Why approval, not autopilot
An outreach agent acts under your name. When it's wrong, the mistake is public and hard to undo: a message to the wrong person, a detail that isn't true, a reply in the wrong tone. Approval turns those into drafts you fix instead of messages you apologise for, and it trains the agent on your voice with every edit. Read do AI SDRs work for the failures it prevents.
Classify actions by consequence
The simplest policy sorts every action an agent can take into four kinds:
| Kind | Examples | Approval? |
|---|---|---|
| Read | Researching a lead, reading your calendar, summarising a thread | No |
| Reversible | Drafting a message, creating a task, adding a CRM note | No, but visible |
| External | Sending an email or DM, a connection request, a post, a reply, a booking proposal | Yes, every time |
| Destructive | Deleting records, cancelling meetings, removing contacts | Yes, every time |
Two rules make it hold up: an action can't lower its own risk level, and approvals never expire into automatic sending.
Some teams give an agent a standing go-ahead for one kind of external action instead, with hard limits on who it may contact, how many a day and on which channels. That's how Siela's AI SDR works once you start it, while its social agents let you choose whether each post waits for you.
What a good approval shows
- The exact content, not a summary.
- Who it reaches, and how many people.
- When it goes, and on which channel and account.
- What happens next, such as follow-ups or a booking if they say yes.
- Why, such as the research or signal behind the message.
An example approval card
| Field | Example |
|---|---|
| Action | Send the follow-up to Priya |
| To | priya@northwind.co (1 person) |
| Content | The full email, as it will be sent |
| Also | Moves the deal to Proposal; creates 2 tasks |
| Why | Priya asked for Q1 pricing by Friday on Tuesday's call |
| Choices | Approve, or reject with a note |
Keeping approval fast
Draft ahead
The agent does the research and writing before you look, so approving takes seconds.
Approve where you already are
Phone, chat apps, the desktop. Approval that needs a separate login gets skipped.
Approve a sequence once, then each send
Approve the plan (who, what, how many steps), then confirm each message as it comes up.
Reject with a note
"Too formal" or "don't mention pricing" fixes this draft and the next ones.
Keep a record
Every action, approval and decision should be written to a log you can read: who approved what, when, and what was sent. It answers "did we send that?" in seconds, and it's how you spot patterns in what gets rejected. In Siela, every tool call and decision is written to an audit log.
What to relax first, and when
If you ever loosen review, do it by consequence, not by convenience. Internal notes and drafts can be visible without approval from day one. Anything that reaches a customer, a prospect or the public should stay behind approval until you have weeks of drafts approved unchanged, and even then only for low-risk, repeatable messages to existing contacts. First messages to new people, posts, and anything destructive are worth approving indefinitely.
Common objections
- "It slows everything down." Approving a prepared draft takes seconds; fixing a wrong message takes much longer.
- "I'll just trust it once it's good." Many teams loosen review for low-risk steps over time. The consequence-based classes above tell you which ones are safe to relax.
- "Nobody has time to review." Then the volume is too high. Fewer, better messages beat many unchecked ones.
For the same principles applied to people, see approvals that don't slow everyone down.
Frequently asked questions
What is human-in-the-loop AI outreach?
Outreach where an AI agent researches and drafts, but a person approves each message, request, post or booking before it's sent.
What should an AI agent need approval for?
Anything that leaves your business or can't easily be undone: emails, DMs, connection requests, posts, replies, booking proposals and deletions.
Does human approval slow AI outreach down?
Very little if the agent drafts ahead and you can approve in one tap where you already work. It also cuts the time spent fixing mistakes.
Should AI approvals expire?
They shouldn't expire into automatic sending. Silence should never count as a yes.
How does Siela handle approvals?
The social agents can show you every post and reply first, or post without asking, as you choose. The AI SDR sends on its own once you start it, within the daily limits, channels and exclusions you set, or saves leads without sending if you choose Find leads only. Every decision is logged.