If the turn budget is spent, hand the task to another channel

What it is

Hotel booking has already asked city, dates, and party size; turn four is still chasing bed type and a loyalty number. The task has not failed. The dialogue budget is spent: filling the rest by mouth grows recognition, memory, and occupancy cost together, while the same form on a screen can be seen at once. Turn-budget channel handoff treats “not done yet” as a designed state — when the budget is gone, change channel, rather than extending the same spoken path without limit. The cap is a written beat count, not a slogan about user patience.

Why it happens

Each turn spends three things at once: attention from an occupied primary task, slots that may be written wrong into common ground, and the user’s forecast of how much longer this will be spoken. A budget cuts away before that product turns unkind, not after error has already accumulated into a submit. Handoff is not a species of failure: failure cannot continue; handoff can continue, but should not continue in voice. The switch itself costs one transfer — carrying filled slots onto the new channel and letting the user know they have not been dropped. A blank form on the other side wastes the spoken turns, and people refuse to open their mouth next time. Budgets are per skill: a timer’s budget is one or two turns, a booking’s maybe three or four, never “until the form is full.”

Studying it

Plot completion against user-turn count, split by skill. Find the kink where completion drops or wrong submits rise; the beat before the kink is a budget candidate. The other arm is a handoff experiment: at turn N offer “finish this on the phone” versus keep chasing by voice; record acceptance, post-handoff completion, and whether filled slots are still there on the new channel.

Wizard-of-Oz can test the handoff move itself before the recognizer is stable — whether people want it, and on which turn. In logs, split “picked up the phone unprompted” from “switched after being offered”; the former is usually a budget already overrun with no exit in the product. Do not use mean turns as the budget: the long tail is who handoff should catch.

Where it stops holding

Screenless, with no reachable second channel, handoff has nowhere to go; the budget becomes “must live in voice,” which means cutting slots and confirming earlier, not pretending a phone exists. In access contexts dialogue may be the only channel; the turn cap yields to access, but the budget point should still change strategy (closed choice) rather than chase openly forever. When the user asks to “keep talking,” the budget can extend one beat, marked as still transferable. Making handoff a rebuke (“you can’t use voice”) drives the people who should leave back onto a path that no longer pays.

Applying it

• Write a hard user-turn cap per skill, reviewed with the required slot count. If slots are still missing at the cap, offer handoff by default and pre-fill what was collected.
• The handoff utterance does two jobs: acknowledge what is already done, and say that the other channel opens the same form, not a new task.
• A refused handoff can have one more turn; if still unclosed, settle into a recoverable draft rather than a fifth and sixth chase.
• How to check: the skill’s turn histogram. Median over the cap, with “opened the app mid-task” rising near the cap, means the handoff point is late or did not carry filled slots.

Related

• Same group: M1.03.1 Each extra turn stacks recognition and understanding error · M1.03.2 Every turn must make clear whose floor it is
• Nearby: M1.01 When Voice-First Is Appropriate · C7.17 Switching Between Voice and Keyboard · M3.05 Voice and Screen Complementarity
• Search terms: turn-budget channel handoff · voice-to-screen handoff · dialogue length cap