When to graduate a client from supervised send to autopilot: an agency checklist

Supervised send is not a personality trait. It is a risk control.

Autopilot is not a maturity trophy. It is a mode you earn per client workspace when the evidence says the human gate is mostly rubber-stamping good work.

IBM's human-in-the-loop framing is useful here: keep humans where ambiguity, ethics, or edge cases still bite; remove them where the loop is stable. Agencies need a checklist, not a vibe.

Default posture

New client workspaces start supervised:

• Human reviews outbound before first send windows go live
• Voice brief, ICP, and do-not-say lists are still settling
• Domains may still be warming
• Client stakeholders are still learning what "good" looks like

GTM supports supervised vs autopilot sending for agency teams. Graduation should be explicit, reversible, and logged.

Graduation checklist (all must pass)

Only move a client workspace to autopilot when every row is true:

1. Identity and authentication are boring

• SPF, DKIM, and DMARC verified for each sending domain
• No open DNS tickets on the client side
• Seed tests show acceptable inbox placement for the domains in use

MailReach and ActiveCampaign both treat authentication and consistent volume as table stakes. Autopilot on a half-warmed domain is just automated spam-folder training.

2. Compliance gate is quiet for the right reasons

• Unsubscribe URL renders on every active step
• Verdict failure rate is at baseline for 10+ consecutive send windows
• No unresolved false-positive incident in the last 14 days
• Suppression sync from CRM is tested with a seed address

3. Voice and ICP are versioned and stable

• Active voice brief has a version pin and client approver
• ICP has not hard-shifted in the last two weeks (or a new brief version was cut and sequences re-bound)
• Do-not-say list reviewed with the client in the last 30 days

4. Write quality is boring in review

• Supervised reject rate is low and falling (operators mostly approve)
• Reject reasons are logged (tone, claim, ICP mismatch, compliance)
• Research notes are separated from sendable copy — no note bleed in the last sample of 25 sends
• Personalization claims carry provenance

5. Volume and caps are intentional

• Per-client daily send caps documented
• LinkedIn connection caps (if used) do not starve email
• Send windows match prospect-local agreements
• No surprise backlog dumps planned for "catch-up week"

ProspectingManual's multi-client ops stress linear volume and isolation. Autopilot should inherit those limits, not override them.

6. Reply and booking path is staffed

• Shared-inbox ownership and after-hours SLA are named
• Sequence stops on reply is verified in a live test
• Calendar routing / AE handoff packet exists
• Client knows who owns a positive reply in the first business hour

IBM notes that HITL matters most when stakes and ambiguity are high. Positive replies and edge-case objections stay human even after send autopilot.

7. Client consent is written

• Client stakeholder signs the mode change (email or ticket is enough)
• Rollback trigger is defined (complaint spike, block-rate spike, brand incident)
• Weekly scorecard still ships (booked + held, credit burn units, exceptions)

How to graduate without theater

1. Run a shadow autopilot week: system would have sent; humans still approve; compare outcomes
2. Flip one campaign first, not the whole workspace
3. Keep random human sampling (e.g. 5–10% of sends) for two weeks
4. Keep always-human paths: first touch on a brand-new offer, executive titles if the client requires it, any sequence after an ICP hard shift

AWS's tenant pause/resume pattern is the right mental model for rollback: if reputation findings spike, disable sending for that tenant quickly without touching other clients.

What is not a graduation signal

• "We are behind on volume"
• "The client asked why humans are slow"
• "The model feels better this week"
• "We need to hit a booked-call goal" (GTM's ~$100/booked call figure is an optimization goal, not a reason to skip QA)

Re-enter supervised mode when

• Complaint or block-rate spikes
• Domain authentication breaks
• Voice brief or ICP hard-shifts
• New offer or regulated claim set ships
• Client stakeholder changes and wants a reset

Autopilot is a privilege the workspace holds, not a permanent setting.

What GTM expects you to keep honest

GTM runs find → qualify → write → send/reply/book in client-isolated workspaces with a compliance gate that stays on. Supervised and autopilot are operating modes around that motion. Graduate with a checklist, keep humans on replies and edge cases, and roll back without drama.

If you are building that motion now, join the waitlist.