What Voice AI Should Leave Blank

A useful Voice assistant should reduce typing without turning uncertainty into a confident-looking record. If you say what happened but do not mention how long it took, the duration should stay blank. If two properties could match the description, the draft should ask for a choice. Nothing should be saved until you review it.

That boundary matters because a clean form can still contain a bad assumption. The goal is an editable starting point, not an automatic declaration of work performed.

A transcript is raw material

The HostHours Voice flow begins on the phone. Speech is converted to text on the device, then the transcript is sent for structured extraction with limited matching context such as the current property, available people, categories, locale, and time.

The model output is treated as an untrusted suggestion. It may organize facts into fields, but it does not become a saved ActivityEntry on its own.

Consider this spoken note:

Talked with the plumber about the leaking kitchen shutoff at Oak Street and scheduled the repair.

The note contains a task, a property clue, and a useful outcome. It does not contain the actual duration. A responsible draft can propose the likely property and a factual description, but the time field should remain incomplete. The user might have spent seven minutes or forty-five minutes. The recording itself cannot answer that question.

The same principle applies when a value is ambiguous. “Yesterday” needs a date interpretation that the user can see. “Alex” may refer to a vendor or a property manager. “Checked the unit” may need a more specific category. Structure can help, but visible uncertainty is safer than an invisible guess.

What must be reviewed before saving

A complete activity record needs a property, activity date, actual duration, person, category, and factual description. Voice can help prepare those fields. You remain responsible for checking each one.

Use this order for a quick review:

1. Confirm the property and tax-year workspace.
2. Check the date against when the work occurred.
3. Enter the actual time spent. A missing duration blocks saving.
4. Confirm who performed the work and add the exact helper or company name when relevant.
5. Choose the category that best describes the completed activity.
6. Read the factual note once for invented details, missing outcomes, or plans described as completed work.
7. Add a photo only when it gives useful context.

The review should be short because the draft has already reduced repetitive entry. It should not be skipped. A user-confirmed record is different from model output, even when every proposed field happens to be correct.

Blank is a valid result

Product forms often treat completeness as the only sign of quality. A review system needs a second measure: whether each value has a factual basis.

A blank duration is better than one copied from a calendar event. An unresolved person is better than assigning the owner by default. An empty evidence recommendation is better than implying that every task should have a photo. The draft can direct attention to missing information without manufacturing it.

This also keeps manual entry available as the dependable fallback. If Voice cannot process a transcript, the user can retry, type instead, or finish the record through the regular form. A provider problem does not erase the local input.

Privacy should match the narrow job

Voice drafting does not keep a permanent audio archive. The device creates the transcript, and the server receives text for structured extraction. Raw audio is not sent to the Voice Edge Function or saved as supporting evidence. Temporary local audio is deleted after successful transcription.

The input and draft are account-scoped and recoverable on the device. The user can retry processing or discard the draft. A stable request identity limits duplicate processing and duplicate saves.

These controls do not make model output authoritative. They keep the feature focused on one narrow task: turning the user’s words into fields that are easier to review.

The standard is user-confirmed facts

The right test for a Voice feature is not how many fields it can fill. Ask whether it leaves unsupported values visible, preserves the original meaning, survives an interruption, and waits for the user to save.

Try Voice first on a recent task whose facts you still remember. Compare the draft with what actually happened. Correct anything uncertain, then save only when the property, person, date, duration, category, and description are all yours.

HostHours organizes factual records. It does not determine tax status, material participation, deduction eligibility, or how a specific activity should be treated.