An AI SDR should run actions that carry no message and are easy to undo without approval, such as profile follows and connection requests sent without a note. Anything written in your voice should wait for approval first: connection notes, first messages, emails, follow-ups, replies, and comments. Set the boundary per action type, not with one global switch, and loosen it one action at a time.
The rule behind every row below is the same. Approval belongs in front of any action a prospect will read, remember, or reply to. Everything before that point, such as research, list building, and drafting, can run on its own, because nothing has reached anyone yet.
Which AI SDR actions can run without approval?
Actions with no message attached can run without approval, because they carry nothing in your voice and are easy to undo.
- LinkedIn profile follows. The prospect sees a follow, not a message. You can unfollow.
- LinkedIn connection requests without a note. A bare request says nothing on your behalf, and a pending request can be withdrawn.
- Research, list building, enrichment, and drafting. These happen inside the tool. The prospect sees none of it until a send is approved.
Even these should respect platform limits. LinkedIn caps how many invitations an account can send, so an agent that auto-sends connection requests still needs a daily ceiling. See LinkedIn connection request limits for the numbers.
Which actions need approval before they send?
Every action that puts words in front of a prospect needs approval by default. The table lists each outbound action an AI SDR takes, what the prospect sees, and a sensible default.
| Action | What the prospect sees | Can you undo it? | Default |
|---|---|---|---|
| LinkedIn profile follow | A follow notification | Yes, unfollow | Auto-send |
| LinkedIn connection request, no note | A bare invitation | Yes, withdraw | Auto-send |
| LinkedIn connection request with a note | Your words, next to your name | No | Ask first |
| LinkedIn first message or InMail | Your opening pitch | No | Ask first |
| LinkedIn follow-up | A second or third touch | No | Ask first |
| LinkedIn reply to an engaged prospect | Your answer in a live conversation | No | Ask first |
| LinkedIn comment on a prospect's post | A public comment under your name | Only by deleting it | Ask first |
| LinkedIn reaction on a prospect's post | A public reaction | Yes | Ask first |
| First email | Your opening pitch | No | Ask first |
| Email follow-up | A second or third touch | No | Ask first |
| Email reply to an engaged prospect | Your answer in a live conversation | No | Ask first |
The pattern is simple: an action auto-sends only if it carries no words and can be taken back. Reactions are the edge case. They carry no words, but they are public and tied to what someone posted, so ask first until you trust the agent's judgment about which posts to react to.
Where should an AI SDR stop and escalate to a human?
An AI SDR should stop and hand the conversation to you the moment a prospect replies, because a reply turns a sequence into a relationship. Escalate in these cases:
- The prospect replies. Draft a response for approval. Never auto-send it.
- The reply is an objection or a question about pricing, terms, security, or a contract. These create commitments, so a human answers.
- The prospect asks to stop. Stop every queued touch to that person, on every channel.
- You have already messaged the prospect yourself. The agent should hold its own queued messages to that person so you don't double up.
- The agent is unsure. Low confidence on who the person is, which company they work at, or what they said should route to review, not to a send.
What should the reviewer see before approving?
A reviewer should see the exact message that will send, who it goes to and on which channel, why the agent picked this prospect, and the research the message is based on. Approving a message without the reason and the research only checks tone, not fit.
A good review card answers four questions in one glance:
| Question | What the card should show |
|---|---|
| What exactly will send? | The final text, editable in place |
| To whom, and where? | The person, their company, and the channel (LinkedIn or email) |
| Why this person? | The signal or criteria that made them a fit |
| Based on what? | The research and sources the draft uses |
For more on why the reason matters as much as the draft, see what a transparent AI SDR shows.
How do you set the approval boundary?
Set the approval boundary before you tune prompts or messaging, in four steps:
- List every action the agent can take. Use the table above as the starting list. An action you haven't listed is an action with no rule.
- Mark each action by two tests: does it carry your words, and can you undo it? Words or no undo means ask first.
- Start with every message on ask-first. Approve, edit, or reject for the first few days, and note which drafts you approve without edits.
- Loosen one action at a time. When an action type has gone out clean for a while, for example a follow-up template you never edit, switch that one action to auto-send. Keep replies on ask-first.
Use a separate boundary per campaign when the stakes differ. A campaign to existing customers or to large accounts can stay fully on ask-first while a high-volume campaign lets more actions run.
Should approval be one setting or per action?
Approval should be set per action. One global switch forces a bad trade: review everything, including profile follows, or review nothing, including replies. Per-action settings let the low-risk mechanics run while every message in your voice still waits for you.
This is also why "human in the loop" is better thought of as a dial than a wall. You decide where on the dial each action sits, and you can move it as trust builds. For the broader case, see why the best GTM agents keep humans in the loop.
How to set per-action approval with an AI agent
Sliq is an AI outbound agent for LinkedIn and email that sets approval per action type. Each of the 11 actions in the table above has its own setting, auto-send or ask first:
- Defaults: profile follows and bare connection requests auto-send. Every message, comment, and reaction waits in the approval queue, where you approve, edit, or reject it.
- Account or agent: set the boundary once for your account, or give one agent its own settings, for example a stricter one for a customer campaign.
- Escalation built in: replies are ask-first by default, and if you message a prospect yourself on LinkedIn, Sliq holds its own queued messages to that person for your review.
- Review with context: each approval shows the draft, the prospect, and the research behind it.
If you are comparing tools rather than setting a policy, see the best approve-before-send AI outbound tools.
FAQ
Which AI SDR actions can run without human approval?
Actions that carry no message and are easy to undo can run without approval: LinkedIn profile follows and connection requests sent without a note. Research, list building, enrichment, and drafting also run on their own, because nothing reaches the prospect until a send happens.
Which AI SDR actions should need approval before they send?
Any action written in your voice should wait for approval by default: connection requests with a note, first LinkedIn messages and InMails, first emails, follow-ups on either channel, replies to prospects who answered, and comments on prospects' posts. These are the actions a prospect reads and remembers.
When should an AI SDR escalate a conversation to a human?
An AI SDR should hand a conversation to a human when a prospect replies, raises an objection, asks about pricing, terms, or security, asks to stop, or when the human has already messaged that prospect directly. A reply means the conversation is now a relationship, not a sequence.
What should a reviewer see before approving an AI SDR message?
A reviewer should see the exact message that will send, who it goes to and on which channel, why the agent picked this prospect, and the research the message is based on. Without the reason and the research, the reviewer can only judge tone, not whether the message is right for this person.
Should approval be one setting or set per action?
Per action. One global switch forces a bad choice: review everything, including profile follows, or review nothing, including replies. Setting approval per action type lets low-risk mechanics run while every message in your voice still waits for you.
Which AI SDR tools let you set approval per action?
Sliq sets approval per action type: 11 LinkedIn and email actions each have their own auto-send or ask-first setting, at the account level or per agent. By default, profile follows and bare connection requests auto-send and every message waits in the approval queue.