Feature

For the people who will not type

Technicians, kitchen staff and field engineers send voice notes far more readily than they type. Hold the button, say what happened, carry on. The rest works exactly as it does for a typed message.

Transcribed automatically Repeats back what it heard Any language you set

Book a free demo See pricing

In short

A voice note sent to the assistant is downloaded, transcribed and then treated as though it had been typed: it creates jobs, closes jobs, answers questions and is chased in exactly the same way. The reply opens with what the assistant heard, so a mis-heard registration or part number is visible immediately rather than turning into a wrong job nobody notices.

What happens: Transcribed, then handled as text
The reply: Opens with what it heard
Language: Set once, never guessed
Cost: About $0.002 for a 20-second note
Needs: An OpenAI key, even on Claude

One note, what happens

They hold and talk: 12 seconds
Downloaded: From WhatsApp
Transcribed: About a second
Handled as: An ordinary message
Reply opens with: What it heard

Why this one matters

The last thing standing between the product and the people it is for

Every objection this product has ever had comes down to one thing: will your team actually use it. Removing the app removed most of that. What was left was typing - and a technician with dirty hands, standing under a car, is not going to type.

They will hold a button and talk, because they already do it a hundred times a week. That is the whole feature: the thing they already do, turned into a job list.

Nothing new to learn - it is the same button they already use
Works one-handed, walking, in a workshop, in a van
The transcript becomes the message, so search and job matching see it
Everything downstream - chasing, the briefing, the weekly review - works on it unchanged

The risk, and what is done about it

It repeats back what it heard

Transcription mis-hears things, and the things it mis-hears are exactly the things that matter here: registrations, part numbers, names. A wrong job created silently from a wrong transcript is the one failure this feature could introduce.

So every reply to a voice note opens with the transcript, in quotes, before anything else. If it heard "Mrs Diaz" as "Mr Diaz" the sender sees it in the same second and can correct it. The note is also marked as spoken on the dashboard, so anybody reading the thread later knows those words were heard rather than typed.

The reply starts: I heard: "…"
The thread marks a spoken message, because a transcript is not the same as typed words
A note it could not make out is refused, not guessed at
A note that fails to download is answered, not silently dropped
The language is set once in the admin and always sent, so the model never guesses it

The practical details

Worth knowing before you switch it on

It needs an OpenAI key

Even if your assistant runs on Claude. Anthropic has no speech-to-text, so this one job goes to OpenAI on your own key. The admin says so rather than quietly doing nothing.

About a fifth of a cent

A 20-second note costs roughly $0.002 to transcribe. It appears in the same spend figures as everything else, so you can see it.

Your language

Set a two-letter code once. A model left to guess the language of a ten-second note gets it confidently wrong, so one is always sent.

Notes, not recordings

Anything over five minutes is refused with an explanation instead of being billed for. You can change the limit.

Off until you turn it on

With it off, a voice note gets a polite reply asking for text - the behaviour before this existed.

Same rules as everything

A number not on your roster is still refused. Voice changes what a message is made of, not who may send one.

Works together with

More of the assistant

Reads the daily plan Someone types what they are doing today, the way they always have. It becomes separate tracked jobs with owners and deadlines - no forms, no commands to learn. Keeps the photo with the job A picture of the damage, the reading or the finished work is attached to the job that message created - not lost in a chat nobody can search. Creates the work that repeats Temperature logs, cleaning, safety checks, calibration, a weekly recall list. You describe the pattern once and the job appears on its own, chased like any other. Reads the end-of-day update "The rotors came in, the Ford is stuck waiting on parts." Matched against that person's open jobs by meaning - one closed, one blocked with the reason recorded. Chases what is late Escalating reminders to the person, never the group. After your chosen number of attempts it stops nagging and tells the owner instead. Briefs the owner Ask for today and get a prioritised list: blocked first, then overdue, then due today, plus who has not reported. Built from the data, not written by a model.

See every feature