Independent implementation review · 8 September 2026

CHAU Voice AI Outbound Caller Pools

A map of how caller pools choose and reserve a number before every dial, why the safety boundaries exist, how campaign success is attributed, and how pool health changes automation.

End-to-end flow

Select a step to see its safety responsibility.

Shared physical-number cap

A number may belong to several pools, but it has one policy row and one daily allowance across those memberships. Twenty-one simultaneous reservations against cap 20 yielded exactly 20 grants.

Existing behavior preserved

Legacy single-number campaigns keep their dispatch path. Their success metrics are synchronized too, so moving to a pool is not required to see truthful performance.

Response-failure simulator

A network result is not always proof of whether a stranger's phone rang.

ReservedAllowance committed before the request.
One requestNo retry may cross this boundary.
Record responseQueue and acceptance ledger move atomically.
Final stateCalled, released, or held for review.

What earns campaign credit?

Choose an evidence pattern. The implementation accepts event identity, not coincidence.

Visible campaign evidence

Resolved calls · classified calls · human answers · attributed appointments · refresh time

Rates are hidden below 80% classification coverage. Unknown evidence is not converted into a false zero.

Deliberately absent

Revenue and sold results

These wait for a stable current-state source shared across integrations and verifiable against each client's own CRM.

Health evidence explorer

Adjust evidence quality to see why low volume or poor classification cannot justify automation.

120
110
48
55%
Insufficient evidence
Active
Watch
Cooldown
Trial

Automation needs repeated evidence

One supported decline produces Watch. Cooldown needs a second fresh observation. Re-reading the same hourly sample does not count.

No invented spam score

Performance deterioration can change allocation. It cannot prove a carrier spam label without a trusted reputation source.

Decision record

The choices most likely to be “simplified” into a regression.

DecisionEvidence / reasonRejected alternative
Reserve before dialConcurrent quotas and selected caller must be one committed fact.Dial, then increment a counter.
Hold uncertain acceptanceThe partner path may return 500 after provider acceptance.Release every non-2xx response.
Retry persistence onlyA failed ledger write does not authorize a second ring.Retry the whole network operation.
One policy per physical numberMembership in two pools cannot create two daily caps.Put limits on pool membership rows.
Exact call ID attributionProvider event identity survives audit against the canonical call record.Same contact plus a time window.
Preserve manual parkExplicit operator/reputation state outranks inventory freshness.Resume automatically when inventory returns.
Coverage before ratesUnknown classification is absent evidence.Count unknown as machine/no answer.
No new outbound flagThe existing Voice Agent Studio gate is the approved boundary.Add a second toggle and split behavior.