When to make an AI employee ask before it sends

Put a person between an AI employee and the send button when the message leaves your company, when it commits you to something, or when one wrong item would be expensive. An approval step stops the run at that point; the people you choose get a notice and an email, look at the items, and approve or decline. Everything before the step has already run, so they judge real results, and nothing after it happens until someone answers.

This article explains when approval earns its delay and when it only slows things down, the choices the Approval step offers, how to approve item by item and rewrite AI drafts, and what happens when nobody answers.

Four situations where approval pays for itself

• First contact with a stranger. An introduction to a business owner carries your name. A person reading it catches the wrong first name, the closed shop, the cafe that is really a chain's flagship.
• Anything that changes a plan. A rain alert that moves an outdoor event indoors should be confirmed by whoever owns the event.
• Lists that go to a client. An agency's monthly competitor map is checked by the account manager before the client sees it.
• The first weeks of a new employee. Until you trust its output, approve its sends, then loosen.

And where it does not: daily internal alerts that lose their value if they arrive late, reports read by your own team, and anything a conversation is waiting on, such as a chat assistant's replies.

The choices on an Approval step

The Approval step's settings in xMap Flows.

Approving item by item

With Each item on its own, the approval page lists the items with a tick box each, up to 50 in one approval. The approver ticks the ones that may go and presses Approve the ticked; the rest stop there. A batch of twenty introductions can become fourteen good ones in a minute, without anyone editing the flow.

Rewriting a drafted message

When an AI step writes a subject and a body for each item, name those fields in The approver can change. On the approval page each item then shows them as boxes. The approver fixes a phrase, shortens a paragraph or corrects a name, and the message that is sent is the one they approved, word for word. The step records that the items were changed.

A worked example: a recruiter in Toronto

Say a recruitment firm in Toronto has an AI employee that finds companies hiring for warehouse roles, looks up the hiring manager at each, and drafts a short note offering candidates. The figures are made up. On Tuesday it finds 18 companies; xMap AI drafts 18 notes. The approval request reaches the team lead on Slack. She unticks three companies the firm already works with, rewrites the opening line of two notes that read stiffly, and approves fifteen. Outreach sends them during working hours in each recipient's time zone, follows up twice a few days apart, and stops for anyone who replies.

1. Job postings finds companies hiring for the role around the city.
2. Apollo, on the firm's own account, finds the person to write to at each.
3. xMap LLM drafts a subject and a short note for each.
4. Approval, each item on its own, with the subject and body open to change.
5. Outreach sends the first message and up to two follow-ups.

The write to companies hiring and full outbound engine templates follow this shape.

Where approvals reach people

The people asked get a notification in the app and an email. If you add Slack, Telegram or WhatsApp, the request arrives there too, with Approve and Decline. Approving several items one by one, or rewriting drafts, happens on the approval page, which the message links to. Organizations with strict sign-in rules answer in the app only.

What a declined or unanswered run looks like

A declined run stops at the step. That is recorded as stopped, not failed, so it does not raise a failure alert or ask to be fixed. Who answered, when and with what note stays in the run's history. If nobody answers in time, the run stops or carries on as the step says; both are recorded the same way.

The quick way: Ask me first

You do not have to add the step by hand. Setting an AI employee to Ask me first puts an approval before each of its sending steps, item by item, with drafted fields open to change. See how much an AI employee should do on its own for choosing between that and its looser settings.