Skip to content
Orion Intelligence Agency
← Back to Insights

Set Up AI Inbox Help in Draft Only Mode Before You Let It Send Anything

Tue Oct 20 2026

Yes, if it starts in draft only mode. Sort mail into five categories, let a short named list of routine replies send on their own, and review the drafts you rewrote every week before giving it more freedom.

Yes, an AI assistant can sort your inbox and draft your replies without sending something you would regret, as long as it starts in draft only mode. Begin with five sort categories. Allow only a short, named list of routine replies to send without you. Every week, look at the drafts you rewrote and use what you find to improve the instructions. Give the assistant more freedom only after those reviews show it has earned it. Mistakes still happen with this setup, but they land in your drafts folder, where you catch them, and not in a customer's inbox.

The regret comes from sending, not from sorting

A reply you never meant to send hurts you. A badly sorted email mostly doesn't. If the assistant files a supplier invoice under "newsletters," you lose a few minutes when you go looking for it. If it tells a frustrated customer "Thanks for reaching out, we'll be in touch!" when they asked for a refund for the second time, you can lose the customer. These two failures are not the same size, so they should not get the same permissions.

My position is that sorting can be fully automatic from day one, and sending cannot. Drafting sits between them. The assistant writes, you read, and you press send. Every rule below follows from that split.

You can test this claim in your own business. Run draft only mode for four weeks and count the drafts you would have been embarrassed to send unedited. If that count is zero, you've lost nothing by being careful. If it's above zero, draft only mode just saved you from each one.

Pick five categories and stop

Five categories are enough to sort a small business inbox. More than five makes the assistant less reliable and your review slower. Each extra category adds a boundary the AI has to guess at, and email that sits near a boundary is the email that gets misfiled. Here is a set that fits most service businesses and shops:

1. **Customers needing a reply.** Questions, complaints, booking changes, anything where a real person is waiting on you. 2. **Money.** Invoices, payment confirmations, failed payment notices, quotes you sent that got a response. 3. **Routine requests.** Hours, directions, "did you get my form," appointment confirmations. These are predictable and low stakes. 4. **Vendors and partners.** Suppliers, contractors, your accountant, the platforms you sell through. 5. **Everything else.** Newsletters, promotions, notifications, cold pitches.

The fifth category is the important one. A catch-all bucket lets the assistant say "not sure" without forcing a wrong guess into a category that matters. If you make it choose between "Customers" and "Vendors" for a weird email from a wholesale buyer, it will sometimes choose wrong. If "Everything else" exists, the wrong guesses land in a low-stakes pile you skim once a day.

Put each category definition in writing, one or two sentences each, and include an example sender. "Money: anything from Stripe, QuickBooks, our bank, or a customer mentioning a payment, refund, or invoice number." The assistant follows written definitions far better than a category name alone. Writing the definitions also makes you decide things you had been deciding on the fly.

Draft only mode, set up properly

In draft only mode, the assistant can write a reply and save it, but it has no permission to send. That has to be a hard limit, not just a prompt instruction. Check the permission settings in your tool. Many inbox assistants built into Gmail, Outlook, and help desk products like Front or Help Scout separate "suggest a reply" from "send a reply." If your tool asks for full mailbox access including sending, check whether a narrower scope exists. If no narrower scope exists, use a tool that offers one.

Each draft should carry a short header the assistant fills in, something like:

- Category it assigned - What it thinks the sender wants, in one line - Anything it was unsure about

That third line will save you the most time. "Unsure whether the 15% discount mentioned is still valid" tells you exactly where to look before you hit send. Without it, you have to read the whole draft line by line looking for the problem.

Write the drafting instructions as a short document, not a clever prompt. Include your tone ("plain and friendly, no exclamation points, sign off with first name"), things to never promise ("no delivery dates, no refunds, no discounts without checking"), and a few real replies you've sent that you liked. Those example replies help more than anything else on the page.

The short named list of replies that can send on their own

After the drafts have been reliable for a while, a few reply types can go out without you. Name each one specifically. A loose category like "routine stuff" is too broad. A good starting list looks like this:

- **Hours and location.** "Are you open Saturday?" gets your posted hours and address. - **Receipt confirmations.** "Did you get my form?" gets "Yes, we received it on [date], and we'll reply within [your stated window]." - **Appointment confirmations** for bookings that already exist in your calendar system. - **Unsubscribe or "please remove me" requests**, answered with a confirmation once you've actually removed them.

Every item on that list has the same three properties. The correct answer is already written down somewhere, the reply promises nothing new, and a wrong reply costs very little. If a reply type fails any of the three, it stays in draft.

Refunds, pricing questions, complaints, anything from a customer who has written twice about the same issue, and anything mentioning a lawyer, a review, or a cancellation stay in draft forever. You don't need to revisit that rule. Those emails are where your judgment is the whole value, and the assistant's job there is to save you typing, not to save you thinking.

Put a guard on every auto-send item: if the incoming email also contains anything outside the routine request, it drops back to draft. "Are you open Saturday? Also my last order arrived broken" is not an hours question. Tell the assistant this in plain words: "If the message contains any request beyond the listed routine one, do not send. Draft instead and flag it."

The weekly rewrite review is the whole system

Your weekly review should cover the drafts you rewrote, because those show you exactly where the instructions are weak. Set aside a fixed half hour each week. Friday afternoon works for most people because the week's mail is still fresh.

Pull every draft you changed substantially before sending. Ignore typo fixes. For each one, ask a single question: **why did I rewrite this?** The answers usually fall into a few groups:

- **Wrong tone.** Too formal, too cheerful, too long. Fix the tone section of the instructions and add the rewritten version as a new example. - **Wrong facts.** It quoted an old price, a discontinued product, last season's hours. Your source material is stale. Update what the assistant reads from, not just the prompt. - **Promised something.** It offered a callback, a date, a discount. Add that to the "never promise" list, word for word. - **Missed the point.** It answered the question the customer asked and missed the problem they actually have. Add a line telling it to look for the problem behind the question, and flag this category for continued drafting. - **Wrong category.** The email was sorted wrong, so the draft was framed wrong. Sharpen the category definition with this sender as an example.

Make one or two instruction changes per week, not ten. If you change everything at once, you can't tell which edit fixed the problem and which one broke something else. Keep a dated log of the changes, even a single line each in a shared doc. In a month, that log will be the best record you have of how your business actually talks to customers.

As an illustration, suppose that in week one you rewrite most of the customer drafts because they open with "I hope this email finds you well." You add one line to the instructions: "Open with the answer, not a greeting." By week two, the opener is gone, and the remaining rewrites are about substance. Now you're fixing the problems that actually matter.

When to give it more freedom

Add one reply type to the auto-send list only after it has gone several consecutive weekly reviews with no meaningful rewrites. Calibrate the exact number for your volume. A shop that gets three hours questions a week needs more weeks of evidence than one that gets thirty. Whatever number you choose, write it down before you start so you're not deciding based on a good mood on a Friday.

Promote one type at a time. After a promotion, spot-check what actually went out for the next two weeks: open the sent folder, filter by that reply type, and read them. If anything there makes you wince, demote it back to draft that day. Demoting is cheap and it shouldn't feel like a failure. It means the review process is doing its job.

Some things never get promoted, regardless of the track record. Anything involving money, anything from an unhappy customer, and anything that creates an obligation for you stays in draft. A long streak of good drafts on those emails doesn't change how much damage a single bad one can do.

What this looks like as a weekly rhythm

The whole setup takes a short weekly routine once it's running:

- **Daily, a few minutes:** Skim the "Everything else" pile for misfiles. Send or edit the drafts waiting in "Customers" and "Money." - **Weekly, half an hour:** Review rewritten drafts, make one or two instruction changes, log them. - **Monthly, at 30/60/90 days:** Decide whether any reply type has earned auto-send, and spot-check anything you promoted last month.

To find out whether it's working, track one number of your own: how long you spend in your inbox per day, measured before you start and again after a month. If it hasn't dropped, look at the rewrite log. Usually the instructions are too vague, or one category is catching mail it shouldn't. Fix those two before you blame the tool.

Start with what you already have

Check the AI already built into your email or help desk before buying anything new. Most current mail and support tools offer suggested replies, smart labels, or a drafting assistant, and the draft only setting is usually the default. The work that makes any of these tools useful is the same: written category definitions, a written tone guide, a never-promise list, and a weekly review. That work carries over to whatever tool you end up using.

Your inbox is where customers decide whether you're paying attention. Keep the sending decision in your hands until the drafts have shown, in writing and week after week, that they deserve otherwise.

- write-down-your-process-so-ai-can-follow-it - measuring-time-saved-on-one-ai-workflow - use-the-ai-already-built-into-your-software-first

Need help shipping this for your team?

Ready to get started?

Tell us where work is slowing down. We will map the clearest first workflow and the path to ship it.

Request a Readiness Scan