AI recruiting outreach needs a human yes because every message goes out under your company's name to a person who may become a colleague, and a mistake repeated at scale costs more than the minutes approval takes. Approve before send means the agent finds the people and drafts the messages overnight, and a person reads them and switches them on once before anything reaches a candidate.
That is the rule in the 100Networks Workspace: nothing reaches a candidate until your team approves it. This post explains why we think it is the right default, what it looks like in practice, and how to make approval quick enough that nobody is tempted to skip it.
What goes wrong when AI outreach runs on autopilot?
The arguments for full autonomy are real. Agents are fast, they do not get tired, and the first message to a passive candidate is usually the biggest delay in the whole process. If the agent can write a decent message, why wait for a person?
Because the failures of unsupervised outreach are not about message quality. They are about judgement and context, which the agent has less of than your team does:
- The wrong person. A profile that was right two years ago, or a near-match that reads well but misses the one requirement the hiring manager cares about.
- The wrong claim. A message that implies a remote role that is not remote, a level the team has not approved, or a salary range nobody signed off.
- The wrong moment. A note to someone who works at your largest client, or a follow-up to a person who already said no on another channel.
- The double approach. Two teammates' sequences, or two roles, reaching the same engineer in the same week.
- The wrong voice. Copy that is grammatical and polite and nothing like how your company speaks.
Any one of these is small. At the volume an agent makes possible, a small error rate becomes dozens of people with a slightly worse view of your company. Senior engineering markets are small; people compare notes. A human read before the send catches most of these, because a person on your team knows things the agent does not.
What does "human in the loop" actually mean here?
The phrase gets used loosely. In a hiring team it should mean three specific checkpoints, each owned by a person:
- The brief. A person confirms what the role is before anyone is searched for. The agent drafts it; you correct it.
- The shortlist. The agent says how well someone fits; only a person shortlists them. In the Workspace the agent's view is in words (Strong fit, Good fit, Partial fit, Poor fit), and the decision (Shortlist, Maybe, Not a fit) is yours alone.
- The send. A person reads what will go out, to whom and when, and switches it on.
What it should not mean is approving every message one by one. That design fails in a predictable way: people approve in bulk without reading, and the check becomes theatre. One considered approval per sequence works better than fifty careless ones.
What should a good approval screen show?
Everything you need to say yes with confidence, on one screen. In the Workspace, switching a sequence on opens a view built from the real people in it:
- Each person, each step. "Email to their address on file." "Connection note, then the message once they accept." "WhatsApp, only if a phone number is on file."
- What will be skipped, and why. No email on file means that step is skipped for that person. A channel that is not connected yet says so before you switch on, not after.
- The timing. Each channel runs once, in order, with the wait between steps you set.
- Calls, honestly described. "AI call within calling hours. It says it is an AI assistant."
Then the rules that keep it safe after you have said yes:
- A reply stops the sequence for that person. People who replied, finished or bounced get nothing more.
- Editing live copy turns the sequence off. If anyone changes the words of an active sequence, it stops until someone switches it on again. Approval is for the copy you read, not whatever the copy became later.
- You can pause at any time. Anything not yet sent will not go.
For phone calls the same principle applies: a call list in the Workspace dials nobody until someone presses Start calling. We cover the calling side in detail in AI phone screening in India.
How does AI draft outreach in your team's voice?
Approval is only quick if the drafts are close to right. Three things get them there.
Context from the role. Outreach writes from the brief and the shortlist, not from a template. For a Senior Backend Engineer, Payments role, that means the message can talk about the actual system, the actual team and why it might be a step up for this person.
A reason for each person. Each message can carry a line on why this person fits, drawn from the evidence the Sourcer found in their own profile. "Your work on the reconciliation service" lands; "your impressive background" does not.
What your team has taught it. Things your team has written down about how you hire (for example, "we never mention salary in a first message" or "always say the role is hybrid, three days in the office") sit in the Workspace's knowledge and shape the drafts.
When a draft is not right, you tell Outreach what to change in plain words ("shorter, and lead with the team, not the company") or edit it yourself. You can preview each step against a real person on the list before you switch it on. When Outreach cannot do exactly what you asked, it says so in a note rather than quietly doing something else.
Doesn't approval slow everything down?
Less than you would think, because approval is not where the time goes.
On a typical role, the slow parts of outreach are finding the right people and writing something specific to each of them. Both happen overnight in the Workspace. What is left for the morning is reading a sequence and its people, which for a shortlist of twenty is a few minutes.
Compare that with the alternative. An autonomous agent that sends at 2 am saves you those minutes, and costs you the chance to notice that one of the twenty works at your client, two were approached last month by a colleague, and the third paragraph promises a level the hiring manager has not agreed to.
A checklist for approving an AI outreach sequence
Before you switch a sequence on, it is worth a quick pass:
| Check | What you are looking for |
|---|---|
| The people | Anyone at a client, partner or your own company; anyone approached recently |
| The claims | Location, work model, level and pay framing match what the role actually is |
| The voice | Would your hiring manager sign this? |
| The why-fit line | Specific to the person, from their own work |
| The channels | The order makes sense; nothing skipped that you expected to send |
| The call step | Says it is an AI assistant; calling hours are respected |
| The stop | A reply ends the sequence for that person |
Most sequences pass in a minute or two. The ones that do not are exactly the ones that justify the check.
What about Pilot and other "ask in plain words" assistants?
The same rule should apply to any assistant that can act on your behalf. Pilot, the assistant across the Workspace's agents, lets you ask for anything in plain words, such as "add the people I shortlisted today to the payments sequence". By default every action it proposes waits for your confirmation. Asking is instant; doing still needs a yes.
How this works in the 100Networks Workspace
Outreach is one agent on an AI hiring team (see what an AI hiring team is). It drafts email, LinkedIn and WhatsApp messages in your team's voice, from your own connected accounts, for the people you shortlisted, and holds them for approval. The Caller can be a step in the same sequence. Nothing is sent and nobody is called until someone on your team has read it and switched it on. For more on what agents should and should not decide, see what an AI recruiting agent is.
The Workspace is for companies' in-house teams, with access by request. See how Outreach and the rest of the team work on the 100Networks Workspace page.