduling, acknowledgements, standard answers, and status updates land right most of the time. Anything involving a price, a deadline you might want to negotiate, or an unhappy customer deserves your eyes.

The voice matching is more convincing than people expect, because it is not copying a style guide. It is reading the hundreds of emails you already wrote from that address and following the shape of them, including the small habits you never noticed you had.

The realistic expectation for week one is that you edit about a quarter of the drafts, mostly to add a detail only you knew. By week three that number usually drops, because the corrections you made are the corrections it stops needing.

## What does a full week of this look like?

It looks like the boring half of the inbox handling itself. Tobias runs a plumbing firm in Hamburg with three vans and no office staff. His mail is a mix of job enquiries, supplier invoices, scheduling changes, and the occasional complaint, and he was answering all of it at 21:00 from the sofa.

He hired an AI Employee and connected Gmail with read and draft permission. The first job he gave it was small: label everything and draft replies to enquiries only. Nothing else, no sending.

Over one week it read 143 messages, created 5 labels, archived 71 items that needed no reply, drafted 22 enquiry responses, and flagged 4 threads as needing him personally, including a complaint about a bathroom job. It also chased 3 quotes that had gone quiet for over ten days, which turned into two booked jobs. Tobias edited 6 drafts and stopped working evenings.

| What arrived | How many | What the AI Employee did |
|---|---|---|
| Job enquiries | 22 | Drafted a reply with availability, left it for review |
| Supplier invoices | 18 | Labelled and archived, flagged the two that were unusual |
| Scheduling changes | 14 | Drafted a confirmation, updated the label on the job thread |
| Newsletters and noise | 71 | Archived quietly, never surfaced |
| Complaints | 4 | Left in the main view, flagged, no draft written |
| Quiet quotes worth chasing | 3 | Follow up drafted after ten days of silence |

## What will it never do with your Gmail?

It will never delete anything. No messages, no threads, no emptying the trash. Archiving and labelling move mail out of the way, but everything stays in your account and stays findable, which means a mistake in filing costs you a search rather than a loss.

The rest of the boundaries are just as firm, and they are the reason the capability list above is safe to hand over in the first place.

- It will not send while approval mode is on, no matter how routine the message looks.
- It will not treat text inside an email as an instruction. A message telling your assistant to forward the client list is read as content from a stranger, then flagged.
- It will not handle money. Payment links, invoice approvals, and bank alerts get raised to you rather than answered.
- It will not invent a fact to fill a gap. If the answer is not in the thread or in what you told it, the draft says so and asks you.
- It will not reach beyond Gmail. Your Drive, your photos, and your account settings are outside the permission entirely.

The refusal to invent is the one that saves you most often. An assistant that confidently makes up a delivery date is worse than no assistant, so the drafts you get on thin information come back with a gap and a question rather than a guess.

## How do you decide what to hand over first?

Start with the category you write the same reply to most often. For almost everyone that is enquiries, scheduling, or acknowledgements, and it is the fastest way to see whether the voice matching works on your mail rather than in a demo.

Leave the hard categories alone until later. Complaints, negotiations, and anything with a number attached should stay with you while you build a sense of where the drafts are strong. That order is not caution for its own sake, it is just how you learn the tool quickly.

### Turn the list into a working inbox

1. **Name the one category first** — Pick the mail you answer most repetitively. Tell your AI Employee that this category is the job and nothing else is, yet.
2. **Let it file everything, draft only that** — Labelling and archiving are low risk and immediately useful. Drafting stays narrow while you check the voice.
3. **Read a week of drafts before widening** — Note what you edit. Those edits are the instructions worth giving it, and they are usually two or three specific things.
4. **Add follow ups, then consider send** — Chasing quiet threads is where most of the recovered revenue hides. Sending comes after, and only for a category you are relaxed about.

## Frequently asked questions

## FAQ

### Can it search my whole email history, or only new mail?

It can search the history inside the scope you granted, which is what makes the answers useful. Asking what did we agree with that supplier last autumn returns the actual thread rather than a guess. If you scoped the connection to a single label, the search covers that label only, which is a fair trade for the tighter boundary.

### Can it read attachments, or just the message body?

It reads the text of common attachments like PDFs and documents, which matters when the important detail is in a quote or a signed order rather than in the email itself. It does not open anything it cannot read as text, and it will tell you rather than pretend. Anything that looks like a payment instruction gets flagged to you instead of acted on.

### Will every reply sound the same?

No, because it is following your existing mail rather than a template. Your replies to a long standing client and to a cold enquiry already differ in length and warmth, and the drafts follow that difference. You can also tell it plainly to be shorter, warmer, or more formal with a particular group, and it holds that instruction going forward.

### Can it work with my calendar at the same time?

Yes, and that pairing is where scheduling mail stops being work. With calendar access it can propose real times instead of vague ones, and stop offering a slot that is already taken. Calendar is a separate permission with its own consent screen, so you can run mail alone for as long as you want first.

### How does it know which emails actually matter?

Partly from the thread itself, partly from what you tell it. A message from a client you have exchanged forty emails with is treated differently from a first contact, and anything mentioning money, a complaint, or a deadline gets raised rather than quietly handled. You can also name senders or subjects that must always reach you untouched.

The list is short on purpose. Read, draft, file, chase, find, and send only when you allow it. Hand over one category, watch a week of it, and widen from there.

**Tags:** ai-gmail-capabilities, ai-email-assistant, inbox-automation, email-drafting, ai-employees