Skip to content
Capture · 7 min read

Voice to Action

Say what needs doing and have it done — an opted-in recording starts a workflow that reads the transcript and acts on it.

You finish a call. Four things were agreed, two need looking into, one needs a reply by Thursday. Normally the next twenty minutes go on writing that down.

LumaVista can do that part. Finish the recording, and a workflow reads the transcript and files what it finds — the commitments become tasks, the open questions become research, the follow-up gets drafted.

You talked. It did the filing.

The chain

Four steps, each of which exists on its own and is documented on its own:

  1. A recording completes and is transcribed — Recording and Transcribing.
  2. A trigger fires, matching that recording against workflows configured to accept it.
  3. A workflow runs — the same engine as everything else, no special case — Building Workflows.
  4. A skill does the work — extracting action items, starting research, drafting — Skills.

Nothing here is a separate subsystem. It is the ordinary automation machinery with a recording as its starting gun, which is why anything you can build a workflow to do, you can build it to do from a conversation.

Whatever it produces links back to the recording it came from. A task filed this way carries the sentence that created it, so “why is this on my list” is always answerable.

Two keys, both yours, both off

This is the most autonomous thing LumaVista does. The permission model is deliberately more awkward than it needs to be for convenience, and that is the design.

Two independent opt-ins, ANDed:

  1. This recording may be acted on. Set per recording, not once for all recordings.
  2. This workflow accepts recordings. Set per workflow.

Either one off, nothing happens. A recording you did not opt in is transcribed and stored and nothing further; a workflow that never opted in is never offered a transcript.

The obvious design is one switch — “act on my recordings” — and it is wrong. A single switch means enabling automation for one meeting enables it for every meeting, including the one next week you would not have chosen. Two keys make the default outcome nothing, and make each act of automation a thing you decided rather than a thing you forgot to prevent.

There is no bulk opt-in for past recordings. Turning this on today does not reach backwards.

What it cannot do without you

It cannot read your transcript while you are away. If your account is encrypted, the key is on your device. A workflow triggered while you are offline has nothing to open the audio or transcript with, and it waits rather than acting on partial data. Making that work is a separate, third decision — background execution — which seals your key for scheduled runs without us ever holding a usable copy. There is no recovery path, by design.

It cannot contact anyone without approval. If the workflow’s last step mails the summary or posts it to a channel, that step is gated. Automation reaching outside your account requires a person to have seen what it will say — three approval choices, all described in Unattended Actions and Approvals.

So the honest summary of what “just talk and it acts” means: it acts inside your account, on recordings you individually opted in, through workflows you individually opted in, and it stops at the boundary where something would leave.

Availability. This is in beta and off by default, and composes with recordings and workflows, which have their own switches. Ask and we will turn it on.

Setting one up

  1. Turn on recordings, and record something.
  2. Create or install a workflow that accepts recording triggers. The Voice Note → Action Items template is the worked example — instantiate it rather than starting empty.
  3. Opt the workflow in.
  4. Opt the recording in.
  5. Finish a recording and watch what it files.

Start with a workflow whose output you can throw away. The first run tells you more about whether the shape is right than any amount of configuring.