Dev.to AI 🤖 Ai 👁 0 📖 9 min read

Handing my inbox to an AI employee: automate what can be undone, approve only what cannot

This is part 1 of the series "Building an AI Office — Handing Email, Admin Work and Alerts to AI Employees." Series contents Handing my inbox to an AI employee until it reached zero (this article) Don't let the AI gue

Handing my inbox to an AI employee: automate what can be undone, approve only what cannot

This is part 1 of the series "Building an AI Office — Handing Email, Admin Work and Alerts to AI Employees."

Series contents

  1. Handing my inbox to an AI employee until it reached zero (this article)
  2. Don't let the AI guess your company details: a 139-item company profile (coming soon)

I set up my AI mail worker so that it could not touch a single email without my approval. After a day of that, my inbox held 113 emails.

Approving everything was safe, and it was the reason the inbox kept growing. Nothing moved until someone clicked Approve, and approvals could not keep up.

What got the inbox to zero was one question asked of every operation: can this be undone? Labels and archiving can, so they run every hour without asking. Sending and spending cannot, and the trash empties itself after 30 days, so those still wait for me (except trashing GitHub notifications, which I approved in advance as a standing rule).

The rule that got my inbox to zero

Operation Can it be undone? Handling
Add a label, archive Yes Every hour, no approval
Hand alerts, candidates and admin tasks to a worker, label and archive Yes Every hour, no approval
Trash GitHub notifications that need no action Only for 30 days Approved in advance as a standing rule
Pass pull request review requests to the review worker Yes Every hour, no approval (approving the pull request itself waits for me)
Trash anything else Only for 30 days Needs approval
Create a label outside the fixed four Yes Needs approval (no labels piling up)
Send, forward No Needs approval (the mail worker cannot do it at all)

The four fixed labels are alerts, candidates, admin tasks and needs-me. Instead of approving each email, I read the run report and comment only when something is wrong. Later, even needs-me mail was archived, and every processed email was marked as read. Anything that costs money was left to me, whatever the amount. A rule added along the way, letting small amounts through, was withdrawn.

Why approving everything filled the inbox

Under the propose-and-approve setup, the mail worker only posted a numbered proposal every hour and touched nothing. I approved with one click on Paperclip's approval screen, after writing the numbers of any wrong items in a comment ("all OK except 3 and 7"). If I did nothing, nothing happened.

It worked, as far as it went. The first approval covered 70 emails: 19 archived, 23 trashed, 28 left in the inbox. A later one-by-one check of each instruction against what was done found that 67 matched, 2 matched with conditions and 1 did not. The one mismatch had been trashed because it shared a subject with a different numbered item, and it came back out of the trash only because the one-by-one check caught it.

The Tasks screen in Paperclip. The lower half is a column of hourly "inbox tidying proposal" tasks, most of them withdrawn. The upper half shows the tasks after the switch: "inbox tidying, first run after the switch, aiming for zero" and "second run after the switch, 56 left," with their subtasks for alerts, blog candidates and skill candidates (the screen is in Japanese). Two rows (a booking and another company's pull request) and most of the subtask titles (email subjects and article names) are blurred.

After about a day, I opened Gmail and thought: this is not tidy. The inbox held 113 emails, 60 of them from the past seven days. The top senders were note.com with 36 and github.com with 8.

Approvals had been made lighter: proposal messages in Slack only twice a day, and only one pending request at a time. All that did was make approvals easier to put off. Proposals piled up, and the mail stayed where it was. An empty inbox and approve-everything did not fit together, so I split the operations by whether they can be undone.

Before that: fully automatic

Approve-everything was itself a reaction. The first setup ran fully automatically.

After a read-only trial, Claude Code, which was building the setup, asked me some questions. I answered "1 ok" and "once an hour." Question 1 had been whether the worker could add labels. Claude Code took the reply as approval to run in production. The hourly run started, and its first pass processed 40 of the 57 unprocessed emails that had piled up since March.

What labels had it created? What had it archived? I had never looked. I had it stop, and the AI undid the run. By then 14 emails had been archived. All 14 went back to the inbox, and the labels the worker had added came off, including the ai-office/processed marker. The result was verified by reading Gmail through a separate path, not through the worker.

The undo worked only because the worker had done nothing but label and archive. Its instructions forbade sending, forwarding and trash from the start.

The sorting was not the problem. The problem was that I did not know the effect before it ran. To an AI, "1 ok" and "once an hour" can mean everything, including the backlog, starting now. The missing step was showing a person what would happen, the whole backlog or only new mail, before running it.

Two runs after the switch

The first full run under the new rule processed 60 emails: 10 tidied, 3 alerts, 32 candidates, 2 admin tasks, 13 needs-me, none deleted. It archived 47, and the inbox, which had taken in new mail in the meantime, went from 116 to 69. The next run processed the remaining 56.

A Paperclip task, "inbox tidying, second run after the switch, 56 left." The instruction says to follow the "inbox is always zero" rule and process up to 60 emails outside the needs-me label, oldest first. The mail worker's reply, in Japanese, reads: 56 processed; 18 tidied (10 archived, 8 deleted), 14 alerts, 3 blog, 1 skill, 1 admin task, 19 needs-me. Inbox: 32 needs-me, 0 unprocessed. 29 archived in total, 2 subtasks created.

A read of the inbox through a separate path confirmed 32 emails left, all needs-me, with zero unprocessed. The two runs took about six minutes and about 2.75 million input tokens in total, most of it read from cache.

Getting to zero

Once needs-me mail was archived too, the inbox should have been empty. Nine emails stayed. There were two causes.

  • One worker's instructions still said "keep actions waiting for Uehara (me) in the inbox." That line was missed when the rule became "the inbox is always zero." A worker does exactly what its instructions say, so a policy change means searching the whole instruction file for any line that contradicts it.
  • The mail worker kept stopping on Gmail's RATE_LIMIT_EXCEEDED error. All the workers together made 333 Gmail calls a day, not a volume that called for spacing out. The error body named defaultPerMinutePerProject: a per-minute limit shared by everyone using OpenAI's Gmail connector, so spacing out my own calls would not avoid it. The instructions said "if a call fails, stop and report," and the worker did exactly that. From then on it retried three times, after 60, 90 and 120 seconds, and carried the work over to the next run if all three failed.

With both fixed, the worker handled the last nine.

A Paperclip task marked Done: "mark the ai-office needs-me emails left in the inbox as read and archive them (one-time)." The instruction asks for the nine labeled emails to be marked read and archived with their label kept, then for the inbox to be counted to confirm zero. The mail worker's reply, in Japanese, reads: done; 9 needs-me emails marked read and archived, label kept, 0 new tasks. Inbox searched again, 0 remaining. 0 unprocessed.

A fresh search showed zero emails left and zero unprocessed. From then on, new mail was sorted by the hourly run, labeled and archived, and what I looked at was decision cards (questions that workers raise through Paperclip's decisions feature, answered by picking an option), not the inbox.

An approval has to be decidable on the screen

Keeping the irreversible operations behind approval raised a new question: what is in the approval?

The admin worker handles needs-me mail that requires an outside procedure. All 16 tasks handed to it were blocked, because the only runbook it had covered one procedure, downloading one vendor's invoice PDF. Zero approvals were waiting. From where I sat, nothing was happening.

Its instructions were rewritten so that a task without a runbook is exactly the kind it should research itself. Each task was to end in one of four ways: information only (summarize and archive), one approval request with the facts for me to decide, a drafted reply that is sent only after approval, or the official procedure written out with where to click and what to enter. The first line of every comment had to say what I have to do: nothing, one approval, or a specific action.

On the rerun, all 16 moved. One needed nothing, one needed me to paste text that only existed on another site's screen, and 14 became approval requests with facts and a recommendation. I approved 13 and rejected 1 in a few minutes. The burden was never the count. It was proposals with nothing to decide on.

Replies follow the same idea. In the first design, a reply's approval request carried only a summary such as "send one reply confirming the open items," and the full text sat in the task's document. From then on, every reply got its own send approval, separate from approving the plan, carrying the recipient, the subject and the full text. "See the task for details" is an approval in name only.

Corrections that carry over

If I have to repeat the same correction every run, handing the work over loses much of its point. Paperclip stored approvals and comments, but nothing carried them into the next run. The one thing that reliably reached a worker every time was its instruction file.

So a small script was set to run every 15 minutes, collect only my approvals, rejections and corrective comments, and write them into a marked section of each worker's instructions. Each decision was tagged as standing or one-time, so that "approved just this once" does not turn into policy. That is why commenting only when something is wrong is enough.

What it runs on

The workers ran in Paperclip, a tool that runs a group of AI workers, on my Mac as my personal assistant, not as a company system. I call the workers AI employees because I hand them work, but they are not counted as my company's AI staff; an earlier post explains where this setup sits in my company. I had Claude Code build the setup and made the calls myself. The agent doing the thinking and the tool calls is OpenAI's Codex, running under a ChatGPT subscription.

Building a custom Gmail API tool was already on the table, because Paperclip's own Gmail connection could only read and draft and needed a Google review. Then it turned out my Codex already had OpenAI's official Gmail plugin, and even headless codex exec could use its 21 operations, including search, read, label, archive, draft, send, forward and trash. Don't trust a conclusion you checked only one way.

The mail worker's first instructions sorted every email into five groups: just tidy it, a blog topic candidate, a skill candidate (a reusable AI work procedure), a cost or monitoring alert, or something only I can handle. It was never to follow instructions written inside an email. A read-only trial sorted the last 24 hours, 13 emails, in about four minutes. That run used 560,000 input tokens, and the worker had inherited my whole Codex setup, database writes and GitHub included, so it got a dedicated config with the Gmail plugin and nothing else.

Blog, skill and alerts workers came when the fully automatic runs began, and admin and code review workers later. The blog and skill workers had no Gmail and could only write to their own folders. A small entry script refused to start Codex once 80% of the plan's usage limit had been used, because one run could reach 500,000 to a million tokens.

The Agents screen in Paperclip. Six agents are listed: skill, blog, mail, review, admin and alerts workers (the names are in Japanese), all running on Codex. Only the mail worker shows Live and running; the rest are idle.

Takeaway

Decide what an AI worker may do from the goal, not from how few approvals you can get away with. With an empty inbox as the goal, automate what can be undone, approve only what cannot, and give every remaining approval what it takes to decide on the spot.

In this series, I record, with real screens and real counts, how I hand our everyday work to AI workers ("AI employees").
If this was useful, please like and follow.

Our products are listed at https://www.eln.ne.jp/products

EarthLink Network builds its own products to run the whole company on AI — each one born from a need along the way. See the full index of what we build:

→ The 18 products EarthLink Network builds in-house

For more about the company and each product, visit www.eln.ne.jp.

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.