How to Build a Human-in-the-Loop AI SDR for LinkedIn (Safe Automation)
· how-to
The number-one objection to AI SDR agents is trust: "I don't want a bot sending unsupervised DMs from my LinkedIn account." Item 12 is built around that concern — every LinkedIn action from agents, Playground, or MCP queues for human approval before send. This tutorial shows how to configure a human-in-the-loop AI SDR your team and legal counsel can stand behind.
Why human review is non-negotiable on LinkedIn
- Account safety — LinkedIn restricts accounts that blast generic or high-volume messages
- Brand voice — AI drafts need a human eye for tone, accuracy, and compliance
- Deal context — reps catch wrong-company or already-customer contacts before send
- Agency liability — client accounts need manager sign-off on outbound copy
Step 1: Configure what enters the review queue
In Item 12, all agent-generated connection requests, DMs, and follow-ups queue in Review by default. The same queue catches MCP and Playground writes — one approval surface for the whole workspace. Open Review each morning to batch-approve proven copy or edit one-offs before they send.
Step 2: Set static vs AI copy policies per step
- Lock connection notes as static templates when legal requires exact wording
- Allow AI prompt mode on follow-up DMs where personalization adds value
- Document approved prompt templates in a shared doc — reps paste into agent steps
- Reject and edit bad AI drafts in Review — the model learns from your team's edits over time via prompt refinement
Step 3: Team approval workflow
SDRs enroll contacts and AI drafts messages; managers review and approve in the shared Review queue. Item 12 Teams ($105/mo) adds seat permissions and team analytics — see which rep enrolled whom and what was approved. For agencies, switch orgs per client workspace so each client's review queue stays isolated.
Step 4: Rate limits and daily caps
Configure conservative daily caps when launching a new agent — start with 10–15 enrollments per day, measure accept and reply rates, then scale. Item 12 caps prevent runaway automation even if someone bulk-enrolls a list. Pair caps with stop-on-reply rules so active conversations never get automated bumps.
Step 5: Audit trail and accountability
- Mirrored inbox — full thread history on each contact record
- Workflow run logs — which agent step generated each pending action
- Review history — approved vs rejected drafts with timestamps
- Pipeline moves — tie replies to stage changes for manager reporting
Next in series
Ready for power-user setup or full auto-prospecting? Continue with How to Build a LinkedIn AI SDR Agent with Claude and Item 12 MCP and How to Build an AI SDR That Finds and Messages LinkedIn Prospects Automatically.
Can I turn off human review in Item 12?
Review is the default safety model. Disabling approval removes your guardrail against bad AI copy and account restrictions — most teams keep it on.
Do MCP-generated sends also hit the review queue?
Yes — workflow_create and contact_enroll_workflow produce the same queued actions as the visual builder. MCP does not bypass review.
How do agencies handle review per client?
One Item 12 org per client LinkedIn account. Switch orgs in the app; each workspace has its own review queue and agent config.