AI / Wanderings 2026
Voice
A restaurant in Italy exposed the design problem. An AI voice skill needs five layers before it speaks for you.

I wanted to reserve a restaurant in Italy. The restaurant did not accept online reservations. It expected a call.
I do not speak Italian.
That left me with a small and rather old-fashioned problem. The internet stopped at the telephone number.
Before the call, I approved the number, recipient, objective, permitted information, and negotiating limits. I also consented to the synthetic version of my voice.
The assistant identified itself as an AI assistant that made the call on my behalf. It spoke Italian and made the reservation.
After the call, it returned a transcript, a summary, and a list of commitments.
The call worked. My Italian did not improve, but I got the reservation.
The experiment gave me a design rule. An AI voice skill needs five layers before it speaks for me.
Each layer controls one part of the call. Together, the layers keep authority with the person who approves the call.
1. Voice layer
The voice layer contains a synthetic version of your spoken voice. You must consent before the assistant uses it.
I tie that consent to an approved call and purpose. I do not give a synthetic voice unlimited authority.
I treat access to my synthetic voice like access to my signature. Both can represent me to another person.
The voice layer controls how the call sounds. The other layers control behavior, authority, and evidence.
2. Call skill
The call skill tells the assistant what to do during the conversation.
It defines your tone, objective, permitted commitments, forbidden topics, and the rules to stop or transfer the call.
Tone matters because the assistant speaks in your name. The objective limits the call to one task.
Permitted commitments define what the assistant can accept or promise. Forbidden topics define what it must not discuss.
If the conversation leaves those limits, the assistant must stop or transfer the call to you.
That last rule matters. A fluent voice can still make a very efficient mess.
3. Calling integration
The calling integration is a custom tool. It connects Codex to telephony and real-time voice services.
Codex applies the call skill. The telephony service connects the call. The voice service sends and receives speech in real time.
A live call can change after every reply. The custom tool lets Codex listen and answer during the call.
Without this integration, the call skill remains a set of instructions with no telephone.
The integration must let the assistant stop or transfer the call. The assistant needs an exit when the call leaves its approved path.
4. Approval layer
Before every call, I approve the number, recipient, objective, permitted disclosure, and negotiating limits.
That approval belongs to one call. It does not authorize another call. An old permission cannot become a permanent mandate.
The approval screen must show the complete request in plain language.
I need to know who receives the call, what the assistant wants, and what information it can share.
I also define its limits for negotiation. If the recipient asks for more authority, the assistant must stop or transfer the call.
This layer keeps the decision with me. The assistant acts after I approve the call.
5. Disclosure and record
The assistant starts every call with a clear disclosure. It identifies itself as an AI assistant that makes the call on my behalf.
The recipient knows who speaks and why.
After the call, the assistant returns a transcript, a summary, and a list of all commitments.
The transcript shows the words. The summary shows the result. The commitment list shows what now needs action.
The record matters because a call can create expectations. I need to know what the assistant said and promised.
Disclosure gives the recipient clear information. The record gives me evidence of the call.
Start with one real call
Start with one real call. Write the objective, permitted information, commitments, forbidden topics, and transfer rule.
Approve the number, recipient, and negotiating limits before the assistant dials.
After the call, examine the transcript, summary, and commitments.
If a limit remains vague, do not connect the voice skill to a telephone.
I still do not speak Italian. I now know exactly what must happen before the next AI call starts.