Office workflows worth automating first

Using ChatGPT at work is safe when you automate the draft and keep the decision. The four workflows below are built around that rule.

The chatpt workbench can shape the first draft; the rules below are what keep it safe to send.

Most “AI at work” advice is a list of jobs you do not have. You do not write 40 blog posts a week. You do not run a customer support team. You write the same four kinds of text over and over and lose an hour a day to them.

These four workflows are chosen because they are common, repetitive, and safe to draft with a review step at the end.

The rule that keeps you out of trouble

Draft inside the tool. Decide outside it.

A model can produce the first version of an email, a summary, or a status update. It cannot know which client is about to churn, which number is politically sensitive, or which decision your boss has already made. The moment a workflow asks the model to decide something, you have automated the wrong part.

Workflow 1 — Inbox triage

The goal is not to answer email automatically. It is to stop reading the same message three times while deciding what it is.

Here are five emails I need to deal with today. For each one, give me:
- a one-line summary,
- what the sender actually wants (a decision, information, or nothing),
- the smallest reply that closes it,
- a suggested priority: now / today / this week / archive.

Do not write the full replies yet. [paste emails, with names removed if needed]

Then pick the two that matter and ask for a draft with your voice:

Draft the reply to email 2. I want to say no without closing the door.
Keep it under 90 words. No apology paragraph. Offer one alternative date.

Why it works: triage is a classification task, and classification is what the tool is good at. Writing five full replies in one go is where it produces generic politeness.

Workflow 2 — Meeting notes into actions

The failure mode of meeting notes is a transcript with no owner and no date. Fix that with a strict output contract.

Turn these meeting notes into:

1. Decisions made — only things that were actually decided, with who decided.
2. Actions — owner, verb, due date. If an owner or date was not stated, write TBD, do not guess.
3. Open questions — the things that were raised and not resolved.
4. One line: what the meeting was supposed to achieve.

[paste notes or transcript]

The “do not guess” line is doing the work. Without it, you get a confident action list with names that were never in the room.

Then send the action list to the attendees. The list, not the transcript, is the artifact.

Workflow 3 — The status update nobody wants to write

Weekly updates are a formatting problem disguised as a writing problem. Give the model the raw material and the shape.

Write my weekly update from these notes.
Shape: three sections — Shipped, In progress, Blocked.
Rules: one line per item. Lead with the outcome, not the activity.
No adjectives. No "excited to". If something slipped, say it slipped and by how long.
Under 180 words.

[paste notes]

Why it works: the shape is fixed, so the model cannot pad. The “outcome, not activity” rule is the difference between “worked on onboarding” and “onboarding drop-off cut from step 3 to step 2”.

Workflow 4 — Document review before a human review

Do not ask it to approve a document. Ask it to find the four things a reviewer will find.

Review this document as [role: a finance reviewer / a legal reviewer / an impatient executive].

Return a table: location | problem | severity (blocker / should-fix / nit) | suggested fix.
Look for: unsupported claims, numbers without a source, internal contradictions,
undefined terms, and anything that commits us to something we cannot deliver.
Do not rewrite the document.

[paste document]

That last instruction matters. If it rewrites, you lose the ability to see what changed. A review table is auditable; a rewritten document is a new document.

What never goes into the box

This is not legal advice; it is a risk rule you can apply without a lawyer.

• Credentials. Passwords, API keys, tokens, connection strings.
• Other people’s personal data. Customer lists, HR records, health information, anything that identifies a person who did not agree to this.
• Unreleased financials and deals. If the number would move a share price or a negotiation, it does not belong in a chat window.
• Privileged legal material. Attorney-client communication, investigation notes.
• Anything you would not email to an outside contractor.

If your employer has an AI policy, it overrides this list. Read it once; it is usually shorter than you expect.

The review loop

Every workflow above ends with a human. Make that explicit in the process, not in your head:

1. Draft — model.
2. Check — the person who owns the outcome reads it against the source.
3. Sign — the same person sends it.

The failure case is a draft that gets forwarded because it looked finished. If a document leaves your team without a named owner reading it, you do not have a workflow; you have an outage waiting for a date.

What not to automate

• Performance reviews and anything that decides a person’s rating or pay.
• Hiring rejections and disciplinary messages.
• Medical, legal, or financial advice to a specific person.
• Crisis communication.
• Anything where the tone is the message.

In those cases the model can help you prepare — collect facts, list options, rehearse a conversation — but the words that go out should be yours, and preferably a second human’s.

A small setup that saves time

• Write custom instructions with your role, audience, and formatting rules. It removes half the per-prompt setup.
• Use a project for a recurring body of work (a client, a product, a quarterly plan) so the files and instructions stay together.
• Keep a snippet file with the four prompts above, filled in for your context. Copy, paste, replace the brackets.
• Review the setup once a quarter. Plans, features, and your job change.

Related

• Use it like an editor, not a vending machine — the same draft-review discipline for longer writing.
• Why it forgets, and what to do about it — custom instructions and projects, explained.
• Prompts for work — 17 prompts built for these workflows.
• Open the workbench — turn a rough request into a work brief.