FlowsBook a callSign inCreate free accountSign up
AI employees and automation

An AI employee that books nothing but drafts everything

Let the AI do the research and the writing, and keep every send, booking and promise with a person. How to set up a drafting-only AI employee and review its work fast.

Part of the guide AI employees

By the xMap team · · 6 min read

A drafting-only AI employee does all the preparation and none of the committing. It finds the businesses, reads what the data says about each, writes the email, the note or the report, and then stops and hands everything to a person, who edits, approves or bins it. It never sends on its own, never books a meeting, never promises a price. In xMap Flows you get this by setting the employee to Ask me first and putting an Approval step, item by item, before anything leaves.

This is for managers who want the hours an AI saves but are not ready to let it speak for the company: sales leaders with new reps, agencies writing to their clients' customers, regulated teams in health, finance or law, and anyone whose brand rests on a careful tone. It covers why drafting is a sensible place to start, how to set it up, what the reviewer sees, a made-up week in Cape Town, the mistakes that make a drafter more work than help, and when to let it do more.

Why draft first

Most of the time in outreach, reporting and client updates goes into the preparation: finding who to write to, checking what they do, gathering a reason to write, and putting words on the page. The last step, pressing send, takes seconds but carries all the risk. A drafting employee takes the slow part and leaves the risky part with a person.

  • Mistakes stay private. A wrong name or a clumsy sentence is caught on the review page, not in a customer's inbox.
  • Your voice stays yours. The reviewer fixes the tone once or twice, and the instructions get better from those edits.
  • Nothing is booked or promised. Meetings, prices and commitments stay with people who can stand behind them.
  • Trust is earned on the record. After a few weeks of reviewing, you know exactly how good the drafts are before you give the employee more freedom.

How to set it up

  1. Build the finding part as usual: a trigger, an area and a data block such as Nearby places or Job postings, with Only new so each item is drafted once.
  2. Add xMap LLM to write the drafts: tell it, in plain words, who you are, who you write to and why, and ask for named fields such as email_subject and email_body. Tell it to use only the facts in the data and to say so when it lacks one.
  3. Add Approval, set to approve each item on its own, and list email_subject and email_body as fields the approver can change.
  4. Put the sending block after it: Outreach for a first message with follow-ups, Email, or an app such as Slack. Only what is approved, with the approver's own words, goes on.
  5. On the employee's card, choose how much it does alone: Ask me first. Its own approval step then sits before every send.

What the reviewer sees

The approver gets a notification and an email with a link to a review page. Each item is a row with a tick box, the facts the draft was based on, and the draft itself in editable boxes. The reviewer ticks the drafts that are right, rewrites the ones that are nearly right, and leaves the rest unticked. Approve the ticked ones, and only those carry on, with the edited words. Nobody answering means nothing goes, unless you set the step to carry on after a waiting time.

A worked example (made up)

Say a small agency in Cape Town writes to newly opened restaurants on behalf of a linen supplier. Every weekday at 07:00 its AI employee finds restaurants that are new to its list, reads each one's category, rating, website and neighbourhood, and drafts a three-sentence email that names something specific about the place. By 07:05 the account manager has twelve drafts on the review page.

Over coffee she ticks eight, rewrites the opening of two, and leaves two that misread a hotel kitchen as a cafe. The ten approved drafts go out through Outreach, which follows up twice and stops if someone replies. No meeting is booked by the AI; replies come to her, and she books the calls herself. These figures are invented to show the rhythm, not results from a real agency.

Mistakes that make a drafter more work than help

  • Vague instructions. "Write a nice email" gives generic drafts. Say who you are, what you offer, the length, the tone and one fact to mention.
  • Letting it invent. Tell it to use only the fields it receives. A draft that guesses an owner's name is worse than one that says "Hello".
  • Approving in bulk. If every draft is ticked without reading, the step is theatre. Read, edit and decline honestly.
  • Never improving the prompt. When you make the same edit three times, put it in the instructions.
  • Hiding the drafts. Send the review link to the person who owns the relationship, not to a shared inbox nobody checks.

When to let it do more

After a few weeks, look at how many drafts went out unchanged. If almost all of them do, you can move the employee to Act and tell me for that kind of message, keeping Ask me first for anything sensitive. The setting is per employee, so a cautious drafter and a trusted sender can work side by side. The Check the work block adds a second, automatic check against your rules before anything reaches a person or a customer.

The drafting employee in one view

Its blocks, in order: Schedule on weekdays, an Area of interest, Nearby places, Only new, Sort and limit to the best twelve, xMap LLM writing email_subject and email_body, Approval item by item with both fields editable, then Outreach. What lands in the inbox each morning is one link to a page of ready drafts, each beside the facts it used; what leaves is only what a person approved. The templates An intro and two follow-ups to new dental clinics and Write to companies hiring, with two follow-ups work this way, and New in town: a weekly column draft drafts an article instead of emails. More in when to make an AI employee ask before it sends.

Questions

Can the AI employee book meetings for me?

Not in this set-up. It drafts the messages; replies come to you and you book the calls. Outreach forwards each reply and stops the sequence.

Can I change a draft before it is sent?

Yes. With Approval set to each item on its own, the fields you choose, such as the subject and body, are editable on the review page, and your words are what get sent.

What happens if nobody approves?

Nothing is sent. The step can also be set to carry on after a waiting time, but for a drafting employee leave it to stop.

Does Ask me first slow everything down?

Only by the time it takes to review. The research and writing are done before the reviewer opens the page, so a daily batch takes minutes.

Templates that do this

Read next

More on ai employees and automation · Every article · Templates

Put it to work

An AI employee in xMap Flows does this on live data, on its own timetable. Start from a template or describe it in plain words.