Disclosure: we build Mailient, one of the tools in this category. This is written as the explainer we wish we had found while evaluating — including the parts that argue against buying one.
The Three Jobs, and Which Ones Tools Actually Do
“AI email assistant” is used for products that do very different amounts of work. There are three distinct jobs, and most tools do one:
- Deciding. Reading everything that arrived and working out the handful that genuinely need a human. This is the rarest and the most valuable, because the most expensive email in your inbox is the one you never opened.
- Writing. Producing a reply that sounds like you rather than like support. This is the most common feature and the easiest to demo.
- Coordinating.Turning “when are you free?” into real times on a real calendar, and chasing the thread when the other person goes quiet.
When you evaluate, work out which of the three you are buying. A tool that only writes will not stop things slipping, and a tool that only files will never save you the drafting.
Which tools do which job
Names, so this is concrete rather than abstract. Sorted by the job above rather than by rank, because rank is meaningless until you know which job you are buying.
| Job | Tools built around it |
|---|---|
| Mostly writing | MailMaestro, Gemini in Gmail, Superhuman AI, Serif |
| Mostly deciding / cleaning | SaneBox, Inbox Zero, Gmail's own priority sorting |
| Reading and writing inside a faster client | Shortwave, Superhuman, Spark |
| Deciding, writing and coordinating | Fyxer, Jace, Mailient |
Placement reflects what each tool leads with, not a verdict on quality — several do more than one job, and we have used some far more than others. Where we have not used something enough to judge it, we would rather place it by its own positioning than invent an opinion.
What It Will Ask For, and Why
The consent screen is where most people hesitate, and it is worth understanding rather than clicking through.
- Read your mail. Unavoidable. Nothing can tell a customer from a newsletter without reading the message. Any tool claiming to triage without read access is not triaging.
- Create drafts. Separate from sending. This is what lets a reply appear in your drafts folder, ready to review.
- Send mail. A distinct, much larger permission. Ask why any tool needs it. If it does not visibly send on your behalf as a feature you asked for, it should not have this.
- Calendar. Only needed if the tool proposes or books times. Skip it otherwise.
You can revoke all of it from your Google account's security settings in about fifteen seconds, and that fact should inform how nervous you are about trying one.
The Design Question That Matters Most
What does it do without asking?That single question separates a tool whose worst failure is a bad draft you delete from one whose worst failure is a bad email your customer reads. Mailient's default is approval: replies land in your Gmail drafts and wait. Autonomous sending exists, but you have to turn it on yourself, per agent.
This is not us being cautious for its own sake. An assistant you cannot trust to stay in its lane is worse than no assistant, because you end up double-checking everything, which costs more time than writing the replies yourself would have. Approval is what makes delegation restful rather than anxious.
There is a second question worth asking, and almost nobody does: what reaches the model provider? Nearly every tool in this category sends your text to a model it does not own. The better ones strip identifying details first — Mailient replaces names and email addresses with placeholders on the way out and restores them in the result, so the provider never learns who you correspond with. Ask any vendor whether they do this, and separately whether the provider is contractually barred from training on what it receives. Those are two different questions and vendors often answer only the second.
How to Judge One in Ten Minutes
Do not evaluate on the feature list. Every tool in this category lists the same words. Do this instead:
- Connect it and look at what it surfaced. Did it pick the threads you would have picked? If it surfaced a newsletter and missed the customer asking about pricing, no amount of writing quality saves it.
- Read one draft to someone you actually know. Not a sample, not a demo — a reply to a real thread. Ask whether you would send it after editing one sentence, or whether you would start over. Starting over means the voice is not there yet.
- Check what it does with a thread that needs a human. A good assistant should decline to draft some things. One that confidently drafts a reply to a legal notice or a co-founder disagreement does not understand what it is looking at.
This is why Mailient shows you real drafts to your real threads during setup, before you pay anything. It is the only test that means much.
The Honest Limits
Things this category is genuinely not good at yet, ours included:
- Anything needing information that is not written down. If the answer lives in your head or in a call, no assistant will produce it.
- High-stakes relationships. The email to your biggest customer about a mistake is not a delegation candidate. Good assistants flag these rather than drafting them.
- Your voice on day one. A profile inferred from your website and recent threads is a starting point. It sharpens from what you approve and reject, which means the first week is the worst week.
- Deciding what your business should do. Triage is about attention, not strategy.
Setting One Up
Whichever you choose, the sequence is roughly the same and takes under ten minutes: connect Gmail, let it read a few weeks of mail, give it your website so it learns how your company talks, review what it found, and read a draft. If a tool wants payment before you have seen it work on your own inbox, that is a reasonable thing to object to.
Mailient connects to Gmail, Calendar, Slack and Notion, drafts in your voice, and shows you the real replies before the paywall. Start free — or read the founder's buying guide.