How to put an AI agent in charge of your email: step by step, on video

What email triage solves

The problem isn't "having email". It's that your inbox mixes three things that don't deserve the same treatment: a client waiting for an answer, an invoice with a date on it, and a newsletter. If you deal with them in the order they arrive, the ones that matter get buried.

Triage puts your AI agent to work reading that inbox with your criteria. It sorts. It flags what's urgent. And when a reply is due, it doesn't send: it leaves you a draft. You send what counts.

It doesn't replace your judgement. It takes away the work that comes first: opening, deciding whether it matters, starting to write from scratch.

Turn on triage

You turn it on from the hablo Customer area, not from Telegram. Go in, Channels, scroll down to Email triage and press Activate. The agent takes a minute to get ready (it restarts the service) and the panel is set.

The link to the panel carries a token that expires. That's why you always open it from Channels, and not from an old saved URL. If the token has expired, go back to Channels and open it again.

1. Go into the Customer area.
2. Channels.
3. Email triage card → Activate.
4. Wait for it to go from "Activating…" to on.
5. Open the panel from that same place.

The panel: a map of what you're going to set up

The panel is the inbox-helper. Five places that matter, in this order:

• Status. Whether it's alive: last pass, emails seen, rules, whether the model is answering. From here you can also trigger a pass by hand.
• Accounts. Which mailbox it reads and whether it can leave drafts or send. Always start on drafts only.
• Instructions. What matters to you, in free text. That is what scores each email (0–10).
• Processed. The history: what it did with each email. For reviewing and correcting.
• Automations. The rules ("if this happens, do that"). This is where the real work lives.

Config and Log are for later on. You don't need them for the first flow.

Connect your mailbox

Without a mailbox, the panel has nothing to read. Go to Accounts → Add account. An internal name (for example "Gonzalo's email"), the address, the type (Gmail in this case) and the password.

With Gmail you don't use the account password. You create a sixteen-letter app password in your Google account — you manage it from myaccount.google.com/apppasswords. You can revoke it whenever you like, without changing your normal access to your email.

1. Accounts → Add account.
2. Name, address, type Gmail.
3. Paste the app password (create it here — never your login one).
4. Create and verify. The "verified" stamp has to show up.
5. Sending: Drafts only. Agent reading: Allow.

Drafts only is the important decision. The agent can write into Gmail Drafts. It doesn't press Send. You open the draft, touch it up and send it.

On_score: noise, marked as read

On the account's card, below Drafts only, is the on_score table. The model already gives every email a score of 0–10. Here you say what to do with that score, inside the mailbox itself: mark as read, archive, label, move, notify by Telegram.

The action that cleans up the inbox most: mark_read. You set a range, for example min 0 and max 2, and anything that is noise gets marked as read in Gmail. You open your email and it isn't shouting at you. Anything on 8 or 9 stays unread, so you deal with it.

Every row is evaluated. You can have several actions (one range for read, another to notify). Anything that fits no range is left alone.

Soul: who you are and how you want replies written

Before the rules, the agent has to know who it works for. That doesn't go in the triage panel: it goes in the soul of the agent, in OpenClaw. Go into the interface, Agents, pick the email agent (not the one for another job), Files, SOUL.md, and write it in plain language.

That is where the character is set: whose mailbox it is, what they do for a living, tone, signature, and above all what it must not make up. If this is left empty or talks about a different business, the drafts come out in someone else's voice.

If you have several agents, each one has its own soul. The one handling WhatsApp enquiries isn't the email one. You pick it in the Agents dropdown before opening SOUL.md.

Instructions: what matters in your inbox

In the panel, the Instructionstab. It's free text that goes into every classification. This isn't where you write the personality (that's the soul). You write the scale: what a 9 is and what a 1 is.

A 0–10 scale is enough. The urgent at the top, the noise at the bottom. If an email asks for something with a date coming up, the score can go up. You save. From then on every email comes out scored with your criteria, not a generic one.

Rules: what to do when X happens

Instructions score. Automations act. Go to Automations → new rule. Condition (subject or content) and action (leave a draft, notify, ignore…).

The first useful case is a narrow one: a word you already use with clients. "quote" in the subject. Action: draft a reply, professional, short, without sending. Create rule. It's live.

You can have several. Order counts: the specific ones first (quote, invoice, a particular client). At the end, a safety net: if the email scored 8 or more and hasn't matched any earlier rule, generate a draft with the suggested reply. That way you don't miss an important client who didn't say the magic word.

Time-based automations, not just when an email arrives

The rules above react to an email. There is another type: scheduled by time. It doesn't wait for an email. It runs on its own, on a date and at a time.

New automation → type Scheduled by time. You choose the frequency (daily, weekly, monthly), the day, and "not before" that time. In Instructions you write what it has to do that day: review, summarise, leave you drafts. Sending doesn't change: drafts only. You send.

A real email, from start to finish

To watch it go round: an email arrives with "quote" in the subject. In Status you can trigger a pass by hand (no need to wait for the next one). The agent reads it, classifies it with the instructions, applies the rule and leaves the proposal.

In Processed you see what it did. The text of the draft has to answer that email, not an invented one. If the client gave no price and no deadline, the agent doesn't make them up: it leaves them in square brackets for you to fill in.

Then you open Gmail → Drafts, touch up a couple of lines, send. The agent hasn't spoken to the client in your name.

Processed: you teach it what matters

Processed holds the history: each email, the summary, the assessment, the score and the label. At the bottom of each card: open in Gmail, correct the score, save, automate.

If the system gave it an 8 and for you it's a 9, you change it and press Save correction. That doesn't rewrite the past: it teaches the criteria for the next ones. Automate opens a rule already filled in from that email.

Corrections stay "unconsolidated" until you press Consolidate context in Instructions: the model rewrites the instructions taking in what you have corrected, and keeps the previous version in the history.

Recap

In this order:

1. Customer area → Channels → turn on triage and open the panel.
2. Connect the mailbox. Verify. Drafts only.
3. Soul of the email agent: who you are and how replies are written.
4. Instructions: what a 9 is and what is noise.
5. On_score: the noise (0–2) gets marked as read.
6. Specific rules (quote → draft) and, at the end, the safety net for anything scoring 8 or more. Plus the ones scheduled by time.
7. Processed: you correct the score, save, consolidate in Instructions.
8. A test pass and a look at the draft.

With that, the agent is working your email. Later on you can add rules, mailboxes or a finer tone. Control stays with the sending: until you press the button, nothing goes out.