Human-in-the-loop AI for candidate outreach

Three places the human can sit

"Human in the loop" is not one design. In outreach it resolves to three distinct patterns, and choosing between them is the whole decision.

• Approve each message. Slowest, highest control. Correct while the model is uncalibrated or the audience is senior.
• Approve the template and the audience, sample the sends. Practical steady state for high-volume sourcing.
• Review after the fact. Only defensible for internal drafts that never leave the company.

What gets logged, and why

The log is the control. For every message the system should retain the source data the draft was built from, the template version, the human who approved it, the timestamp, and the final text as sent. Without the final text as sent, you cannot answer the only question that matters after a complaint: what did this person actually receive?

Reply handling deserves the same treatment. If an agent classifies a reply as "not interested" and closes a thread, that classification is a decision with consequences and needs to be reversible, visible, and attributable.

Failure modes worth monitoring

Four failure modes recur, and each has a cheap monitor.

• Personalisation drift: the draft asserts something about the candidate that is not in the source data. Monitor by sampling drafts against their source records.
• Duplicate contact: the same person approached from two roles in the same week. Monitor with a contact-frequency cap enforced outside the agent.
• Silent template decay: response rates fall while volume holds. Monitor per template version, not in aggregate.
• Approval theatre: approvals recorded faster than a human could read. Monitor approval dwell time.

Volume is the risk multiplier

Every failure above is survivable at ten messages a day and reputational at a thousand. This is why we treat the sending rate as a governed parameter rather than a growth lever, and why the approval pattern should tighten, not loosen, as volume rises.

It also explains why the workspace matters as much as the model. Consent state, prior contact history, and reply threads have to live in one reviewed system, or the controls above cannot be enforced by anything except discipline.