AI outbound works when confidence—not volume—decides what happens next. This field guide
distills the decision model behind my Claude + Clay systems—including when high-value
targets should stay human-in-the-loop even when the evidence is strong.
4 routes
automate · review · monitor · suppress
6 stages
source through measurement
1 boundary
evidence must earn the action
human-in-loop
required for high-value targets
Core assertion
THE ONE THING
The quality of an outbound system is defined by its refusal logic.
Generating research and copy is easy to demo. The harder—and more valuable—work is deciding whether the evidence has earned an action at all, and whether the account is valuable enough to require human judgment before that action happens.
The shift
BEFORECan Claude generate this?
Optimize the prompt, add more sources, and push the output downstream.
AFTERWhat evidence earns which action?
Define eligibility, preserve provenance, route ambiguity, and stop weak work upstream.
I first framed this lesson around a Slack-triggered Jina account-research workflow because
it made an obvious demo. I removed it because I did not trust the component enough to
recommend it. The systems I did trust shared a more durable pattern: rules controlled the
boundary, Claude handled ambiguity, and confidence plus account value decided the route.
The operating model
Select a stage to inspect the decision it contributes.
01 / DECISION CONTRIBUTION
Source
Capture an observable change with its source and timestamp.
Four routes—not one sequence
Confidence is not a vanity score, and it is not an automation switch. It establishes how much the system trusts the evidence. Account value and the cost of being wrong determine whether a human must approve the next action.
SELECT A SCENARIONo route selected
Choose one of the four evidence states above to see how the system responds.
Decision architecture
Evidence moves forward. Rep feedback moves back.
Evidence sources
crmSalesforce context
->
signalClay + public signals
->
voiceGong context
source · timestamp · reason code preserved
Confidence-routing decision layer
Deterministic control
rulesICP + identity checks
->
rulesFreshness validation
->
rulesSuppression policy
eligible evidence advances
Probabilistic judgment
claudeSignal classification
->
claudeGrounded context
->
claudeConstrained angle
confidence + account value determine the safest useful action
Decision route
routeAutomate · human review
->
routeMonitor · suppress
->
feedbackAccept · reject · learn
The systems behind the lesson
01 / PIPEGEN
Turn scattered signals into explainable actions.
Clay handled ingestion and orchestration. Claude classified triggers, checked fit against defined criteria, and generated constrained angles. Reason codes and source evidence traveled with every recommendation; strategic targets stayed behind a human approval step.
02 / TRUST LAYER
Make CRM data safe enough to automate.
Verified employment, contact data, opportunity state, and freshness before activation. Uncertain records went to review; stale or ineligible records were refreshed or suppressed.
Is the workflow ready to interrupt a seller?
Run this check before any recommendation becomes a task, sequence, or send.
[1]The signal is observable—not inferred from generic language.
[2]Account fit, contact identity, and data freshness are verified.
[3]Source, timestamp, confidence, and reason code travel with the recommendation.
[4]Suppressions run before a draft or seller task is created.
[5]The route reflects account value and the cost of a false positive—not confidence alone.
[6]The rep can answer: why this account, why now, and what next?
What teaching revealed
The hardest part was separating what was visually demonstrable from what I genuinely
understood. Removing the Jina workflow made the lesson less tool-centric and more honest.
It also exposed a gap in how I had talked about automation: I had spent more time explaining
what systems produced than why they were allowed to produce it.
Explaining the system clarified that “do nothing” is not a failure state. Monitoring and
suppression are valuable outputs when they protect rep attention, customer experience, and
trust in the system. The durable skill is not prompting. It is designing the decision boundary.
“A signal is only useful if it changes what the rep does next.”