One preview, then go

Supervised mode is gone. Every campaign now sends on its own, on email and on LinkedIn, from day one. You can still read any message before it goes, edit it or stop it, and ask for changes in plain words; every later message in the workspace follows them. Since 27 October 2026, leads come from lead searches with a personal email, a business email, a phone number and a LinkedIn URL already attached, so there is no separate reveal step and no heat threshold before a sourced lead hears from you, and no keyword or signal is watched on their behalf. The rest of this post describes the design as it stood in August.

Per-message approval sounds safe and works badly. By the fortieth draft you are not reviewing, you are clicking, and a control that everyone clicks through is not a control. So GTM asks for one approval that actually gets attention: the ICP and the first three messages, on one screen, before anything sends.

After that, each campaign carries its own mode. Autopilot, or Supervised. Per campaign, not per account, so a proven motion runs hands-off while a new one earns trust on review.

In approval mode, the queue shows the exact words that will go out, plus the plan around them: how the first touch leaves, how many follow-ups on what cadence, what stops them, and where the lead lands if nothing comes back. You are approving a sequence, not a sentence.

There are three decisions, not two. Approve covers the sequence. Request changes sends the draft back with your note, and it returns rewritten. Remove takes a categorized reason. Wrong ICP, not qualified, bad timing, and that judgement trains the same telemetry the sourcing gate is tuned against. Every human overrule makes the next list better.

Approving is also the moment the product spends. A sourced lead arrives without an email address. Enrichment is charged per person, and buying details for a whole market to write to a fraction of it is a bill with no argument for it. The approval is when someone says this person will be written to, so it is when the reveal runs.

An approved first touch still waits for intent. It leaves when the contact's heat crosses the campaign's entry threshold. They engaged a monitored keyword, they replied to something, the signal fired, so the message lands while the evidence it was written from is still warm. A workspace can switch this off per campaign, but the default is that timing is part of the message.

Full auto means no human sits between a generated message and a real person. That is exactly why the pre-send check is architecture rather than a feature flag: the suppression list is checked on every send, a reply that says stop ends outreach on every channel, and the database itself refuses to mark a message sent without a recorded verdict. An approval flow you can turn off, backed by a gate you cannot.