Abstract
This publication discloses, enables, and dates a protocol by which a human, while already speaking with an AI voice agent over a live telephone call, verbally delegates a second, downstream phone errand — and the platform autonomously performs it end to end and reports back. The mechanism has four coupled parts. First, in-call task capture: a function/tool call the model issues mid-conversation (schedulecall) converts a spoken delegation into a typed task row in the agent's own workspace, and the confirmation is spoken back within the same turn. Second, LLM workflow synthesis: a cognitive loop later dequeues the task and an agent session runs analyze → generate → execute → report, where "generate" prompts a language model to emit an executable, typed, multi-step workflow (steps drawn from a fixed action vocabulary — aigenerate a call script, placecall, callbackowner, CRM lookups/creates, email — carrying {{variable}} templating and per-step conditions). Third, first-utterance-locked outbound placement: the outbound call is enriched by parallel CRM and caller-memory lookups, and its synthesized opening line is bound by a first-utterance lock that is re-asserted at Realtime session start, tying the synthesized script to a deterministic opening disclosure; the outbound-to-owner leg auto-enters an asymmetric owner mode; and toll-fraud rails (feature flag, destination prefix allow-list, and a reserve/release rate + daily-volume limiter) gate every placement. Fourth, owner callback report: on completion the agent phones the delegating owner back and states the result. The combination is disclosed defensively to establish prior art and prevent third-party patent monopolization; a clean-room, dependency-free reference implementation accompanies it.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Assuncao, gustavo matthew, "Spoken-Delegation Phone-Errand Protocol", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10927