What can go wrong when drafts send themselves
Picture a small service business that turns on an AI tool to write follow-up emails after estimates and appointments. For two weeks it feels like a gift. Then a draft goes out on its own telling a customer their work is "scheduled for Tuesday," when Tuesday was only a suggestion in an internal note. Another one answers a Spanish-speaking customer in English. A third thanks someone for a visit that was canceled. Each message was reasonable on its face. None of them was checked by a person before it left.
That scene is an original illustration, not a TwoQM client and not a measured result. The lesson is simple: letting AI draft is useful; letting AI send is a different decision, and most small businesses are better off keeping a person between the two.
This article is not about who gets a follow-up or when. It covers one control: a send-approval step that holds every AI-drafted message until a named person on your team reviews it.
AI drafts are usually fluent, which is exactly why mistakes slip through. The common problems:
Facts that aren't confirmed. A date, amount, or scope detail pulled from a note that was only a guess.
Promises nobody approved. A discount, a refund, a "we'll take care of it," or a callback time.
Wrong person or wrong channel. A message to a contact who asked not to be contacted, or a text to someone who only wants email.
Wrong language or tone. English to a customer who writes in Spanish, or a cheerful note after a complaint.
Stale context. A draft written yesterday that no longer fits after today's phone call.
None of these needs a complicated fix. They need a person to look before anything leaves.
What a send-approval step looks like
The core rule: AI drafts land in a review queue, never in the outbox. Nothing sends until someone with the right role approves it. There is no "auto-approve if nobody responds" setting.
Each draft in the queue should show:
The draft itself, editable in place.
What it was based on. The record, note, or message the AI used, side by side, so the reviewer can check facts quickly.
Highlighted details. Dates, amounts, names, and commitments marked so they are easy to verify against the record.
Recipient, channel, and language, plus any contact limits on that customer.
Simple actions. Approve and send, edit then send, send back for a new draft, or reject with a short reason.
Decide who can approve what
Not every follow-up carries the same risk. Write a short rule list your team agrees on:
Routine confirmations and thank-you notes can be approved by front-desk or office staff.
Anything that mentions money, schedule commitments, scope changes, or a complaint goes to a manager or the owner.
Some topics are never drafted by AI at all, like sensitive disputes or anything your industry's rules restrict.
Your rules, not the tool's defaults, decide where each draft goes for review.
Keep drafts fresh, and keep the history
Drafts should expire. If a draft sits in the queue past the point your team sets, or the customer record changes after it was written, it gets marked stale and redrafted instead of sent.
Keep a record of who approved each message, when, and what changed between the AI draft and what was sent. That history is useful for answering a customer question later, and the edits show where your instructions to the AI need work. It is a quality record, not a way to grade staff.
Two related pieces: cross-sell flags a person must review before outreach covers review before an idea becomes a contact, and duplicate contact merge rules before marketing sends covers cleaning the list a message goes to.
Where TwoQM fits
AI integrations can be set up so drafting and sending are separate steps, with the approval queue built in rather than bolted on. A CRM can hold contact limits, language preferences, and the approval history next to each customer. Business automation can remind reviewers about waiting drafts and expire stale ones. Where your email, texting, or scheduling tools allow it, connected systems work can pull confirmed facts into the draft and send only after approval; what is possible is confirmed during scoping.
Where AI helps, and where it stops
AI can write a first draft from your notes, match the customer's language, suggest a shorter version, and point out details it could not confirm.
AI does not approve its own drafts, send messages, or decide that a commitment is acceptable. A person on your team makes every send decision.
Where to start this week
Pick one kind of follow-up your team writes often. Write down who may approve it and which details must be checked. Run AI drafts for that message type into a review queue for a week, approve or edit each one by hand, and note the edits. Then decide whether to add a second message type.
TwoQM builds AI integrations, CRM workflows, and automation for small businesses across industries. If you want AI to help write follow-ups without letting it send them, TwoQM can scope a draft-review workflow around the tools you already use. Pricing explains how business apps, systems, and custom software are scoped.
Describe Your AI Draft Review Workflow · Explore TwoQM AI integrations