AI inbox management for a small sales team: triage, draft, approve
How we run a sales mailbox with AI: labels pick what enters the flow, a model drafts, a person approves. What runs on a schedule and what stays manual.
- AI inbox management works as a pipeline: triage, draft, approve, send, reconcile. Each stage does one checkable thing.
- A label decides what enters the flow. Nothing is picked up by accident.
- The model writes drafts in the mailbox, never sends. A person approves on a small board.
- Repeat messages from the same customer are merged before drafting. Simple and complex cases are queued separately.
AI inbox management, done properly, does not mean a chatbot answering your customers. It means the mailbox is sorted for you, each enquiry that needs an answer already has a sensible draft waiting, and a person spends their time on the decision instead of the typing. This is how we run the sales mailbox of a KH Group company: what the pipeline looks like, what runs on its own, and what we deliberately keep manual.
The pipeline in five stages
- Triage. Incoming messages that belong to the sales flow get a label,
agent-sales. A classifier workflow runs four times a day (07:00, 11:00, 15:00 and 19:00) and picks up only labelled messages. Unlabelled mail is left alone. - Draft. A sales agent loads the company's knowledge base, analyses the enquiry and writes a reply. The reply is saved as a draft in the same thread, and a case is recorded in a database.
- Review. A person gets a notification with a link to a review screen: the customer's message, the agent's analysis and the draft side by side. They can ask for changes in plain words, and the draft is rewritten in place, as many times as needed.
- Send. Approval is a button on the board. It sends the draft and records the result in the same request. Only a logged-in person can press it. Scripts and agents cannot.
- Reconcile. A separate workflow reads the mailbox's sent folder and records what actually left. It is the only source of truth for "the customer has received this".
The whole thing runs on n8n, with Gmail as the mailbox and a Cloudflare D1 database behind the board. Why the send step is human, and the three rules that keep the board honest, are explained in Human-in-the-loop automation.
Why a label, and not "all unread mail"
A sales mailbox holds more than enquiries: supplier invoices, newsletters, replies in ongoing deals, messages from colleagues. A flow that drafts replies to everything produces a lot of drafts nobody wants, and every useless draft trains the reviewer to skim.
A label is a cheap, visible gate. A filter or a person applies it, and the flow only touches what carries it. When the flow has finished with a message, it removes the label, so the same message is never processed twice. You can see at a glance in the mailbox what is waiting for the flow, and you can take something out of it by removing one label.
A schedule, not a live trigger
Our classifier could react to every new message within seconds. We chose four runs a day instead. A sales enquiry does not need a reply in ninety seconds. It needs a good reply the same day. Batches mean the reviewer works through a short queue a few times a day instead of being interrupted constantly, and a problem in one run affects a handful of messages, not a stream.
Housekeeping runs on the same rhythm. A clean-up job moves orphaned and duplicate drafts to the bin, which can be undone, and writes notes about anything suspicious. It never decides on its own that a case is finished.
Merging repeat messages before drafting
Customers often write twice: a question, then a correction an hour later, or three short messages in two minutes. In the first version each message became its own case with its own draft, and the drafts did not know about each other. Twice a customer would have received two replies that contradicted each other, had a person not caught it.
Now the inputs are merged, not the replies. When a new case arrives and the same customer has another open case younger than 14 days, the flow gathers their original messages in order and gives them to the agent as one request: answer every point in a single message. The 14-day window is deliberate. A forgotten case from months ago should not be glued to a fresh enquiry. A person can also merge cases by hand on the board, with a preview that shows the combined input before anything changes.
Two queues: simple and complex
Not every enquiry is equal. A first message from a web form asking about product options is a good fit for a drafted reply. A request for a formal offer, a return-on-investment calculation or a question from an existing customer about their installation needs a salesperson, or at least a much better-informed agent.
Each case therefore carries a tier, basic or complex, and the board shows them as two tiles. Complex cases do not block the simple ones, and the reviewer can clear the easy queue in a few minutes before turning to the cases that need thought.
What stays manual, on purpose
- Sending to a customer. Always a person, always from a logged-in session.
- Stopping a case. Deciding not to answer is allowed, but it needs a written reason and a person.
- Anything outside the flow. Drafts in the mailbox that our flow did not create, such as those from another tool or written by a colleague, get their own status, and the board refuses to send them.
- Permanent deletion. Agents may read, label, archive and write drafts. Deleting and sending are reserved for people.
How to start with your own mailbox
Do not start with the whole pipeline. Start with one label and one workflow that reads labelled messages and writes a draft, and do nothing else for a week. Read the drafts. You will learn more about what your knowledge base is missing in that week than in a month of planning. Add the review board when drafts are good enough that you want to approve them quickly, and the merge logic when you see the first pair of contradicting drafts.
The Academy lesson Your first automation builds the first step in n8n: a message arrives, a model summarises and prioritises it, a person is notified. Working with an AI agent covers the loop of describing, checking and refining that you will use to tune the drafts. How to make those drafts sound like your company is the subject of AI email drafting that sounds like you.
Frequently asked questions
Can AI answer customer emails on its own?
It can draft them well. Whether it should send them is a separate decision. We let AI triage and draft, and a person approves every message that goes to a customer. Only reversible steps run unattended.
How often should an AI inbox flow run?
Ours runs four times a day on a schedule. That is frequent enough that an enquiry gets a draft within hours, and calm enough that a person reviews in batches instead of being interrupted all day.
What happens when the same customer writes several times?
If a customer has another open case younger than 14 days, the messages are merged in order and the agent writes one reply that covers all of them, instead of several drafts that contradict each other.
Which tools do you need?
A mailbox with labels and drafts (we use Gmail), an automation tool that can call a model and keep a log (we use n8n), a place to keep the knowledge base, and a small board where a person sees the queue.
Written from our own production systems. Drafted with AI assistance and reviewed by Kamil Hrbáček.
Want this running in your company?
Flowsmith designs and operates automations like this for small and mid-sized teams — with a person approving anything that leaves the building.
