The AI governance layer for RevOps.
Your team already reviews AI-written code before it ships to production. Outbound is the last place where AI output goes straight to the outside world, to prospects, under your brand, from your domain, with no review layer at all. cadohq is that layer: policy is authored and versioned, a gate enforces it in the send path, overrides are logged, and leadership sees the posture of the whole program.
One principle underneath all of it: the system that generates the message can never be the system that clears it. Sequencers are paid on volume sent. A governance layer is paid on judgment, so it has to stand outside the send path it governs.
Three forces converged. Generation became free, so outbound volume multiplied. The mailbox providers turned deliverability into a machine-enforced compliance regime with a two-tenths-of-a-percent margin for error. And the teams deploying AI are scaling something they say they do not trust. Each number below is sourced and named, because a governance product should govern its own claims first.
The asymmetry that makes this a governance problem: AI generates per message, but the punishment lands per domain. One cadence past the complaint ceiling degrades inbox placement for every sequence the company runs. Microsoft matched the regime for Outlook in May 2025. The governable unit is the sending program, not the single message.
Where sequencers do ship governance controls, they govern access, not content: admin toggles that switch AI features on or off, and usage logs written after the send. Access control and monitoring are necessary. Neither one reviews what the AI wrote before a prospect reads it. Pre-deployment review of the content itself is the unowned layer, and it is the layer regulators and risk frameworks keep converging on: NIST AI RMF, ISO 42001, and the EU AI Act all require documented human oversight of AI output, and the AI Act's transparency obligations take effect August 2, 2026 with fines up to 15 million euros or 3 percent of worldwide turnover.
Sources: Google email sender guidelines · Microsoft high-volume sender requirements, April 2025 · Belkins cold email response rates, 2025 · Sopro State of Prospecting 2026 · Default, The State of AI in RevOps, 300+ teams · EU AI Act, Article 50. Vendor-published figures are attributed to their vendor and are their measurements, not ours.
Rules are machine-readable, individually owned, and versioned. Compliance rules are deterministic pass or fail. Quality rules carry the model score. Each rule has an enforcement level: block cannot be overridden in product, revise needs an approved exception, advisory annotates and ships.
The gate fails closed. A cadence ships only with a current passing verdict tied to its exact version. Not yet scored counts as hold. Any failed block rule forces hold regardless of score.
The author is never the approver. A held or revise cadence ships only when a different role clears it, with a written reason. Every override is a logged, time-boxed event, not a global off switch. This is what keeps the control on instead of routed around.
Reason given: "Verbal proof point came up on the discovery call, adding the named reference next week."
Append-only and tamper-evident. Every score, gate decision, override, and rule change is recorded with who, what, when, why, and the ruleset version in force. This is the record that answers, months later, why a sequence shipped, and the evidence a security review asks for.
The "are we governed?" view for leadership, distinct from per-rep coaching. Pass rate against the threshold, exception rate, coverage, and prevented sends, trended over time.
Governance only holds if nothing escapes it. A send that never reached cadohq is ungoverned, so coverage is tracked as a first-class number and shadow sends are flagged. Deliverability makes this concrete: one rep past the 0.3 percent complaint line degrades inbox placement for every cadence the org runs, so the governable unit is the sending program, not the single message.