Gmail AI Agent: What Gemini in Gmail Still Cannot Do

A client sends the signed contract on a Tuesday. It arrives in Gmail as a PDF named scan_0042.pdf. You read it on your phone, think you will file it properly later, and then you do not file it later.

Three weeks on, someone asks where the signed copy went. You search Gmail for the client name, get a couple of hundred threads back, and spend ten minutes on a job that should have cost you nothing.

Gemini was sitting in that inbox the entire time. It could have summarised the thread for you. It could not put the file in the folder.

That is the real question behind "Gmail AI agent" in 2026. Whether AI is in your inbox is settled. It is. What is left over is the part worth paying for, and this post draws that line in one table.

What native Gemini covers, and what needs an agent

This is the table I would want if I were the one buying. The test in the last column is mechanical: count how many products the job touches, and ask whether anything has to happen while nobody is looking.

Google ships Gmail AI features on a rolling basis and changes which plan carries which feature, so treat the left column as the shape of what an in-thread assistant does rather than a frozen feature list. The right column has not moved in two years and is unlikely to.

Read the right-hand column and one word repeats: elsewhere. Drive, the CRM, the sheet, the calendar, the recipient of the digest. An assistant that lives in the compose window is very good at the message and structurally unable to act on any of that, because acting in another product is not what an in-thread assistant does, and it has no reason to wake up.

The label row is the interesting one, and it is the row I would expect to cause the most confusion. Deciding that a message belongs under Clients / Acme / Contracts is a reading task, and reading tasks are exactly what a language model is good at. Doing it to every message that lands overnight, consistently, for six months, is a scheduling and permissions task. Same intelligence, completely different plumbing.

Is it worth it if you already pay for Google Workspace?

Sometimes not, and I would rather say that plainly than sell you something twice.

I am not going to print a Workspace price in this post. Google moves these features between plans on a rolling schedule, and a figure written in October is routinely wrong by February. Read the current number and the current feature list off Google's own Workspace pages on the day you decide, not off a blog post, mine included.

What I can be precise about is the test. Write down the chore that annoys you most, how many minutes it takes, and how often it happens. Twenty minutes twice a week is a little over eight and a half hours a quarter. That number is what you are buying back, and it is the only one that should touch the budget conversation.

Then check three conditions. If all three hold, the extra spend is defensible. If any fails, stay with what you have.

• The job crosses a product boundary. It ends in Drive, a CRM, a sheet or a calendar rather than in a compose box. If everything it needs is already in the thread, you have the tool.
• It repeats on a rhythm. Daily, weekly, or every time a particular kind of message lands. One-off cleanups are an afternoon with a spreadsheet.
• Something must happen while you are not watching. The whole value of a chase-up is that it fires on day three whether or not you remembered.

On price, here is our side of it so you can do the arithmetic against whatever Google quotes you. Gravity runs your first agent free, with no card. Autopilot is $20 a month worldwide and ₹1,999 a month in India, with a lot of usage included. Minipilot is ₹399 a month in India with a smaller allowance. Buy more usage if you run out. If you want the wider market at this level, our comparison of the cheapest AI agent platforms covers who charges for what.

One structural note that matters more than any headline price. Suite AI is normally sold per user per month, so a thirty-person company pays thirty times over for the four people who actually wanted the automation. Check that ratio in your own org before you compare anything.

Five Gmail jobs worth handing to an agent

These are the five I would start with, ordered by how little has to go right for them to be worth the setup. Each one has a longer walkthrough behind it.

1. Filing attachments into the right client folder

This is the scan_0042.pdf problem from the top of the post, and it is the fastest win because the failure is so cheap to fix and so expensive to leave. An agent reads the message, works out which client and which matter it belongs to, renames the file into whatever convention you use, and writes it into the correct Drive folder. The intelligence is in the classification. The value is in the write.

Two rules I would set before the first run. Give it a fallback folder for anything it is unsure about, so ambiguity produces a small pile to review rather than a wrong filing. And have it leave the original in Gmail untouched, because a copy in the right place costs you storage and a move you did not intend costs you an afternoon.

The full setup, including the naming conventions that survive contact with real attachments, is in our guide to AI agent attachment filing for Gmail.

2. The morning triage pass

The agent runs before you wake up, reads everything that arrived overnight, and hands you a short digest: what needs an answer today, what can wait, what is noise. You open one message instead of sixty.

Keep the first version read-only and keep it short. A digest with four sender groups and the three most urgent items on top is enough, and it is far more likely to survive than an elaborate scoring system nobody trusts. Our inbox triage walkthrough has the prompt structure and the thirty-day review that tells you what to change.

3. Labels that hold up after six months

A label system is usually fine for a fortnight and then rots, because applying them depends on a human remembering at the exact moment a message arrives. Filters catch the easy cases and nothing else. Six months in you have a Clients label holding a fraction of client mail, and no way to tell which fraction, which is worse than no label at all, because you now trust a view that is lying to you.

An agent classifies by meaning rather than by keyword match, which is why it copes with the variety in real email where a filter cannot. Start it on the three labels you actually search, not the twenty you created optimistically, and read the first week's output before you widen the taxonomy. The mechanics are in the Gmail label automation guide.

4. Who chases the thread that went quiet?

Nobody, usually. That is the point. The quote you sent on the 4th got no reply, and by the 20th it is buried under four hundred newer messages and has quietly stopped being real.

An agent watching for outbound threads with no inbound reply after a set number of days is the closest thing to free money in this list. It costs almost nothing to run and it recovers work you had already done. The output should be a draft follow-up sitting in your drafts folder, addressed and ready, not a message that goes out on its own.

Two caps are worth setting on day one. Put a hard limit on how many chases it can queue per day, because a system that surfaces forty at once is the kind you stop opening. And exclude anyone who has replied in the last week on any thread, so you do not nudge someone you spoke to yesterday.

Set the threshold at three business days for sales conversations and seven for everything else. Those two numbers are reasonable defaults to start from, then tune them from your own reply patterns after a month.

5. Email ends, the record updates, nobody retypes

The customer replies with a new timeline, a changed scope or a decision. Somebody is supposed to reflect that in the CRM. Often nobody does, and the pipeline slowly becomes fiction.

This is the clearest case in the whole post for an agent over an assistant, because the input is in Gmail and the output is in a completely separate product with its own permission model. Gemini can read the reply. What it will not do is go and update the CRM record behind it.

Scope it tightly. The agent may write to notes and activity fields. It should not touch deal value, close date or stage, because those encode decisions that live in someone's head and an agent asked to infer them will produce something confident and wrong. If the concept of an agent doing this is still fuzzy, what an AI agent actually is covers the definition underneath all of it.

Give it the narrowest access that works

This is the ten-minute step that decides whether your security review approves any of the above, and it is worth doing before you connect anything.

Google publishes a graded list of Gmail API scopes, and the grading is the useful part. There is a read-only scope. There is a separate scope for managing labels. There is a modify scope that covers changing messages. And there is full account access, which includes permanent deletion.

A triage or labelling agent needs the modify scope, because applying a label is a write and modify already includes the reading it needs. If the agent only reads and never writes, read-only on its own is the right answer. Full mailbox access is the one to refuse outright, and if a tool asks for it to do something that reads and tags messages, that tells you something about the tool. Sending is its own scope and deserves its own decision, taken later and separately, after the agent has produced drafts you agreed with for several weeks running.

One more that people skip: log where the agent's runs are visible to a human. The failure mode of scheduled automation is not doing something dramatic and wrong. It is silently doing nothing after a credential expires, for weeks, while you assume it is working. Our guide to giving an agent access to your email safely goes through the scope-by-scope decisions and the revocation path.

Which should you start with?

Start from the complaint you actually have. It is a better router than any feature list.

If writing emails is what eats your day, you already have the fix inside Gmail. Use it and buy nothing.

If things go missing, start with attachment filing. It pays back the first time someone asks for a document.

If the volume is the problem, start with the morning triage digest, read-only, for two weeks.

If deals go quiet and nobody notices, start with the chase-up. It is the cheapest to run and the easiest to measure, because you can count the replies it recovers.

If your mail lives in Outlook rather than Gmail, the boundary in the table above holds exactly the same way, and AI agent triage for Outlook covers the Microsoft 365 permission model instead.

Whichever you pick, the honest measure after a fortnight is the acceptance rate. If you are keeping fewer than half of the things it hands you, the first things I would check are your instructions and your inbox hygiene rather than the agent, and switching tools will not fix either.

FAQ

Sources