Your production partner in China sends a clean take: the actor speaks a rehearsed command and the lamp turns on. The proposed headline says, “Just say it your way.” The footage demonstrates a particular request. The headline promises flexibility. Before approving both, establish what variations the product team has actually checked.
ONCE recommends rehearsing three things separately: wording, target selection and conversational context. This helps a marketing team decide whether to demonstrate a specified command, show a bounded set of natural expressions, or revise a broader promise before a full shoot.
A developer opportunity is not a finished product claim
Apple’s November 3, 2026 announcement opens a new Siri AI Developer Preview through the 27.3 operating-system betas, with App Intents exposing app content and custom actions. It describes customer shipping as still ahead and App Store submissions opening in spring. It does not announce universal consumer availability for third-party integrations.
For a hardware brand, that distinction matters before commissioning a launch film. Platform capability does not establish that your app has implemented an action, that your device completes it, or that every paraphrase works. Existing voice integrations should not automatically be presented as implementations of this new preview.
Apple’s WWDC26 session, Build intelligent Siri experiences with App Schemas, distinguishes finding content, taking actions and understanding context, and recommends testing the complete Siri experience. The production method below adapts that distinction to advertising; it is not an implementation guide or a certification test.
The ONCE rehearsal matrix: wording, target, context
A shot list usually records what the actor says and what the device does. A natural-language claim also requires attention to what happens when the input changes.
| Axis | Question to resolve before filming |
|---|---|
| Wording | Does an equivalent request work with different phrasing, or is a specific command required? |
| Target | What happens when two devices could match the request? Can the system select appropriately or ask a suitable clarifying question? |
| Context | What makes “this one” meaningful? What happens after a screen change or when the previous conversational reference is absent? |
Record more than pass or fail. Useful outcomes include correct execution, appropriate clarification, incorrect execution and inability to handle the request. When the target is ambiguous, asking a question may be the right behavior. Removing that exchange to support a one-command headline changes what the audience is being shown.
Keep the actual input, expected behavior, observed outcome, selected target and version conditions. Work with the product team to distinguish interpretation, action dispatch and the device’s final state. A spoken completion message alone does not establish the physical result.
Eight cells can reveal omissions, not a reliability percentage
Consider a fictional appliance brand demonstrating a safe, approved lamp-dimming task. Select two phrasings, two target situations and two context situations: 2 × 2 × 2 gives eight cells. This is a planning illustration, not an ONCE test or a claim about a real Siri integration.
The phrasings might be the documented command and one paraphrase. Target situations might contain one lamp or two selectable lamps. Context might identify the intended lamp already or require it to be named again. Each combination needs a sensible expected response approved by the product team; unsupported requests should not be added merely to manufacture failures.
The exercise could reveal that adding a second lamp introduces a clarifying question. The creative choice is then to include that exchange or explicitly frame the single-device situation. Selecting the easiest cell does not justify expanding the copy to “however you ask.”
Even eight expected outcomes from eight single runs would not establish 100% accuracy. This is not a random sample, and it does not cover all accents, languages, noise levels, permissions, networks or operating conditions. A numerical reliability or speed claim needs a separate, appropriate testing approach.
Match the script to the demonstrated range
Only a specified command has been checked: show that operation without adding a claim that users never need to learn wording.
Several paraphrases have been checked: build dialogue around those observed expressions and their conditions. A few examples still do not establish unrestricted understanding.
Clarification or confirmation is needed: write it into the scene. A short exchange can communicate useful intelligence without pretending the interaction always ends in one sentence.
The integration is a prototype or developer preview: identify its actual stage and intended presentation. An internal concept film, development update and advertisement for a currently available feature serve different purposes.
For overseas teams coordinating with Chinese engineering and production partners, rehearse in the target market’s language. An English command does not validate a Mandarin voice-over, and a localized line is an actual input change, not simply a subtitle change. Recheck the final spoken wording against the relevant build.
Buy the rehearsal before expanding the shoot
Separate interaction planning, input-sample design, engineering-assisted rehearsal, filming and screen capture, editing, and later version maintenance in the proposal. Rehearsal delivers observations and unresolved questions. Production delivers the approved story and agreed assets. Extra short edits do not replace missing interaction validation.
If a specified command is ready but broader language flexibility remains uncertain, a bounded demonstration may be the sensible first production. Keep volatile dialogue and interface sections separately updateable where feasible. That creates a practical reuse discussion without assuming every software change requires an entirely new film. Any savings remain project-specific.
For ONCE’s move toward sustained content services, the useful accumulated record is the connection between an input, its conditions and the wording it supports in an ad. Such a record can inform future scripts and localization reviews. It is not a certificate of model capability.
Before approving “just say it your way,” ask the product and creative teams to choose one real task, map the three input axes and agree the expected responses. That defines the promise before production scales. See ONCE’s production planning articles for related preparation methods.
Sources and scope
Apple Developer, “Siri AI is now open to all developers,” November 3, 2026; WWDC26, “Build intelligent Siri experiences with App Schemas.” Both checked November 4. The rehearsal matrix, eight-cell example and purchasing recommendations are ONCE’s proposed methods. No client app, hardware or developer beta was tested, and no recognition accuracy was measured.
