Workflow 004: Ella, The Inbox Agent

I haven't checked my own email in months.

There's no Gmail tab open on my laptop. It used to be there twenty four hours a day.

An email agent reads your inbox on a schedule, sorts and labels what came in, and writes draft replies to anything it recognizes. It doesn't send. A person opens the drafts, checks them, and hits send. Mine is called Ella, she runs at 7am on Claude connected straight to Gmail, and she was built by my assistant out of the SOP my assistant already had for managing my inbox.

That last part is the whole story.

What this replaces

Two hours a day of an assistant opening an inbox cold. Reading everything. Deciding what matters. Writing the same replies to the same questions. Then finally getting to the work that actually needed a human.

Before this, my assistant started every morning at zero. Now she starts with a sorted inbox and a stack of drafts waiting for review.

And I stopped being a stop in the chain entirely.

Blueprint

Every weekday at 7am, the agent opens the inbox, works through what's new, applies labels, and writes draft responses for anything matching a known pattern. Everything lands in drafts. Nothing gets sent.

Trigger: 7am, every weekday. Inputs: the inbox, plus the SOP describing how these emails get handled. Output: a labeled inbox and a folder of drafts ready for a human to review and send.

It runs at 7am on purpose. My assistant starts around 8:30. Ella starts work before she does, so the inbox is already cleaned by the time a person sits down to it.

Organize

What has to exist first:

• A real SOP for the inbox. Not a vague description, the actual process: open the inbox, look for these categories, here's what we do with each one
• The recurring email types: meeting invites, booking requests, potential client enquiries, whatever repeats in your world
• The standard responses to your common questions, in the words you actually use
• Gmail connected to Claude
• A person who owns the review step

If nobody has written the inbox process down yet, that's the work. The agent build is the easy part afterward.

Systematize

Here's the part most people get backwards.

I didn't design this agent. My assistant did, out of her own process. She already had a documented procedure for managing my inbox: how she opens it, which keywords she scans for, which emails come up constantly, what she does with each one, how she replies. We took that document and taught it to the agent.

No n8n. No Zapier. No code. Just Claude connected to Gmail.

You are [AGENT NAME], my inbox agent. Every weekday at [TIME], go through
my inbox, sort it, and prepare draft replies.

Context:
- Whose inbox this is: [NAME AND ROLE]
- Who reviews your drafts: [NAME]
- Tone for replies: [DESCRIBE, OR POINT TO YOUR ME SKILL]

Do this, in order:

1. Read everything that arrived since your last run.

2. Sort and label into these categories:
   [LIST YOUR CATEGORIES, e.g. meeting invites, new enquiries,
   existing clients, newsletters, invoices, needs the founder]

3. For anything matching a known pattern, write a draft reply and save it
   to drafts. Match the tone of previous replies in the thread.
   Common patterns and how to handle them:
   - [PATTERN]: [HOW WE RESPOND]
   - [PATTERN]: [HOW WE RESPOND]
   - [PATTERN]: [HOW WE RESPOND]

4. Flag anything unusual, sensitive, or that only [FOUNDER NAME] can answer.
   Do not draft those. Label them and leave them.

5. Never send anything. Everything goes to drafts. A person reviews and sends.

Rules:
- If you're not confident about a reply, don't draft it. Flag it instead.
- Never invent facts, availability, prices, or commitments.
- Match how we've replied to similar emails before.
- If a connection fails, say so plainly rather than skipping silently.

How to set it up

1. Write the SOP first, or better, have the person currently doing the job write it. It should describe what they actually do, not what the job description says.
2. Connect Gmail to Claude.
3. Run it manually. It will get things wrong the first few times.
4. Give it feedback out loud after each run: you did this right, do more of this, that reply was wrong. Fold the corrections into the file.
5. Once it's steady, schedule it to run before whoever handles the inbox starts their day.

Assign

The agent sorts, labels, and drafts. A human reviews and sends. That line never moves.

It would be easy to let it send. It's also the fastest way to have an AI commit you to a meeting that doesn't exist or answer a sensitive email with a confident guess. Drafting is reversible. Sending isn't.

Something happened last week that made the point better than I could.

I had gallbladder surgery. Someone had been following up with me, and Ella drafted a reply explaining that I'd been in hospital for a week and would be slower than usual.

Nobody told her to write that.

My assistant had been sending that message manually to other people, and Ella picked up the pattern from the replies going out around her. We sat there going, wait, did you write that or did she?

That's the good and the alarming version of the same thing. The agent learns from what the humans around it do, which makes it better every week, and also means it can pick up context nobody explicitly approved. If that draft had sent itself to the wrong person, it would have shared a medical detail I hadn't chosen to share with them.

It didn't, because it went to drafts, and a person read it first.

Rule of thumb from this build: let AI write, let a human send.

And notice who built this. The person whose job the inbox was is the person who automated it. She didn't lose the role, she moved up the stack: she now owns the review, the judgment calls, and the training of the agent. That's what automating a job actually looks like when it goes well.

Integrate

Ella runs at 7am, and she is the front of a relay:

• She cleans and labels the inbox before anyone opens it
• My morning brief agent reads the result at 8am, so the brief never wastes a line on mail that doesn't concern me
• My assistant arrives at 8:30 to a sorted inbox and a stack of drafts to review and send
• Anything genuinely unusual gets flagged, and only that reaches me

Three steps, two of them automated, one human, and the founder is no longer in the chain at all.

Maintenance is a conversation. Every time a draft misses, my assistant tells the agent what to do differently and the file gets updated. It has gotten steadily better without anyone rebuilding it.

Tomorrow: the meeting processor. What happens to a call after it ends.