Agentforce Voice and the telephony decision

Made once, lived with for years. The options differ more in integration surface and lock-in than in call quality.

1. VideoThe vocabulary: carrier, SIP, media, and where an agent attachesPreview6 min
2. VideoThe options, and what each locks you into6 min
3. ReadingCost per minute against cost per resolution5 min
4. VideoBuild versus buy, for the telephony layer itself6 min
5. VideoWhat switching providers costs you a year in6 min
6. LabLab: score three telephony options against your own constraints14 min

Where the agent sits in the call flow, and where a human can still intercept before it goes wrong.

1. VideoNumber provisioning, routing and failover6 min
2. VideoWhere the agent sits in the call flow, and where a human can intercept6 min
3. VideoFailover: what happens when the agent platform is down and the phone still rings6 min
4. ReadingA routing diagram worth drawing before you build anything5 min
5. LabLab: draw the call flow for your own routing14 min

A caller has no screen and no scroll-back. Silence past about a second reads as a dropped call.

1. VideoThe latency budget a caller toleratesSilence past roughly a second reads as a dropped call. That budget dictates how much you may look up before speaking.6 min
2. VideoWhat you may look up before speaking, inside that budget6 min
3. VideoPre-fetching and filler speech, and where each is honest6 min
4. ReadingMeasuring turn latency in a system you do not fully control5 min
5. LabLab: script one turn and measure its latency14 min
6. VideoWhere the budget breaks first: the tool call, not the model5 min

A caller talking over the agent is normal, not an edge case. Design for it from the first script.

1. VideoBarge-in, silence detection, and talking over the agent6 min
2. VideoSilence past about a second reads as a dropped call6 min
3. VideoA caller who interrupts mid-sentence, and what the agent does with the fragment6 min
4. VideoThree seconds of dead air, and what caused it last time6 min
5. ReadingDesigning the silence-handling state machine5 min
6. LabLab: test barge-in against a scripted interruption14 min

Assume the transcript is wrong somewhere. Confirm anything that gets written, and know when to ask again instead of guessing.

1. VideoDesigning for a wrong transcript rather than a clean one6 min
2. VideoWhat transcription error looks like on a real recording6 min
3. VideoConfirmation: repeating back what will be written6 min
4. VideoWhen to ask again instead of guessing at a mangled word6 min
5. ReadingNumbers, names and addresses: the three things transcription gets wrong most5 min
6. LabLab: run a script against a noisy recording and see what breaks14 min

The escalation list is written before the first call, not after the first complaint.

1. VideoDeciding what a voice agent must never do without a human6 min
2. VideoEscalation to a human, and handing over context6 min
3. VideoA caller who asks for something the agent should refuse on the phone6 min
4. ReadingWriting the escalation list before the first call, not after the first complaint5 min
5. LabLab: write the escalation list for your own agent14 min
6. VideoWhat the human on the other end needs to see when the call lands5 min

Transcription, summary, action items and the record they land on. Conversation intelligence at a defensible cost per call.

1. VideoTranscribe, summarise, extract action items6 min
2. VideoWhat lands on the record, and validating it first6 min
3. VideoCapturing a call into records that are useful afterwards6 min
4. VideoCost per call, and what it buys against a seat-based licence6 min
5. LabLab: test with real accents and real background noise14 min
6. ReadingA defensible cost-per-call number, and how to state it5 min
7. QuizCheck: voice7 min