How a call happens

This page follows one call from the decision to dial, through the safety gate and the conversation, to the recorded outcome. The platform writes each outcome exactly once.

One attempt, one job

Each planned call runs as one supervised job. The job has three parts: the safety gate, the conversation, and the outcome. The conversation is only the middle part. The job controls all three parts.

Two rules apply to every attempt:

• The job is durable. The job keeps a record of its own progress. If a server stops during a call, the job continues from that record. No data about the attempt is lost.
• One attempt is exactly one dial. The platform does not redial inside an attempt. The decision to try again belongs to the lead's journey, not to this call.

The safety gate comes first

Before the platform dials, fifteen safety checks run in a fixed order. The checks ask questions such as:

• Does the lead, the campaign and the caller number all belong to the same client?
• Is consent on file?
• Is the number on a stop list?
• Is the current time a legal calling hour where the person lives?
• Did the platform call this person too often recently?
• Is the caller number healthy?
• Is an emergency stop set?
• Is a live-call slot free, platform-wide?

The first failed check stops the attempt immediately. The full list of seventeen checks has its own page.

Three facts are true for every gate decision:

• The full gate gives its answer in less than one second.
• The gate writes each decision to an audit trail, for a pass and for a block. A silent refusal is not possible.
• A blocked call does not count as an attempt. The lead keeps all of its allotted attempts. The journey records the reason for the block and adjusts its plan.

Placing the call

After the gate passes an attempt, the platform prepares the call in a strict order:

1. Claim a number. The platform claims one healthy phone number from the client's own pool. The claim is immediate. The claim cannot move the number past its daily limit. The platform cannot claim a number at its daily limit.
2. Select the voice. The platform selects the AI agent that the campaign specifies. This agent holds the conversation. The platform also hands the agent the lead's first name, the lead's business name, and the client's own name, read from the lead's file. Thus the agent greets the right person by name, and introduces itself as the right client. If the client has set an application-page address, the platform hands that too, so a request to text or email the application link sends the client's real page, not a placeholder.
3. Dial. The platform tells the carrier, Telnyx, to dial from the claimed number, on the client's own voice application. The request carries the attempt's own id, so a retry can never ring the person twice. Before a retry, the platform asks the carrier for its live calls and never dials beside one. The platform places every call itself. The voice service never dials.
4. Register the conversation when the phone rings. Two outside providers act in a strict order. When the carrier reports "ringing", the platform registers the conversation with the voice service, one call at a time, never in a batch. A batch delays the dial to a time that the provider selects. That delay makes the gate's decision out of date. Thus the platform does not use a batch. This step moved twice: from before the dial to the answer, and on 13 October 2026 from the answer to the ring.
5. Connect. When the person answers, the platform transfers the carrier's call to the AI agent's line. The conversation starts. The voice service decides whether a machine answered. The carrier does no machine detection on a campaign call. How the platform uses Telnyx gives each step from the carrier's side.

Both outside providers allow only a fixed number of calls at once, set by their own account. If either one refuses a dial for being at that limit, the platform does not retry at once. It holds back every further dial to that provider for a short time, lets the wait grow if refusals continue, and sends exactly one test dial when a wait ends, so a large group of waiting leads does not retry together and trip the same limit again. The lead loses nothing from the wait: no attempt is spent, and the lead's own history records no fault.

If the connection from the phone network to the AI agent fails after the person answers — the voice service refuses it, or its own side hangs up before the two sides join — the platform ends the call within twenty seconds, so the person is never left listening to silence. The platform treats this as its own fault, not the person's: the attempt is not counted against the lead's allotted calls.

During the conversation

The platform takes no part in the conversation. Both providers send event reports to the platform: ringing, answered, hung up, and last, "analysis ready". Each report is only a notification from the provider. The platform controls what occurs next.

The platform examines every report before the report counts:

• Each report carries a cryptographic signature. The platform verifies the signature on arrival and rejects a forged report.
• The platform rejects each report that is more than five minutes old.
• The platform accepts a duplicate report but counts it once.

The platform waits for a limited time:

• a maximum of ten minutes for the call to connect or end;
• ten more minutes for a connected call to end;
• a maximum of five minutes for the after-call analysis.

If no report arrives before a limit, the platform stops the wait and continues. A stalled call blocks the phone number that it claimed. Thus the platform sets these limits.

The outcome

Every call gets a seven-part outcome:

• Did the call connect?
• Who did the platform reach?
• What is the person's level of interest?
• Does the person qualify?
• Is there a compliance alert, for example a "stop calling me" request?
• Which actions did the agent take?
• What is the caller's situation?

The platform ranks the sources of the outcome by reliability:

1. Phone-network signals and the agent's logged actions come first. The platform treats them as fact.
2. An AI analysis of the transcript and the agent's own tool log comes second. The platform runs this analysis itself, asking the same fixed questions every time, rather than paying the voice service for its own built-in analysis of the same call. This analysis can only fill gaps. It cannot change what stronger evidence established.

The platform grades its confidence in each outcome:

• An outcome with hard evidence gets the full confidence grade.
• An interpreted outcome gets a lower grade.
• The platform does not guess an outcome below the confidence threshold. That outcome goes to a human review queue.

One rule overrides all others: a compliance alert outranks every other signal. If a person says "stop calling me", that request is the outcome of the call. This is true even if a booking also occurred. The platform suppresses the number in the same minute, with no delay.

The platform writes the outcome exactly once. The first write is permanent. Thus a duplicate report, or two reports that arrive together, cannot count a call twice.

When the platform is not sure

If the evidence cannot decide a call, the call goes to a review queue, and a person decides it. The queue puts the most severe items first. A compliance item goes to the front as urgent.

One known gap: the queue operates fully today, but the screens for the reviewer's work are not complete.