By The Botseon team
One bot, one inbox
"The first job to hand over, and what the first week with it looks like: the job description, the approval card, and the facts the bot keeps."
The promise
By the end of this guide the inbox has a bot triaging it, drafting the routine replies and stopping on a card for everything else, and you will have read its first ten drafts and corrected the ones that needed it.
Who this is for
A support team, or anyone who loses the first hour of the day to mail. One person is enough to start.
Set up
Sign in and create a bot from the five starter templates; pick Inbox Triage. Before it can read anything it needs your mailbox: install Gmail or Microsoft 365 from Plugins and connect your account. The sheet shows the scopes it asks for before you agree, and the bot uses your connection, never a copy of your password.
The jobs
Triage the queue
Inbox Triage ships as one of the five starter bots. Its job description is two sentences, and they are in the bot's details pane once you have made it, where you can change them.
Triages the inbox: labels what's urgent, drafts replies to the routine messages, and leaves the
rest for a human to decide. Never sends a reply without approval.
Label and sort incoming mail: Always allow
Draft a reply, without sending it: Always allow
Send a reply: Ask first
9 credits per run, measured over a mailbox of forty unread messages.
Draft the first reply
The bot works the task through in the transcript, one step at a time: what it read, what it decided, what it drafted. When it wants to send a reply it does not send it. It stops, and a card appears with the draft and the reason it wants to send now.
Shot: approval-card (the approval card, desktop and phone).
Read the first ten drafts closely. They are the bot's guess at your voice, and the guess improves fastest when you correct a draft on the card rather than after it.
11 credits per run, measured over five drafted replies.
Escalate what needs a person
Some messages are not the bot's to decide: a complaint that needs a manager, a request the policy does not cover. The bot lists these with a line on why they need a person, rather than guessing.
Not measured: this job needs a person, a paid account or a real mailbox, so no run of it can be
automated.
Now let it run
Open the bot's Memory view after the first few runs. What the bot learned is there as rows of facts you can read, edit and forget: that newsletters can wait, that anything from your accountant is urgent. Each row says where it came from.
Shot: memory-view (the memory view, desktop and phone).
Once the drafts read like you, give the job a clock: open Routines in the bot's details pane, choose "On a schedule," and pick "Every day," "Weekdays" or a custom line. A routine run is an ordinary run, and the bot is quiet unless a card needs you or a run failed.
What the bot may not do
Every action that sends, spends or deletes stops at a card. Sending offers three answers: "Allow once," "Always allow" and "Deny." "Always allow" creates a standing rule for exactly that kind of action, listed where you can read it and remove it, and it expires on its own. Sending a reply carries the external_send permission tag, which is why it is never always-allowed by default.
The line at the top of the Memory view states the rule the whole product runs on: stored facts are data, not instructions. A bot can use them; it can never be given permission by one.
Where your data is
The bot reads and drafts through your own Gmail or Microsoft 365 connection, never a copy of your password. On the managed service everything it reads and stores follows the same rule as the rest of the product: EU/EEA storage, no non-EEA egress by default, US sub-processors under SCCs.
What can go wrong
| Symptom | Cause | What to do |
|---|---|---|
| A draft sits waiting and never sends | Sending always stops at a card, by design | Open the transcript and answer the card yourself |
| The bot will not always-allow a send | Sending carries a permission tag that is never always-allowed | Keep answering the card; there is no setting that removes it |
| A rule you set stopped applying | Standing rules expire after their term | Re-create the rule from the list if you still want it |
firstSeen: 'First time this bot has acted on this.',
allowOnce: 'Allow once',
alwaysAllow: 'Always allow',
Do it yourself in 10 minutes
- Create a bot from the five starter templates and pick Inbox Triage.
- Connect Gmail or Microsoft 365 from Plugins and read the scopes before you agree.
- Answer the opening question with one of the three starting points it offers.
- Watch the first run in the transcript: what it read, what it decided, what it drafted.
- When a draft card appears, read the
approval-cardmoment on the desktop and on a phone. - Correct one draft on the card rather than after it, and watch the next draft improve.
- Open the bot's Memory view and read the
memory-viewmoment it has written so far. - Edit or forget one fact, and confirm the change is reflected the next time it is used.
- Open Routines and put the job on a daily schedule with a Test run first.
- Let one scheduled run go by unattended, and read the transcript the next morning.