An AI email assistant for a small sales team is easy to imagine wrong. The picture in most owners' heads is an autoresponder with better grammar. A prospect writes in, the AI writes back, deals close while everyone sleeps. That version gets people in trouble. Here's what the good version does, why it drafts instead of sends, and how to tell whether your team is a fit.
The five jobs
1. Triage
Every inbound reply gets read and classified within the hour: real interest, a question, a sample or quote request, an opt-out, an out-of-office, or a message in an active thread. Knowing which pile each email belongs in is most of the triage work a person does today.
2. Drafting in your voice
For anything that needs a reply, the assistant writes one and saves it as a draft in the thread. It's written to your rules: your greetings, your sign-off, your banned phrases, your length limit. A rep opens the draft, edits if needed, and sends. The assistant never sends a client-facing email on its own.
3. Keeping the task board current
Whatever you use to track deals and work, Planner, Trello, Asana, a CRM, the assistant creates and updates the card at the same time it reads the email. New sample request, new card, assigned to the right person, due date set, ship-to address filled in from the thread. Client answers a question, the card's notes update. The board stops drifting from the inbox because the same process touches both.
4. The mechanical replies
Some replies need no judgment. Someone asks to be removed from your list, and the answer is a polite confirmation and a suppression. An auto-reply names a colleague to contact instead, and the answer is a short intro to that colleague. These can run without review, under narrow rules, once you've approved the template.
5. Reporting
On a schedule, the assistant produces a report: what it did on the board, what it drafted, and what it's holding for approval, numbered. The reviewer replies with "send 2 and 5, hold 3." That's the workflow.
Why draft-only is the right default
Owners sometimes hear "it doesn't send" as a limitation. It's the feature.
- Your voice is the product. In a small sales team, clients buy from a person. One off-key email can undo months of outreach. The cost of a bad send is far higher than the cost of a 20-second review.
- Your knowledge isn't written down yet. Pricing, product specs, what a repeat client ordered last year, which accounts belong to whom. In most small teams that lives in two people's heads. Until it's captured, the assistant can't answer those questions correctly, and it shouldn't guess.
- Review is how it learns. Every edit a rep makes to a draft is feedback. Every "don't say that" becomes a rule. In our experience the rulebook grows fast in the first month and then settles. You can't get that from a system that sends before anyone reads.
- Trust is earned in tiers. Board actions are internal and reversible, so they run freely from day one. Emails to clients are drafts. After a category of email has been approved unchanged enough times, it can graduate to auto-send. That's a decision you make later, on evidence.
What should never auto-send
Anything with a price, a promise, or a person's feelings in it. A quote. A delivery date. A reply to a frustrated customer. A follow-up on a proposal that's sitting with the client's decision-maker. These are exactly the emails a person should write, and the assistant's job is to make that fast. The draft is there, the context is there, the card is already updated.
The voice rules that matter most
We've written and refined a lot of these rulebooks now. The rules that make the biggest difference are simple.
- Write like a person answering in under five minutes. If a sentence sounds like AI, rewrite it.
- Answer every question that was asked, directly, and end with exactly one next step.
- Don't narrate that you understood. "Sounds like you'd like option two" gets deleted. "Sounds good, I'll get that moving" stays.
- One exclamation point at most. Contractions on. No em dashes.
- Never manufacture urgency. A follow-up nudge is a warm opener and an open question, then stop.
- Never claim something you don't do. If you don't offer rush shipping, the assistant can't offer rush shipping, and the rule says so in writing.
Put a deterministic check in front of every draft that enforces the mechanical rules automatically. Strip the em dashes, flag the banned words, count the exclamation points. Don't rely on the model remembering.
What it takes to set up
- Scoped mailbox access. With Microsoft 365 or Google Workspace, an app can be limited to specific mailboxes. Do that before it reads anything. The assistant should be unable to read the owner's inbox even if someone asks it to.
- A task tool with an API. Planner, Asana, Trello, most CRMs. If your board is a whiteboard, that's step zero.
- A rulebook in one file. Voice rules, scenarios, who owns what, what escalates. One copy. Two copies drift, and the drift causes bugs.
- A reviewer with a schedule. Someone reads the reports at set times. Four a day works well for a sales inbox. Two might be enough for support.
- A miss log. Every mistake gets a row and a category. A category that repeats means the rule needs to change.
Is your team a fit?
Yes if replies arrive faster than one person can triage them, the same kinds of emails come in every week, and your task board is always a day behind the inbox. Probably not yet if every email is a bespoke negotiation, or nobody on the team will commit to reviewing drafts at a set time.
We built this for a corporate gifting company's two-rep sales team and wrote up what happened, including the order it caught on day one, in our corporate gifting case study. If you want to know what it would look like on your inbox, that's what the free automation audit is for.