How many emails did you open this morning just to find out whether they needed you? And how many of the real to-dos are still sitting somewhere in a long thread?
An email AI assistant promises to sort that for you, and as of this month one may switch on in your team's Gmail without anyone asking. On October 15, Google said AI Inbox is rolling out to paid Workspace users. It pulls suggested to-dos (action items, deliverables and deadlines) out of your email and writes short summaries of the topics running across your threads. The post also says plainly: "There is no admin control for this feature."
For a small business, the useful question is what to do now that an assistant is arriving anyway.
This guide is designed to:
- Explain what an email AI assistant does, using AI Inbox and Copilot in Outlook as examples.
- Help you decide which inboxes should use one first.
- Walk you through a 2-week pilot with a simple scorecard.
- Give you a short set of rules before the whole team turns it on.
After reading it, you will be able to:
- Check whether your Google or Microsoft plan already includes an assistant.
- Pick 2 inboxes and 1 task to test.
- Score the assistant on real email, not a demo.
- Write down what it may read, draft and do on its own.
The basics: what an email AI assistant does
An email AI assistant is software that reads your mail and does something with it. Most do some mix of summarizing threads, drafting replies, finding things ("when did the supplier say the order ships?") and pulling out tasks. The newer ones also act on your mailbox: moving, deleting or filing messages when you ask in plain words.
Here's how the 2 tools most small businesses already have compare, as of October 2026:
| Gmail AI Inbox | Copilot in Outlook | |
|---|---|---|
| What it does | Suggested to-dos and topic summaries in one view | Chat with your mail, plus triage: delete, move, copy, categorize, manage rules and folders |
| Who gets it | Paid Workspace plans, including Business Starter, Standard and Plus | Microsoft 365 Copilot users |
| Where | US only, English (EN-US) | Windows, Mac, web and mobile |
| Admin switch | None | Not covered in the release note |
| Safety check | It's in beta and can't read attachments, delegated or encrypted mail | Asks for confirmation before acting on more than 5 emails or changing inbox rules |
Sources: Google's announcement, the AI Inbox help page and Microsoft's Copilot release notes for September 2026.
Read the help page before anyone opens it. AI Inbox only reads the Primary tab, so Promotions, Social and spam stay out. It skips mail you've archived, deleted, muted or snoozed. And it needs Gmail's smart features turned on, which can take up to 7 days before the first to-dos appear.
On cost, the Workspace pricing page lists "Gemini in Gmail" among the features of its Business plans. Check your own plan and invoice before paying for a separate email tool, since you may already have one.
OK, so the tools are here and some of them are already paid for. Which inboxes should use them first?
Step 1: Map your inboxes before you switch anything on
Larger companies start any AI rollout with a data classification: a list of the information they hold, sorted by how sensitive it is. Your version is a 10-minute list of the inboxes your business runs and what arrives in each.
Write down every inbox: yours, the shared info@ address, accounts, bookings, each staff member. Next to each one, note the most sensitive thing it receives.
Ask yourself:
- Which inboxes get customer personal details, card numbers or health information?
- Which ones get contracts, payroll or bank emails?
- Who has access to each shared inbox today, and does everyone still need it?
- Which inbox wastes the most time on sorting and searching?
Pause and reflect: Look at your list. Which 2 inboxes are busy but low risk? Circle them. Those are your pilot.
Step 2: Pick 1 job for the email AI assistant
A pilot that tests everything at once tells you very little. Choose a single job you can measure.
Good first jobs are the dull ones. Pulling deadlines out of supplier threads. Summarizing a long back-and-forth before a call. Filing receipts into a folder. Drafting a first reply to a common question that a person then edits and sends.
Leave these out of the first round: anything that sends mail on its own, anything that deletes mail, and anything touching the sensitive inboxes from Step 1. Microsoft's triage feature asks before it acts on more than 5 emails, but 4 wrong deletions still hurt.
Step 3: Run a 2-week pilot with a scorecard
Big organizations agree on success measures before a pilot starts, so nobody judges it on a good or bad Tuesday. Do the same, on 1 page.
For a week before the pilot, note roughly how long you spend each day sorting and searching mail in your 2 pilot inboxes. Then turn the assistant on for those inboxes only and keep the same notes for 2 weeks.
| Measure | Before | After 2 weeks |
|---|---|---|
| Minutes a day sorting and searching | ||
| To-dos it found that you'd missed | ||
| To-dos it invented or got wrong | ||
| Deadlines it missed (check a few threads by hand) | ||
| Drafts sent with light edits, not rewritten |
Check its work against the email itself. AI Inbox can't see attachments, so a deadline that only appears in an attached PDF won't show up. That's a gap you'll want to know about before you trust the list.
Pause and reflect: If the assistant missed 1 deadline in 20, would that be acceptable for this inbox? Decide now, before you see the results.
Step 4: Write the rules for your email AI assistant
Email has a risk most AI tools don't. Anyone in the world can send text straight into it. The NIST Generative AI Profile, a US government guide to these risks, describes indirect prompt injection, where attackers hide instructions "into data likely to be retrieved" by an AI system. For an assistant that reads your inbox, that means a stranger's email could try to tell it what to do.
So the rules matter more here than for a writing tool. Here's a short version, adapted from the NIST Generative AI Profile, which also lists data privacy and information security among the risks to manage:
| Rule | Why | Who owns it |
|---|---|---|
| A person sends every reply. The assistant only drafts. | Stops a wrong or injected reply going out | Each user |
| No delete or rule changes by the assistant during the pilot | Mistakes in bulk are hard to undo | Owner |
| Sensitive inboxes from Step 1 stay off until reviewed | Limits what it can read | Owner |
| Payment or bank detail changes are always confirmed by phone | A summary or to-do can repeat a fake request as if it were real | Office manager |
| Check the to-do list against the thread for anything with money or a date | It can miss or invent items | Each user |
If you're in a regulated field (health, legal, finance), ask your lawyer or compliance adviser before an AI tool reads client email.
Step 5: Roll out, then review once a month
If the pilot scored well, add inboxes 1 or 2 at a time, starting with the next lowest risk on your Step 1 list. Tell the team which jobs the assistant is for and show them the rules page before their inbox changes.
Because AI Inbox has no admin switch, each person controls it. The help page says to uncheck "Show AI Inbox" in the Inbox tab of Gmail settings to hide it. Make sure anyone who works in a sensitive inbox knows where that setting is.
Once a month, fill in the scorecard again for 1 day. Vendors change these features often (AI Inbox is labeled beta), so a quick check is worth more than a long review once a year.
Questions owners ask
Do I need to buy a separate email AI assistant?
Maybe not. Check what your Google Workspace or Microsoft 365 plan already includes first. Pay for a separate tool only if your pilot shows a gap the built-in one can't fill.
Can I stop AI Inbox appearing for my team?
Google's announcement says there's no admin control. Each user can hide it in their own Gmail settings. Smart features also have to be on for it to work, so look at how those are set for your accounts.
Is it safe to let it read customer email?
It depends on what's in that email. Start with low-risk inboxes, keep a person sending every reply, and check with your adviser if you handle health, legal or financial information.
Concluding thoughts: start with the inbox list
An email AI assistant can give you back real time, but only once you know which inboxes it reads and what it's allowed to do. Go back to Step 1 and write your inbox list this week.
If you'd like help, an audit can check which AI features your current plans already include and which inboxes carry the most risk, and our training can give your team the rules and prompts to use it well.