The pattern repeats often enough to be worth naming. A team buys the AI add-on that came with the subscription. People start summarizing long threads before handoffs, a couple of them say their inbox feels lighter, nobody asks for the old workflow back. Weeks later somebody notices that the support address, the one the tool was bought for, was never touched. It isn't a misconfiguration. The feature doesn't run on shared mailboxes at all, and the sentence saying so sits in a vendor FAQ nobody reads before signing.
That's the gap this category keeps falling into. Almost everything written about AI email assistants measures one person's inbox: your drafts, your speed, your morning. If the mail that costs you money arrives at an address three or four people answer, most of that writing is about somebody else's problem.
The short version
- An AI email assistant does four separable jobs: it drafts replies, summarizes threads, sorts and prioritizes the queue, and acts on the mailbox. Most teams need one of them and buy all four.
- Three types exist, and they differ by where the AI runs: inside the client you already pay for, in a client or overlay you add on top of a mailbox, or on the shared address itself. Only the third puts what the AI decided in front of the whole team.
- The AI in your Microsoft 365 or Google Workspace plan is real and often enough for personal mail. Microsoft documents that Copilot in Outlook runs on a user's primary mailbox only, and not on shared, group, delegate or archive mailboxes.
- The best evidence on AI drafting comes from clinical studies, not vendors. In one, assisted drafting produced no measurable time saving but a large drop in task load and exhaustion. In another, reviewers judged 7.1% of drafts as posing a risk of severe harm.
- Prices are quoted per seat, per mailbox, per message volume, per resolved conversation, or bundled into something you already pay for. The mechanic decides what your invoice does when the business changes.
- If you're two people and the volume is fine, you probably don't need to buy anything yet.
What an AI email assistant actually does
An AI email assistant is software that reads your mail and then writes, condenses, sorts or acts on it. In practice that means four jobs: drafting a reply, summarizing a thread, prioritizing what lands in front of you, and doing something to the mailbox itself like archiving, moving or categorizing.
They get sold as one feature. They fail differently, they need different amounts of supervision, and the one you need depends entirely on where your time goes.
| Job | What it reads | What you get | Who has to check it |
|---|---|---|---|
| Drafting | The incoming message, the thread, sometimes your past replies | A reply to send, edit or throw away | A person, every time, before it goes out |
| Summarizing | A long thread, an attachment, a day of mail | Four lines instead of forty | Whoever acts on the summary |
| Sorting and prioritizing | Subject, body, sender, past patterns | Labels, categories, a ranked inbox | Usually nobody, which is the risk |
| Acting on the mailbox | Your instruction, plus whatever it decides matches | Archived, moved, flagged, categorized mail | Depends entirely on the confirmation rules |
Two of these are low stakes. A bad summary wastes a minute, and a bad draft that somebody reads before sending wastes seconds. The other two are where money leaks: a message sorted into the wrong bucket stays invisible until a customer complains, and an action taken across a hundred messages is hard to unwind.
Sorting is the job with the most written about it and the least to say here, because it's a topic of its own. The mechanics of what a classifier decides, how it gets priority wrong, and what to measure before you trust it are covered in AI email triage. One point carries over: sorting quality is invisible until it fails, so switch it on last and check it hardest.
Three types, and the question each one answers
The category splits by where the AI runs. That sounds like an implementation detail, and it decides what you can get out of it.
Type 1: the AI in the mail client you already pay for. Your subscription includes drafting, summarizing and some inbox actions. Nothing to install, nothing to migrate, no new invoice. It works on your mailbox, for you.
Type 2: an AI-native client or an overlay you add to a mailbox. A separate app, extension or service that connects to your account and gives you better search, sharper drafting or a priority view. More capable than the built-in features and usually faster, and now you have another tool to adopt, pay for and get people to actually open.
Type 3: an assistant that runs on the address several people answer. Drafts, labels, assignments and notes land in a state everyone on the team can see. It connects to the mailbox you already have rather than replacing your email client, so there's no migration for the people who answer from it. This is the type that can answer the question behind most duplicated and dropped replies: who owns this one.
| Type | Where it runs | What it optimizes | What it can't do | Best for |
|---|---|---|---|---|
| Built into your plan | Inside Gmail or Outlook | Your personal writing and reading speed | Reach a shared or delegate mailbox, or share what it decided | Individual mail, teams already paying for it |
| Added client or overlay | On top of your mailbox | Speed and search for one person | Give a colleague the same view of the queue | Power users with heavy personal volume |
| Shared address assistant | On the mailbox customers write to | Coverage, ownership, consistent replies | Replace personal productivity tooling | Teams answering a support or sales address |
The expensive mistake is buying for the wrong layer: adding a personal AI client to fix a queue problem, or expecting the AI in a subscription to touch an address it was never designed to see.
What the AI in your Gmail or Outlook plan already does, and where it stops
Before evaluating anything new, find out what you're already paying for. Both major platforms ship real drafting and summarization now, and for a lot of teams that's the end of the shopping trip.
In Gmail. Google lists the Gemini features available in Gmail along with who can use them: "Help me write" for drafting in Compose, Proofread, Suggested Replies, conversation summaries, and search overviews. The part worth reading closely is the eligibility column. Google's own feature list shows availability varying by plan, by language and by country: some features are English and US only, others run in eight languages globally, others need a specific Workspace or Google AI tier. "It's in our plan" and "it's available to our team" aren't the same statement.
In Outlook. Microsoft documents the drafting path plainly: in new Outlook, Home > New mail > the Copilot icon > Draft, type the prompt, Generate, then Keep it. Classic Outlook uses Home > New Email > Copilot > Draft with Copilot. After generation you can adjust tone and length or regenerate. Microsoft also documents a set of natural-language mailbox actions: pin, flag with reminders, mark read or unread, archive, delete, move, copy, mark junk, set categories, create folders and manage inbox rules.
That last group deserves attention, because it's the clearest published example of how autonomy gets bounded. Microsoft states that no confirmation is required for actions affecting five or fewer emails, and that above that threshold Copilot shows a confirmation card. Whether a tool has a threshold like that, and whether you can move it, tells you more about what you're buying than any feature list.
Does Copilot work on a shared mailbox?
No. Microsoft's FAQ for Copilot in Outlook states it directly, as of August 2026: "Copilot scenarios in Outlook are only available on a user's primary mailbox. They are not available on a user's archive mailbox, group mailboxes, or shared and delegate mailboxes that they have access to." The same page adds the caveat every vendor should print in the same size as the marketing: "it's important that you review, edit, and verify anything it creates for you".
Read that first sentence against your own setup. If customers write to a shared mailbox, a distribution address or an alias several people watch, the AI in your subscription doesn't run there. It makes each person faster in their own inbox and leaves the queue exactly as it was.
A general version of the same limit applies well beyond one vendor. Anything the platform AI produces for you is personal state: a priority it assigned, a summary it wrote, a draft waiting in your window. None of it is visible to the colleague who opens the queue after you go home. A shared inbox tool like TriageFlow puts that output where the team can see it. If the shared address itself is the thing you haven't sorted out yet, start with how a shared inbox works before adding AI to it.
What these assistants are genuinely good at, and what the evidence says about drafting
Start with the honest positives, because they're real and usually undersold in favor of flashier claims.
Summarizing a long thread before a handoff works well, and it's the feature people keep using after the novelty wears off. First drafts of repetitive replies save real keystrokes when the answer is close to something you've written forty times before. Search is genuinely better with a model behind it: you can describe the message you half remember instead of guessing the keyword. And clearing the obvious sort, the newsletters and receipts and notifications, carries almost no risk.
Drafting is where the claims get big, so it's worth looking at the one setting where somebody measured it properly with trained reviewers in the loop. Two studies from clinical inboxes, which is a high-stakes context and not a support queue. The failure shapes still transfer.
In a five-week prospective study at Stanford Health Care covering 162 clinicians, AI-generated draft replies to patient messages were used about 20% of the time, and there was no statistically significant change in the time spent replying (reply action time increased by 11.8 seconds, P = .19). What changed was how the work felt: task load fell from 61.31 to 47.26 (95% CI -17.38 to -9.50, P < .001) and work exhaustion fell from 1.95 to 1.62 (95% CI -0.50 to -0.17, P < .001). Same clock, less strain.
The second study is the one to read before switching anything on unsupervised. Six board-certified radiation oncologists reviewed drafts written by a large language model, and across 156 survey responses they judged that 11 of them (7.1%) posed a risk of severe harm and one (0.6%) a risk of death, while the same drafts improved subjective efficiency in 120 cases (76.9%). The drafts also ran long: 169 words on average against 34 words for the manual replies.
Both studies are single-setting, clinical, and involve expert reviewers who caught the problems. What transfers isn't the rate, it's the shape of the failure: assisted drafting makes the queue feel lighter without necessarily making it faster, and when it's wrong it's fluent, verbose and mistaken about how serious the situation is.
The criteria that actually decide it
Vendor comparisons collapse under their own feature tables. These are the questions that change the answer, phrased so you can put them straight to a sales call and hear whether the reply is specific or vague.
- Where does the AI act? Suggest only, draft for a human to approve, or send on its own? Three very different risk profiles get sold under one category name.
- What is it grounded in? Your past replies and help content, or the model's general training data? Grounded output sounds like your team; ungrounded output sounds like every other company's email.
- Which mailboxes does it run on? Personal only, shared, delegate, alias, distribution address. Ask this before anything else if customers write to an address more than one person answers.
- How granular is the autonomy? Can you enable it per category, per sender, per queue, or only globally? Is there a threshold above which it asks first, and can you set it?
- Is there an audit trail? When it labels, assigns, archives or sends, can you see afterwards what it did and why? Without that, a wrong decision stays invisible.
- Does its output land in shared or personal state? A draft in one person's window helps one person.
- What happens to your mail? Training opt-out, retention, processing region, and whether the tool inherits your existing access rules or creates its own path to the data.
- How long until it's useful? Some assistants are useful on day one. Some need weeks of corrections first. Ask what "useful" looked like for a team your size.
- What does it do to the client you already use? Connecting to your Gmail, Outlook or IMAP account alongside your existing client is a different project from replacing that client. Replacement means migration, retraining, and a real chance of quiet non-adoption six weeks later.
Those are the AI-layer questions. The container questions, meaning routing rules, seats, channels and how a shared mailbox tool is priced and structured, are a separate decision covered in the shared inbox software buyer's guide. None of it replaces the human process underneath: if nobody has agreed what gets answered first and who owns which category, an assistant will just accelerate the existing mess. Agreeing that is what an email triage pass is for.
What this costs, and how the pricing works
Skip the headline number and look at the mechanic, because the mechanic decides what your invoice does when the business changes.
- Per seat. Predictable, and it taxes headcount growth: every new hire who needs read access costs the same as a full-time agent.
- Per mailbox. Cheap when one address does the work. Expensive once you split sales, billing and support into separate queues.
- Per message or per AI action. Tracks the workload, which is fair, and spikes exactly when a seasonal peak has already stretched you.
- Per resolved conversation. Aligns cost with outcomes, and puts you in a conversation about what counts as resolved.
- Bundled into a subscription you already hold. The cheapest option by a distance, and the reason to check what you have before buying anything.
Four questions move the final number more than the advertised rate: is there a seat minimum, is the discount tied to an annual commitment, what exactly counts as a billable AI action, and what happens when you exceed the allowance (throttled, blocked, or billed at a higher rate).
One anchor for the bundled case, so the mechanics aren't entirely abstract: Microsoft 365 Copilot for business lists $18.00 per user per month, paid yearly as of August 2026, reduced from $21.00 under a promotion that runs to 30 September 2026 and applies to the first year only, and it needs a separate qualifying Microsoft 365 license on top. An earlier version of this article quoted $30 per user per month for the same product three months ago, which is a decent argument for never trusting a price in a blog post, including this one.
Worth knowing when you compare quotes: volume-based pricing does exist in this category, where the bill tracks how much mail arrives and seats are unlimited. For a team that adds people faster than it adds tickets, that mechanic is usually cheaper than per-seat, and it is worth asking for by name.
For the arithmetic on whether any of it pays for itself, headcount against volume and response targets is the calculation that matters, and our support team size calculator will do it faster than a spreadsheet.
Which type fits your bottleneck
"Writing the replies is what takes the time"
Drafting is your job, and the built-in AI in your plan is the cheapest place to test whether it helps. Measure it for two weeks against how the work feels, not just the clock: the Stanford result suggests the honest gain may be strain rather than speed, and that's still worth having.
"Finding what was already said takes the time"
You want search and summarization, not drafting. Both platforms do conversation summaries now, and the added clients in type 2 do it better. Of the three bottlenecks this is the one least likely to justify a new purchase for a small team.
"Deciding who owns a message takes the time"
No amount of personal AI fixes this, because the output stays private to whoever triggered it. You need type 3, and the first question to ask any vendor is whether the thing runs on a shared, group or delegate mailbox at all. If it doesn't, the rest of its capabilities are irrelevant to the problem you're solving. The container that AI layer sits on is its own decision, worked through in the shared inbox software buyer's guide.
If none of the three describes you, that's a real answer too. Two people with a manageable queue should turn on what the plan already includes, agree who covers which days, and revisit when the inbox starts producing missed messages rather than mild annoyance.
Frequently asked questions
What is an AI email assistant?
Software that reads your mail and then drafts, summarizes, sorts or acts on it. It either lives inside the email client you already use or connects to the mailbox from outside. Two things it gets confused with: an email client is where you read and send mail, and the assistant is the layer that decides or writes something on top of it. A chatbot holds a conversation with a customer in a chat window, while an assistant works on mail that already exists in a mailbox. Some products do all three, and they're still separate jobs with separate failure modes.
How do AI email assistants work?
You grant the tool access to the mailbox, usually through your account permissions rather than by handing over a password. It then reads message content and thread history and generates from it: a draft, a summary, a category, or an action. Everything it produces is an estimate, including the parts that sound certain.
Are AI email assistants safe for work email?
Two separate questions hide in that one. What happens to the mail (whether it's used for training, how long it's retained, where it's processed) is answerable from the vendor's documentation, and if it isn't, that's your answer. What the tool is allowed to do unsupervised is the bigger risk, and it's entirely under your control: start with suggest-only, then draft-for-approval, and treat autonomous sending as its own decision with its own review.
Can an AI email assistant send replies without me?
Some can. Whether you should let it depends on what it costs when it's wrong, and the clinical evidence above is worth weighing: fluent drafts, reviewed by specialists, still carried a 7.1% rate of severe-harm risk in that setting. Support queues are lower stakes, but the failure looks the same.
Is this the same as AI email triage?
No, triage is the sorting half of it, and we cover that separately in AI email triage.
How much does an AI email assistant cost, and are any of them free?
Pricing comes in five shapes: per seat, per mailbox, per message or AI action, per resolved conversation, or bundled into a subscription you already hold. The mechanic matters more than the rate, because it decides what happens to the bill when you hire someone or hit a seasonal peak. Free is mostly the bundled case: both major platforms include drafting or summarization on some tiers, so check your plan before you shop. Standalone free tiers are usually rate-limited enough to work as a trial rather than a tool, and a genuinely free product with no subscription behind it is funded by something, so read the privacy policy before connecting it to a work mailbox.