Full autopilot is fast until it isn’t. Full manual is safe until you burn out. The useful middle has a name, and a method.
Human-in-the-loop outreach means AI does the research and the first draft, and a person approves anything that goes out under their name. It’s how you get most of the speed of automation without handing your reputation to a script.
The idea is simple. Doing it well is less so. Approve too much and you’re back to doing everything by hand. Approve too little and you’re rubber-stamping a machine. Here’s the workflow we built Manet around, and that we’d recommend whatever tool you use.
Why keep a human in the loop at all?
Your name is on it. One tone-deaf comment on a grieving founder’s post can cost you more than a month of good outreach earns.
Context is hard. AI is good at reading evidence. It’s less good at knowing that this particular prospect is your cofounder’s ex-colleague, or that the post was sarcastic.
LinkedIn notices patterns. High volumes of identical automated actions are what run into platform limits. Reviewed, varied, human-paced activity looks like what it is.
The edits are the training data. Every time you change a draft, you’re teaching the system how you actually write. Skip review and you skip the learning.
We make the longer argument in the case against fully autonomous LinkedIn outreach.
What to automate and what to approve
The rule of thumb: automate what’s reversible and invisible, approve what’s public or personal.
Action | Automate | Approve | Why |
|---|---|---|---|
Finding people who fit the brief | Yes | Tedious, invisible, easy to correct | |
Checking evidence for each match | Yes | Spot-check the reasons, not every row | |
Choosing who to contact first | Suggest | Glance | Priorities are a judgment call |
Liking a post | Optional | Low risk, low value. Many skip it | |
Commenting on a post | Draft | Yes | Public and permanent |
Connection request with a note | Draft | Yes | First impression |
First message after connecting | Draft | Yes | Sets the tone of the relationship |
Follow-ups | Draft | Yes | Easy to get wrong if they replied elsewhere |
Replies to a real conversation | Suggest | Always you | This is the relationship |
The workflow, step by step
1. Write a brief that a colleague could act on
Who, where, why now, and who to exclude. “Heads of sales at B2B SaaS companies with 50 to 300 people who posted about hiring SDRs in the last 30 days, excluding current customers.” The clearer the brief, the less you’ll reject later. Our 25 AI prospecting prompts are a good starting point.
2. Let AI research, with evidence attached
The research step should never hand you a bare name. It should hand you the reason: the post, the job listing, the funding round. Evidence is what makes a fast approval possible, because you’re checking a claim rather than redoing the research.
3. Queue drafts, not actions
Everything public lands in a queue as a draft with its context next to it. Nothing goes out by default.
4. Review in short, fixed windows
Two ten-minute windows a day beat a constant drip of notifications. Batch the queue, work through it, close it. With Manet, Aria can send the queue to WhatsApp, so a window can be the five minutes before a meeting.
5. Approve, edit, or skip, and say why when you skip
Three buttons, and a one-word reason on skips (“wrong role,” “too salesy,” “already know them”). Reasons tell the system what to change in the brief or the drafting.
6. Let the edits teach the system
If you keep cutting the first sentence, the drafts should stop writing it. If you keep skipping agency founders, the brief should exclude them. This is the step most tools skip, and it’s the one that makes week six better than week one. Manet learns from your edits so that it drafts and decides more like you over time.
A ten-second review checklist
You don’t need to deliberate over each draft. Ask four questions:
Is this the right person? Does the evidence actually support the match?
Is the moment real? Is the signal recent and about their current company?
Would I say this? Out loud, at an event, to their face.
Is the ask small? Could they answer in a line?
Four yeses: send. One no: edit or skip.
When to loosen the reins
Approval isn’t forever for every action. Autonomy should be earned per action type. If you’ve approved fifty profile-view or like actions without a single change, it’s reasonable to let those run on their own. Comments, connection notes and messages are different: they’re public, personal and permanent, so we’d keep a human on them even when the drafts are consistently good. A quick glance costs seconds. A bad message costs a relationship.
What to measure
Metric | What it tells you | Healthy direction |
|---|---|---|
Approval rate | How well the brief and drafts match your judgment | Rising over the first weeks |
Edit rate | How much drafts still sound unlike you | Falling as the system learns |
Time per approval | Whether review is sustainable | A few seconds per item |
Reply rate | Whether the whole thing works | Higher than your old baseline |
Skip reasons | What to fix in the brief | Fewer repeats of the same reason |
Three ways approval goes wrong
Rubber-stamping. If you approve everything in a hurry, you’ve built autopilot with extra steps. Smaller queues help.
Queues that are too big. Forty drafts a day is a chore. Ten to twenty good ones is a habit.
No evidence on the screen. If you have to open three tabs to check a draft, review will become the bottleneck. The context should come with the draft.
For how this compares with fully autonomous tools, see what is an AI SDR?
Frequently asked questions
What is human-in-the-loop AI outreach?
An outreach workflow where AI researches prospects and drafts actions, and a person approves anything that goes out under their name, such as comments, connection requests and messages.
Is human approval slower than full automation?
Slightly, but not by much if the drafts come with evidence. Two short review windows a day are usually enough, and the quality gain tends to outweigh the minutes spent.
Which LinkedIn actions should always be approved?
Anything public or personal: comments, connection notes, first messages, follow-ups and replies. Low-risk, invisible actions like research can run automatically.
Can I approve outreach from my phone?
With Manet, yes. Aria can send progress updates and approval requests on WhatsApp so you can approve, edit or skip between meetings.
Does the AI learn from my approvals?
In Manet, yes. Your edits and skips shape future drafts and decisions, so it writes and chooses more like you over time.
When is it okay to fully automate an action?
When the action is low risk, reversible and you've consistently approved it without changes. Keep public and personal actions reviewed.
Want the research and drafting done, with the final say left to you? Try Manet free.
[
ready to start?
]
