“I know your team is struggling with hiring velocity.”
Written by a system that never knew that. It is the line that makes a prospect distrust every other sentence in the message. GNX is built so it cannot write it.
How that is enforced
Six decisions in how context is gathered, scored, and handed to the model.
The model is told what is missing
Rather than a generic instruction not to hallucinate, the agent receives a concrete list of what is unavailable for this specific lead: no funding stage, no technology information, no headcount, no reliable first name. Naming the gaps is far stronger than warning against them.
Facts and hypotheses are labelled apart
A provider-reported headcount is a fact. “This company is struggling with pipeline efficiency” is a hypothesis. The two are kept separate, and hypotheses are only ever turned into questions — never asserted as things we know.
Readiness is an ingredient list
A lead is reported as “7 of 10 ready” with the missing pieces named, not as an arbitrary quality percentage. You can act on a missing job title. There is nothing you can do with “68%”.
Required facts are a hard gate
Identity, role, company, and a work email are required. Without them nothing is drafted at all. Optional facts change how personalized a message can be; required facts decide whether one exists.
Scored only when enrichment is finished
Readiness is calculated after enrichment reaches a terminal state. Scoring a lead mid-enrichment measures our own impatience rather than the data, and would write a thin email while richer context was moments away.
Every message is explainable
Context is captured as a stable, hashed snapshot. For any message that went out, the exact facts, campaign configuration, and model version behind it can be traced — rather than shrugging at a black box.
A draft is invalid until it proves otherwise
Generated content is not written straight to the wire. Each draft is checked, and a failure is regenerated once — then recorded as a failure rather than quietly shown to you as if it were fine.
Wrong recipientThe greeting names someone other than the actual lead.
Foreign companyThe body describes a different company than the one being written to.
Unresolved placeholdersTemplate variables that never got filled in.
Leaked formattingMarkdown or HTML bleeding into a plain-text message.
Wrong sequence stepFollow-up content saved against the first-touch slot.
Duplicate contentThe same message repeated across steps or leads.
Someone approving thirty emails in a row should not be the last line of defence against one of them being addressed to the wrong person. That is what these checks are for.
Autopilot changes who approves, not what is checked
Turning on autopilot for a campaign means messages are approved automatically instead of by you. It does not skip validation, suppression, or the send-time eligibility checks. Every message still passes the same pipeline — and a lead who replied, unsubscribed, booked, or asked not to be contacted is re-checked at the last possible moment before anything sends.
Outreach you can stand behind.
Choose a plan and see what the agent knows about a lead before it writes a word.